Large-scale climate and the associated phenomena have been defined in Chapter 2 (e.g., Cross-Chapter Box 2.2) as ranging from global and hemispheric, to ocean basin and continental scales. The definition of the regional scale is case specific in the AR6 WGI Report. Section 1.4.5provides definitions of the different regional types adopted by the different chapters. In this chapter, regional scales are defined as ranging from the size of sub-continental areas (e.g., the Mediterranean basin) to local scales (e.g., coastlines, mountain ranges and cities) without prescribing any formal regional boundaries. These spatial length scales range from a few thousand down to a few kilometres and the relevant driving modes and processes at regional scales are summarized in Figure 10.3. In contrast to Chapters 11, 12 and Atlas, which make a region-by-region assessment of climate change, this chapter does not necessarily restrict itself to the use of the AR6 WGI Reference Regions (Section 1.4.5 and Atlas.1.3). Different regional definitions have been used in sections 10.4 and 10.6, selected for their adequacy to illustrate methodological aspects (e.g., for the attribution of long-term regional trends, regions that display such trends have been selected). Typological regions (Section 1.4.5 and Atlas.1.3) are used in Box 10.3 and Cross-Chapter Box 10.4.

Figure 10.3 | Schematic diagramto display interacting spatial and temporal scales relevant to regional climate change information. Figure adapted from Orlanski (1975). The processes included in the different models and model components considered in Chapter 10 are indicated as a function of these scales. The various types of models (including global and regional climate models) for constructing regional climate information are assessed in Section 10.3.1 and Box 10.3.

Open figure

The concept of a unified and seamless framework for weather and climate prediction (A. Brown et al., 2012; Hoskins, 2013) provides the context for understanding and simulating regional climate across multiple spatial and temporal scales. This concept is embodied in the subseasonal-to-seasonal (Vitart et al., 2017) and the seasonal-to-multi-annual (Smith et al., 2020) prediction activities that generate regional climate information across temporal scales. The seamless framework benefits from the convergence of methods traditionally used in weather forecasting and climate projections, in particular the role of the initialization in climate models and the strategies for the evaluation of physical processes relevant at different temporal scales.

The relatively short observational record (Section 10.2) is a primary challenge to estimate the forced signal and to isolate low-frequency, multi-decadal and longer-term internal variability (Frankcombe et al., 2015; Overland et al., 2016; Bathiany et al., 2018). Because only one realization of the actual climate exists, it is non-trivial to extract estimates of internal and forced variability from the available data (Frankcombe et al., 2015). As an alternative, approaches that use large observational ensembles can be applied (Section 10.4; McKinnon and Deser, 2018).

There is a close relationship between spatial and temporal scales (Figure 10.3). For example, an individual convective storm may exhibit scales of variability ranging from metres and seconds to kilometres and hours, while for El Niño–Southern Oscillation (ENSO) the scales of variability are regional to hemispheric in extent and multi-year in length. These scales interact and the interactions are represented in climate models, although the ability of current models to simulate regional phenomena and even large-scale climate drivers still leaves room for improvement (Section 10.3) and limits their capability to represent the interactions across spatial and temporal scales.

It is important to note that in this chapter and subsequent regional chapters, including the Interactive Atlas, the baselines and reference periods used for climate change estimates from regional models may vary from those used in Chapters 1 to 9. In these chapters three main time baselines are defined for the past, for example, pre-industrial (before 1750), early industrial (1850–1900) and recent (1995–2014), while the future reference periods are 2021–2040 (near term), 2041–2060 (mid-term) and 2081–2100 (long term) (Section 1.4.1 and Cross-Chapter Box 1.2). Regional climate simulations used in the recent literature have been performed with different baselines. The differences are often due to the availability of the boundary conditions from global simulations, leading to periods chosen for those simulations like 1950–2005, in line with the CMIP5 historical simulations followed by projections from 2005 onwards (Vaittinada Ayar et al., 2016; Zhang et al., 2017; L. Cai et al., 2018). For simulations that use CMIP3 boundary conditions other periods have been used. As a consequence, these regional simulations mix for the recent period historical simulations with projections. The mismatch needs to be considered when assessing results obtained from both global and regional models in the context of the climate information distillation process, or when linking the regional chapters to the assessments performed in previous chapters. The choice of baseline provides a source of uncertainty for the assessment of climate impacts (e.g., for the response of bird species in Africa; Baker et al., 2016). Besides, a range of different baselines may need to be considered to satisfy a variety of users, since this choice affects the perceived result (Dobor and Hlásny, 2019). The influence of the different baseline periods can be explored using the Interactive Atlas where different baselines are available, for example, 1986–2005 (according to AR5), 1995–2014 (this Report), and both 1961–1990 and 1981–2010 (WMO).

One way of overcoming the baseline uncertainty is to define the results for a given model based on specific global mean temperature changes from the pre-industrial period (e.g., Sylla et al., 2018 for West Africa; Kjellström et al., 2018 for Europe; Taylor et al., 2018 for the Caribbean; Montroull et al., 2018 for South America). The specific global mean temperature is known as global warming level (GWL; Sections 1.6.2 and 10.6.4, and Cross-Chapter Box 11.1). The GWL is a useful dimension of integration because important changes in regional climate, including many types of extremes, scale quasi-linearly with the GWLs, often independently of the underlying emissions scenarios (e.g., Hoegh-Guldberg et al., 2018; Beusch et al., 2020; Seneviratne and Hauser, 2020), always taking into account caveats described in Cross-Chapter Box 11.1. In addition, GWLs allow a separated analysis of the global and regional climate responses associated with a warming level (Section 10.6.4; Seneviratne and Hauser, 2020). The choice of global temperature goal in the context of the 2015 Paris Agreement means that there is an increasing desire for the regional climate information to be expressed as a function of GWLs.

There are important differences in the processes affected by greenhouse gases (GHGs) over land and ocean. Notably, this leads to preferential warming of the land regions, which are themselves skewed towards the Northern Hemisphere (NH).

Variations in solar forcing (Section 2.2.1) could influence regional climate through its modulation of circulation patterns, although this research field is still hampered by large observational and modelling uncertainties. The 11-year solar cycle has been suggested to affect the leading atmospheric circulation modes of the North Atlantic region in model-based studies (Gray et al., 2013; Thiéblemont et al., 2015; Sjolte et al., 2018). In particular the solar cycle has been suggested as an important source of near-term predictability of the North Atlantic Oscillation (NAO; Kushnir et al., 2019), while other studies have not found evidence for links between the solar cycle and NAO in observational records (Ortega et al., 2015; Sjolte et al., 2018; Chiodo et al., 2019). On centennial time scales, solar fluctuations were found to be correlated with the Eastern Atlantic Pattern (Sjolte et al., 2018). Possible influences on winter circulation and temperature over Eurasia (Chen et al., 2015) and North America (Liu et al., 2014; Li and Xiao, 2018) have also been identified.

An updated assessment of past changes in stratospheric ozone can be found in Section 2.2.5.2. The AR6 assesses that both GHG and stratospheric ozone depletion have contributed to the expansion of the zonal mean Hadley cell in the Southern Hemisphere (SH) for the period 1981–2000 with medium confidenceSection 3.3.3; Garfinkel et al., 2015; Waugh et al., 2015; Grise et al., 2019). There is medium confidence that stratospheric ozone depletion contributed to the strengthening trend of the summer Southern Annular Mode (SAM) for the period 1970–1990, but this influence has been weaker since 2000 (Section 3.7.2). The poleward shift of the SH westerlies has also been explained by stratospheric ozone depletion (Solman and Orlanski, 2016). Section 10.4 assesses its role in the multi-decadal increase of rainfall in south-eastern South America and Section 10.6.2 does so for the occurrence of the Cape Town drought.

Both natural and anthropogenic aerosols are often emitted at a regional scale, have a short atmospheric lifetime (from a few hours to several days; Section 6.1), are dispersed regionally and affect climate at a regional scale through radiative cooling/heating and cloud microphysical effects (Chapter 8; Rotstayn et al., 2015; Sherwood et al., 2015). The majority of aerosols scatter solar radiation, but with strong regional variations (Shindell and Faluvegi, 2009) that lead to regional radiative effects of up to two orders of magnitude larger than the global average (B. Li et al., 2016; K. Li et al., 2016; Mallet et al., 2016). Black carbon, instead, is known to absorb solar radiation, leading to regional atmospheric warming patterns due to its inhomogeneous spatial distribution (Gustafsson and Ramanathan, 2016). Patterns of forcing generally follow those of aerosol burden. However, temperature and precipitation responses are both local and remote (Z. Li et al., 2016; Kasoar et al., 2018; L. Liu et al., 2018; Samset et al., 2018; Thornhill et al., 2018; Westervelt et al., 2018). For instance, changes in aerosol concentrations in the NH have been reported to modulate monsoon precipitation in West Africa and the Sahel (Undorf et al., 2018; Section 10.4.2.1) and in Asia (H. Zhang et al., 2018; Section 10.6.3).

Natural aerosols include mineral dust, volcanic aerosol and sea salt. The feedback processes between climate and mineral dust as well as sea salt are assessed in Section 6.4, while the volcanic aerosol is dealt with in Cross-Chapter Box 4.1. Mineral dust created by wind erosion of arid and semi-arid surfaces dominates the aerosol load over some areas. The major sources of contemporary dust are located in the arid topographic basins of northern Africa, Middle East, Central and south-west Asia, the Indian subcontinent, and East Asia (Prospero et al., 2002; Ginoux et al., 2012) and emissions are controlled by changes in surface winds, precipitation, and vegetation (Ridley et al., 2014; W. Wang et al., 2015; DeFlorio et al., 2016; Evan et al., 2016; Pu and Ginoux, 2018). Dust both scatters and absorbs radiation and serves as a nuclei of warm and cold clouds (Chapter 6). The surface direct radiative effect is likely negative over land and ocean, especially when the assumed solar absorption by dust is large (Miller et al., 2014; Strong et al., 2015). Surface temperature and precipitation adjust to the direct radiative effect with both sign and magnitude depending on the dust absorptive properties. Dust often cools the surface, but in regions such as the Sahara surface air temperature increases as the shortwave absorption by dust is increased, leading to increases of surface temperature over the major reflective dust sources (Miller et al., 2014; Solmon et al., 2015; Strong et al., 2015; Jin et al., 2016; Sharma and Miller, 2017).

Volcanic eruptions load the atmosphere with large amounts of sulphur, which is transformed through chemical reactions and micro-physics processes into sulphate aerosols (Cross-Chapter Box 4.1; Stoffel et al., 2015; LeGrande et al., 2016). If the plume reaches the stratosphere, sulphate aerosols can remain there for months or years (about two to three for large eruptions) and can then be transported to other areas by the Brewer-Dobson circulation. If the eruption occurs in the tropics, its plume is dispersed across the Earth in a few years, while if the eruption occurs in the high latitudes, aerosols mainly remain in the same hemisphere (Pausata et al., 2015). The global temperature response observed after the last five major eruptions of the last two centuries is estimated to be around –0.2°C (Swingedouw et al., 2017), in association with a general decrease of precipitation (Iles and Hegerl, 2017). Nevertheless, the statistical significance of the regional response remains difficult to evaluate over the historical era (Bittner et al., 2016; Swingedouw et al., 2017) due to the small sampling of large volcanic eruptions over this period and the fact that the signal is superimposed upon relatively large internal variability (Gao and Gao, 2018; Dogar and Sato, 2019). Evidence from paleoclimate observations is therefore crucial to obtain a sufficient signal-to-noise ratio (Sigl et al., 2015). Reconstructed modes of climate variability based on proxy records allow evaluation of the influence on those modes (Zanchettin et al., 2013; Ortega et al., 2015; Sjolte et al., 2018; Michel et al., 2020).

Anthropogenic aerosols play a key role in climate change (Chapter 6). Although the global mean optical depth caused by anthropogenic aerosols did not change from 1975 to 2005 (Chapter 6), the regional pattern changed dramatically between Europe and eastern Asia (Fiedler et al., 2017, 2019; Stevens et al., 2017). Large regional differences in present-day aerosol forcing exist with consequences for regional temperature, hydrological cycle and modes of variability (Chapter 8, Section 10.6). Examples of regions with a notable role for anthropogenic aerosol forcing are the Indian monsoon region (Section 10.6.3) and the Mediterranean basin Section 10.6.4). Anthropogenic aerosols are also very relevant in many urban areas (Box 10.3; Gao et al., 2016; Kajino et al., 2017).

The SRCCL assessed that nearly three-quarters of the land surface is under some form of land use, particularly in agriculture and forest management (Jia et al., 2019). The effects of land management on climate are much less studied than land cover effects although net cropland has changed little over the past 50 years, while land management has continuously changed (Jia et al., 2019). Section 7.3.4.1 assesses the global influence of both land use and irrigation on the effective radiative forcings. Land cover changes and land management can influence climate locally, such as the urban heat island and non-locally as in the case of increased rainfall downwind of a city (Jia et al., 2019; Box 10.3) or the monsoon circulation affected by irrigation (Section 10.6.3). The influence of land cover changes and land management on regional climate extremes is assessed in Section 11.1.6.

It is very likely that the global land surface air temperature response to urbanization is negligible (Section 2.3.1.1.3). However, there is evidence that urbanization may regionally amplify the air temperature response to climate change in different climatic zones (Mahmood et al., 2014), either under present (Doan et al., 2016; Kaplan et al., 2017; X. Li et al., 2018) or future climate conditions (Argüeso et al., 2014; Kim et al., 2016; Kusaka et al., 2016; Grossman-Clarke et al., 2017; Krayenhoff et al., 2018). For instance, in northern Belgium, Berckmans et al. (2019) found that including urbanization scenarios for the near future (up to 2035) have a comparable influence on minimum temperature (increasing it by 0.6°C) to that of the GHG-induced warming under RCP8.5.

Internal climate variability on seasonal to multi-decadal temporal scales is substantial at regional scales. This variability arises from internal modes of atmospheric and oceanic variability, intrinsically coupled climate modes, and may additionally be driven by processes other than those originating the modes. It also interacts with the response of the climate system to external forcing. The teleconnections associated with the modes are useful to understand the relationship between large and regional scales (Annex IV: Modes of Variability). A description of various large-scale modes of variability can be found in Chapters 2, 3 and 8, and in Annex IV, while their future projections are assessed in Chapter 4. The specificities of their regional influence are briefly discussed here. More details of their typical temporal scales and regional influences can be found in Annex IV.

Atmospheric modes of variability may have seasonally-dependent regional effects like the North Atlantic Oscillation (NAO) in European winter (Tsanis and Tapoglou, 2019) and summer (Bladé et al., 2012; Dong et al., 2013). Even though these modes are internal to the climate system, their variability can be affected by anthropogenic forcings. For instance, the SAM (Hendon et al., 2014) is both internally driven (Smith and Polvani, 2017), but also affected by recent stratospheric ozone changes (Bandoro et al., 2014). The teleconnections between these modes of variability and surface weather often exhibit considerable non-stationarity (Hertig et al., 2015).

Due to the large ocean heat capacity and their long temporal scales, multi-annual to multi-decadal modes of ocean variability such as the Pacific Decadal Variability (PDV; Newman et al., 2016) and the Atlantic Multi-decadal Variability (AMV; Buckley and Marshall, 2016) are key drivers of regional climate change. In the case of the AMV both natural (volcanic) and anthropogenic (aerosol) external forcings are thought to be involved in its timing and intensity (Section 3.7.7). These modes not only affect nearby regions but also remote parts of the globe through atmospheric teleconnections (Meehl et al., 2013; Dong and Dai, 2015) and can act to modulate the influence of natural and anthropogenic forcings (Davini et al., 2015; Ghosh et al., 2017; Ménégoz et al., 2018b). The dynamics of the ocean modes is simultaneously affected by other modes of variability spanning the full range of spatial and temporal scales due to non-linear interactions (Figure 10.3; Kucharski et al., 2010; Dong et al., 2018). This mutual interdependence can result in changing characteristics of the connection over time (Gallant et al., 2013; Brands, 2017; Dong and McPhaden, 2017), and of their regional climate impact (Martín-Gómez and Barreiro, 2016, 2017). As with atmospheric modes of variability, the regional influence of ocean modes of variability on regional climates can be seasonally dependent (Haarsma et al., 2015).

Uncertainty and confidence are treated in the same way in regional climate change information as in larger-scale (continental and global) climate problems (Chapter 1 and Section 10.3.4). The degree of confidence in climate simulations and in the resulting climate information typically depends on the identification of the role of the uncertainties (Section 10.3.4). Since the direct verification of simulations of future climate changes is not possible, model performance and reliable (i.e., trustworthy) uncertainty estimates need to be assessed indirectly through process understanding and a systematic comparison with observations of past and current climate (Section 10.3.3; Knutti et al., 2010; Eyring et al., 2019). The observational uncertainty, which is particularly large at regional scales, also has to be taken into account (Section 10.2). These uncertainty estimates are then propagated in the distillation process to generate climate information.

Uncertainties in model-based future regional climate information arise from different sources and are introduced at various stages in the process (Lehner et al., 2020): (i) forcing uncertainties associated with the future scenario or pathway that is assumed; (ii) internal variability; and (iii) uncertainties related to imperfections in climate models, also referred to as structural or model uncertainty. However, the relative role of each of these sources of uncertainty differs between the global and the regional scales as well as between variables and also between different regions (Lehner et al., 2020). One way to address the internal variability and model uncertainties is to consider results from both multiple models and multiple realizations of the same model (Eyring et al., 2016a; Lehner et al., 2020; Díaz et al., 2021). These models are at times also combined with different weights that are a function of their performance and independence to increase the confidence of the multi-model ensemble (Abramowitz et al., 2019; Brunner et al., 2019).

Other elements that play a role are the inconsistency between the global and regional models in dynamical downscaling or the observational and methodological uncertainty in bias-adjustment methods (Sørland et al., 2018). These elements, in addition to those typical of the uncertainty in global and large-scale phenomena (Chapters 1–9), affect the overall confidence of regional climate information. This complex scene with different sources of uncertainty makes the collection of results available from multi-model, multi-member simulations most useful when synthesized through a distillation process (Section 10.5.3).

Surface or in situ observations can come from a variety of networks: climate reference networks, mesoscale weather and supersite observation networks, citizen science networks, among others, all with their strengths and weaknesses (McPherson, 2013; Thorne et al., 2018). Supersite observatories are surface and atmospheric boundary layer observing networks that measure a large number of atmospheric and soil variables at least hourly over a decade or more, ideally located in rural areas (Ackerman and Stokes, 2003; Haeffelin et al., 2005; Xie et al., 2010; Chiriaco et al., 2018). Adequate calibration of instruments, quality control and homogenization are essential in these sites. They produce valuable data needed to diagnose processes and changes in regional and local climate. Many climate datasets have been developed from in situ station observations, at different spatial scales and temporal frequencies (Annex I: Observational Products). These include sub-daily (Dumitrescu et al., 2016; Blenkinsop et al., 2017), daily (Chen et al., 2008; Camera et al., 2014; Journée et al., 2015; Funk et al., 2015; Aalto et al., 2016; Beck et al., 2017a, b; Schneider et al., 2017) or monthly time scales (Cuervo-Robayo et al., 2014; Aryee et al., 2018). Sub-daily data is useful for estimating storm surge (Mori et al., 2014) or river discharge (Shrestha et al., 2015), daily data for carbon-stock dynamics (Haga et al., 2020) or tourism (Watanabe et al., 2018), and monthly data for beach morphology (Bennett et al., 2019).

Satellite products provide a valuable complement to in situ measurements, particularly over regions where in situ measurements are unavailable. They have been discussed in earlier chapters (e.g., Chapters 2 and 8) for large-scale assessment. Currently 54 essential climate variables (ECVs; Bojinski et al., 2014) are defined by the Global Climate Observing System (GCOS) program, and passed on, for example, to NASA programmes through the Decadal Survey, to the Copernicus Climate Change Service of the European Union, to the ESA Climate Change Initiative ESA-CCI, as well as to the international collaborations with geostationary Earth orbit (GEO) satellites. Their observations are valuable (high confidence) for regional applications since they provide multi-channel images at very high spatiotemporal resolutions, typically 16 channels, 1–2 km, every 10 to 15 minutes. The advanced geostationary satellites are: Himawari-8 and 9 (Kurihara et al., 2016), GOES-East and GOES-17 (Goodman et al., 2018), Meteosat-10 and 11 (Schmetz et al., 2002) and FY-4 (Cao et al., 2014). Geostationary satellite networks or constellations form an essential component of the Global Observation System (https://www.wmo.int/pages/prog/www/OSY/GOS.html), providing measurements not only for various cloud properties and moisture but also for air quality, land and ocean surface conditions, and lightning.

Low Earth orbit (LEO) satellites, with orbits typically at 400–700 km, provide advanced measurements of the Earth’s surface. Sun-synchronous polar orbiters can also cover the polar regions, which cannot be observed with GEO satellites. Examples of LEO observations for land surface monitoring are NASA’s Landsat (Wulder et al., 2016), ESA’s Soil Moisture Ocean Salinity Earth Explorer (SMOS) mission (Kerr et al., 2012), the Sentinel missions of the Copernicus programme, and JAXA’s ALOS-2 (Ohki et al., 2019), providing high spatial resolution land surface images. Many kinds of data are accumulated for land use and land cover studies, targeting aspects like urban footprint (Florczyk et al., 2019), land-cover data (Global Land 30; CCI-LC: ESA, 2021; Chen and Chen, 2018), land surfacetemperature data (Landsat, Parastatidis et al., 2017), and surface albedo (Chrysoulakis et al., 2019).

Availability of active sensors on LEO satellites enables measurement of microphysical properties of aerosol, cloud and precipitation, which can advance regional climate studies and process evaluation studies to improve regional climate models (high confidence). An example is the polar-orbiting ‘afternoon-train’ satellite constellation (known as the A-train), incorporating Aqua, CALIPSO, Cloudsat, PARASOL, Glory and Aura satellites. Vertical profiling observations from Cloudsat (with a W-band cloud radar) and CALIPSO (with a cloud lidar) led to considerable advances in measurements of cloud microphysics (Stephens et al., 2018). Precipitation and its extremes are essential concerns of regional climate studies. The GPM (65°N–65°S, 2014–present) and the preceding TRMM (36.5°N–36.5°S, 1997–2015) with Ku-/Ka-band precipitation radars have provided three-dimensional measurements of precipitation with about 5 km resolution and sub-daily sampling (Skofronick-Jackson et al., 2017). Their non-sun-synchronous observation works to cross-calibrate the constellation satellites to produce global high-resolution mapped products of precipitation, such as Integrated Multi-satellitE Retrievals for GPM (IMERG; Huffman et al., 2007) and the Global Satellite Mapping of Precipitation (GSMaP; Kubota et al., 2007), with hourly sampling at about 11 km resolution. The CPC MORPHing technique (CMORPH) has provided 30 min interval global precipitation with about 8 km coverage since 2002 (Joyce et al., 2004). Precipitation estimations from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) is a sub-daily to daily rainfall product that covers 50°S to 50°N globally with 25 km resolution from 2000 to the present (Nguyen et al., 2019), and is used for semi-global-scale precipitation coverage (Benestad, 2018). TRMM/GPM observations have enabled estimates to be obtained for global four-dimensional convective heating (Shige et al., 2009; Tao et al., 2016; Takayabu and Tao, 2020).

The use of these data has enhanced our understanding of precipitation processes at regional scale (high confidence), such as diurnal cycles in a large river valley (H. Chen et al., 2012), and in coastal (Hassim et al., 2016; Yokoi et al., 2017) and mountainous regions (Hirose et al., 2017). Three-dimensional observations revealed the contrasts in regional characteristics of rainfall extremes in monsoon regions and continental dry regions (Sohn et al., 2013; Hamada and Takayabu, 2018). Satellite measurements are also used to evaluate climate model performance, as well as to develop new parametrizations. As a demonstration of the utility of these products in studying model bias, a subtropical cumulus congestus regime has been identified that may be implicated in the unrealistic double Inter-tropical Convergence Zone (ITCZ) found in some climate models (Takayabu et al., 2010; Hirota et al., 2011, 2014). Another example is a parametrization of a land surface model that was developed specifically for a certain soil type. By assimilating satellite brightness temperature observations with their LDAS-UT scheme, Yang et al. (2007) successfully optimized a land surface model for the Tibetan Plateau.

For application at a regional scale, it is important to consider variations in the spatiotemporal resolution of the satellite products. A simple concatenation of data in time can show artificial jumps that are artefacts of changes in calibration and processing algorithms, or related to satellite orbital stability or changing performance of the instruments (Wielicki et al., 2013; Barrett et al., 2014). Recalibration and cross-calibration are then prerequisites for obtaining homogeneous time series of measurements across different or successive satellites that can then be used to produce long series that are valid as climate data records (Kanemaru et al., 2017; Merchant et al., 2017). Scale representativeness is also an issue in utilizing soil observations (Taylor et al., 2012, 2013). Although a variety of technologies to measure soil moisture at the point scale exist (Dobriyal et al., 2012), its spatial representativeness is less than 1 m² (Ochsner et al., 2013; L. Liu et al., 2016). Therefore, to be able to use in situ soil moisture for validating coarser-scale data from satellites or models, networks of point-scale measurements are used (Crow et al., 2015; Polcher et al., 2016). Smaller networks are typically of the size of a single climate model gridcell or a satellite pixel and are suitable for monitoring watersheds, while small numbers of those representing larger areas (>100 km²) are emerging (Ochsner et al., 2013).

Derived observational products are created from raw datasets collected from surface stations, remote-sensing instruments, or research vessels, which are converted into meaningful physical quantities by applying a suitable measurement theory, using either statistical interpolation techniques (Section 10.2.2.4) or numerical atmospheric and land surface models (Bosilovich et al., 2015).

Most global observational datasets are available at coarse temporal and spatial resolution, and do not include all available station data from a particular region, due to data availability problems. Therefore, efforts have been made to develop regional or country-scale datasets (Annex I). Radar and satellite remote sensing are resources that can provide a valuable complement to direct measurements at regional scale. Examples for precipitation have been described already, some of which have been released to the community (Dinku et al., 2014; Oyler et al., 2015; Manz et al., 2016; Dietzsch et al., 2017; Yang et al., 2017; Bližňák et al., 2018; Krähenmann et al., 2018; Panziera et al., 2018; Shen et al., 2018). However, some of these datasets are limited by their short record, varying between one (Shen et al., 2018) and 64 years (Oyler et al., 2015).

Reanalysis products are numerical climate simulations that use data assimilation to incorporate as many irregular observations as possible. These products encompass many physical and dynamical processes. They generate a coherent estimate of the state of the climate system on uniform grids either at global (Chaudhuri et al., 2013; Balsamo et al., 2015), regional (Chaney et al., 2014; Maidment et al., 2014; Dahlgren et al., 2016; Langodan et al., 2017; Attada et al., 2018; Mahmood et al., 2018) or country scales (Rostkier-Edelstein et al., 2014; Krähenmannet al., 2018; Mahmood et al., 2018).

Reanalyses incorporate an increasing volume of observations from a growing number of sources over time, which sometimes presents a difficulty for trend analysis. However, regional reanalyses are valuable for regional climate assessments, since they can employ high-resolution model simulations due to their limited spatial domain. Their accuracy is also better than global reanalyses since they are often developed over regions with a high density of observational data (sometimes not freely available for all regions) to be assimilated into the model (e.g., Yamada et al., 2012). Regional reanalyses can assimilate locally dense and high-frequency observations, such as from local observation networks (Mahmood et al., 2018; Su et al., 2019) and radar precipitation (Wahl et al., 2017) in addition to the observations assimilated by global reanalyses. In some regional reanalyses, satellite-derived high-resolution sea ice (Bromwich et al., 2016, 2018) and sea surface temperature (Su et al., 2019) are also applied as lower boundary conditions. The periods of regional reanalyses are limited by the availability of the observations for assimilation and by the global reanalyses needed as lateral boundary conditions. Most regional reanalyses cover the past 10 to 30 years. There are also regional reanalysis activities that use conventional observations only, which produce consistent datasets over 60 years to capture precipitation trends, extremes and changes (Fukui et al., 2018). Existing regional reanalyses cover North America (Mesinger et al., 2006), Europe (Dahlgren et al., 2016; Jermey and Renshaw, 2016; Kaspar et al., 2020), the Arctic (Bromwich et al., 2016, 2018), South Asia (Mahmood et al., 2018), and Australia (Su et al., 2019). A project for regional reanalysis covering Japan has also started (Fukui et al., 2018), where grid spacing is between 5 and 32 km, although cumulus parametrizations are still needed to compute sub-grid scale cumulus convection. Recently, reanalyses using convection-permitting regional models have been published (e.g., Wahl et al., 2017, for central Europe).

The data assimilation schemes used in regional reanalyses are often relatively simple methods, specifically nudging (Kaspar et al., 2020) and 3DVAR (Mesinger et al., 2006; Bromwich et al., 2016; Dahlgren et al., 2016), rather than the more complex schemes implemented in state-of-the-art global reanalysis systems. This is partly due to limitations of computational resources. Recently, a number of regional reanalyses using more sophisticated methods, such as 4DVAR and Ensemble Kalman filter, have been published (Jermey and Renshaw, 2016; Fukui et al., 2018; Mahmood et al., 2018; Su et al., 2019). The regional reanalyses also incorporate uncertainties due to deficiencies of the models, data assimilation schemes and observations. To estimate uncertainties, some regional reanalyses apply data assimilation using ensemble forecasts (Bach et al., 2016). Another approach compares multiple regional reanalyses produced with different systems covering the same domain, which represents the uncertainties better than single reanalysis systems with ensemble data assimilation schemes (Kaiser-Weiss et al., 2019).

The regional reanalyses represent the frequencies of extremes and the distributions of precipitation, surface air temperature, and surface wind better than global reanalyses (high confidence). This is due to the use of high-resolution regional climate models (RCMs), as indicated by different regional climate modelling studies (Mesinger et al., 2006; Bollmeyer et al., 2015; Bromwich et al., 2016, 2018; Dahlgren et al., 2016; Jermey and Renshaw, 2016; Fukui et al., 2018; Su et al., 2019). Regional reanalyses, however, retain uncertainties due to deficiencies in the physical parametrization used in RCMs and by the use of relatively simple data assimilation algorithms (Bromwich et al., 2016; Jermey and Renshaw, 2016; Su et al., 2019). Regional reanalyses can provide estimates that are more consistent with observations than dynamical downscaling approaches, due to the assimilation of additional local observations (high confidence) (Bollmeyer et al., 2015; Fukui et al., 2018).

The usefulness of any observational dataset is conditioned by the availability and outcome of a quality control (QC) process. The objective of the QC is to verify that data are representative of the measured variable and to what degree the value could be contaminated by unrelated or conflicting factors (WMO, 2017a). Data quality assessment is key for ensuring that the data are credible and to establish trusted relationships between the data provider and the users (Nightingale et al., 2019). QC is performed for all relevant global climate datasets (e.g., Menne et al., 2018). For instance, QC informs users that old reanalysis datasets can be inconsistent in the long term because they assimilated inhomogeneous observations over the reanalyses period (Kobayashi et al., 2015). As a consequence, the evaluation against independent observations suggests that reanalyses should not be automatically regarded as climate-quality products for monitoring long-term trends at the regional level (Manzanas et al., 2014; Torralba et al., 2017). QC needs to be systematically carried out by the institutions responsible for handling the data (e.g., Cao et al., 2016b).

The QC procedure depends strongly on the specific nature of the dataset. It focuses on aspects such as the correct identification of sensor, time and location, detection of unfeasible or inconsistent data, error estimation, assessment of the adequacy of the uncertainty information and the adequacy of the documentation (e.g., Heaney et al., 2016). QC principles also apply to model data (Tapiador et al., 2017). An important piece of information provided is the representativeness error (Section 10.2.1.1; Gervais et al., 2014). When problems in the data representativeness are identified, observational datasets are provided with a quality mask (Contractor et al., 2020), or the problematic dataare either removed or corrected (Ashcroft et al., 2018). These are factors often taken into account in constructing regional climate information (Kotlarski et al., 2019).

Quality-controlled data are now produced widely at the regional level, as in the case of sub-daily precipitation records in the United Kingdom (Blenkinsop et al., 2017) and the USA (Nelson et al., 2016). However, many more datasets and variables lack the same level of scrutiny (Alexander, 2016). Quality-controlled, high-resolution observational datasets are especially needed at regional and local scales to assess models as their resolution increases (Di Luca et al., 2016; Zittis and Hadjinicolaou, 2017), although the awareness and appropriate use of the QC information is challenging (Tapiador et al., 2017) when generating regional climate information (high confidence).

Homogenization aims to make data spatially and temporally ‘homogeneous’. Changes in a homogeneous time series are solely due to large-scale climatic changes (whether forced or due to internal variability). Station data are influenced by factors that act at regional scales, from the mesoscale and local scale down to the microscale (WMO, 2019). Station time series contain inhomogeneities such as artificial jumps or trends, which hamper assessments of regional long-term trends. Typical reasons for this are the urbanization of a station’s surroundings, which can lead to warming (Hamdi, 2010; Hansen et al., 2010; Adachi et al., 2012; Jones, 2016; Y. Sun et al., 2016), or relocations outside of the urban area, which could lead to cooling (Tuomenvirta, 2001; Yan et al., 2010; Xu et al., 2013; Dienst et al., 2017, 2019). Another potential source of inhomogeneity is a change in measurement methods that affect most instruments of an observational network over a limited time span, such as the transition to Stevenson screens (Parker, 1994; Böhm et al., 2010; Brunet et al., 2011; Auchmann and Brönnimann, 2012) or to automatic weather stations (WMO, 2017b).

The above examples have been selected as they are present in many stations and without going through homogenization they could potentially have influenced global land warming estimates (Section 1.5.1). Single-break inhomogeneities tend to have a magnitude comparable to global climate change (Tuomenvirta, 2001; Venema et al., 2012) and are thus important for analyses of small regions. Also station records in national networks often have similar changes, making them important for national climate change estimates, but many of these influences are averaged out at the global scale (Jones, 2016).

The main approach to reduce the influence of inhomogeneities in station observations is statistical homogenization by comparing the data from a candidate station with those of neighbouring reference stations in conjunction with the use of metadata (Trewin, 2010). This is a challenging task because both reference and candidate records normally have multiple inhomogeneities. Three challenges should be considered. First, most of our understanding of statistical homogenization stems from the homogenization of temperature observations from dense networks. Recent studies suggest that our ability to remove biases quickly diminishes for sparse networks (Gubler et al., 2017; Lindau and Venema, 2018a). This affects early instrumental data and observations that are not strongly correlated between stations, such as wind and humidity (Chimani et al., 2018).

Second, in addition to systematic errors, homogenized data also suffer from random errors, introduced by the homogenization process. These errors are largest at the station level but are also present in network-averaged signals (Lindau and Venema, 2018b). These errors are determined by the break time series, as well as the noise series and the performance of the homogenization method, are spatially correlated, and have an impact on activities such as interpolation and statistical post-processing of climate simulations (Section 10.2.3.1). Third, the above discussion pertains to the homogenization of monthly and annual means. Homogenization of daily variability around the mean is more difficult. For daily data, specific correction methods are used (Della-Marta and Wanner, 2006; Mestre et al., 2011; Trewin, 2013; C. Zhou et al., 2021) that are able to improve the homogeneity of test cases, although recent independent validation efforts were not able to show much improvement (Chimani et al., 2018). The difference with homogenization methods of monthly and annual means may stem from assumptions on the nature of inhomogeneities for daily data, which are not yet well understood (Chimani et al., 2018).

It is virtually certain that statistical homogenization methods reduce the uncertainties of long-term estimates. Considering a decomposition of the long-term warming error into a bias and a noise uncertainty around the bias, the (trend) bias especially will be reduced, but also most of the noise uncertainty. This conclusion is based on our understanding of the causes of inhomogeneities and their statistical nature combined with the design principles of statistical homogenization methods, as well as on analytical (Lindau and Venema, 2018b), numerical (Venema et al., 2012; Williams et al., 2012) and empirical validation studies (Hausfather et al., 2016; Gubler et al., 2017; Killick et al., 2020).

The above section is about the homogenization of land stations. Satellite data has its own issues and methods for homogenization (Brinckmann et al., 2013; Huang et al., 2015; Brogniez et al., 2016). The homogenization of radiosonde data and land station data use similar methods (Haimberger et al., 2012; Jovanovic et al., 2017).

Data scarcity arises largely due to the lack of maintenance of observing stations, inaccessibility of the data held in national networks, and uneven spatial distribution of stations that lead to a low density in many regions. This is particularly problematic when trying to assess regional climate change, for which a high density of observational data is desirable. Although in several regions numerous stations provide (monthly) data covering more than 100 years for both temperature and precipitation (GCOS, 2015), large areas of the world remain sparsely covered. The post-1990 decline in the total number of stations contributing to the Global Precipitation Climatology Centre (GPCC) monthly product may be related to delays in data acquisition and not paucity of data (GCOS, 2015). This is because GPCC is the result of a single time scale, single Essential Climate Variable (ECV) and single data collection centre. There is no similar drop-off of the rainfall reports in the Global Historical Climatology Network Daily database (GHCNd, Menne et al., 2012) or the Integrated Surface Database (ISD) at the sub-daily time scale.

Kidd et al. (2017) made some assumptions about GPCC-available gauges and indicated that only 1.6% of Earth’s surface lies within 10 km of a rain gauge, and many areas of the world are beyond 100 km from the nearest rain gauge. Data scarcity is especially critical over Africa (Nikulin et al., 2012; Dike et al., 2018) but the apparent data scarcity could be due to reasons other than actual paucity of data, as stated earlier. For instance, over South Africa, the number of weather stations collecting daily temperature used in the fourth version of the Climatic Research Unit Temperature dataset (CRUTEM4, Osborn and Jones, 2014) has significantly declined since 1980 (Archer et al., 2018). Although CRUTEM4 has now been replaced by CRUTEM5 (Osborn et al., 2021) it has yet to take advantage of the significant international efforts to curate and make available improved global holdings (Rennie et al., 2014) which increased the global available station count for monthly mean temperatures. This includes additional stations from many African countries. The apparent decline in stations since the 1980s could also be due to countries not contributing their data to the SYNOP/CLIMAT networks for reasons other than having non-operational stations.

Even in Europe, precipitation station density in the widely used E-OBS gridded dataset varies largely in space and time across regions (Prein and Gobiet, 2017). This variability is partly due to the reluctance of some data owners to share their data with an international effort. Regardless of the reason, low station density is a major source of uncertainty (Isotta et al., 2015). Kirchengast et al. (2014) and O and Foelsche (2019) found that at least 2 to 5 (12) stations are required for capturing the area-averaged precipitation amount of heavy summer precipitation events on a daily (hourly) basis with a normalized root-mean-square error of less than 20%. Like the E-OBS dataset, gridded daily temperature and precipitation datasets are being developed for other regions of the world. Examples include south-east Asia (SA-OBS, Van den Besselaar et al., 2017), and Latin America and West Africa (ICA&D, Van den Besselaar et al., 2015). Despite the uneven distribution of stations in space and time, the value in these initiatives is illustrated by the large number of studies in which the data product is used. This is the case, for instance, in the work of Condom et al. (2020) over the Andes, a region with prominent data scarcity, and the African Monsoon Multidisciplinary Analyses project over West Africa (AMMA; e.g., Lebel and Ali, 2009). There have been efforts to reduce data scarcity through initiatives such as the International Surface Temperature Initiative (ISTI, Thorne et al., 2011), GHCND, and the Expanding Met Office Hadley Centre ISD with quality-controlled, sub-daily station data from 1931 (HadISD, Dunn et al., 2016).

Data scarcity arising from changing coverage in observation station networks results in substantial problems for climate monitoring (e.g., trend analysis of extreme events requires high temporal and spatial resolutions) or model evaluation (Section 10.3.3.1). It is virtually certain that the scarcity and decline of observational availability in some regions (but not necessarily globally), increase the uncertainty of the long-term global temperature and precipitation estimates. As an example, Lin and Huybers (2019) found that changes in the number of rain gauges after 1975 resulted in spurious trends in extremes of Indian rainfall in a 0.25° gridded dataset spanning the 20th century. In fact, the number of stations used to construct the gridded dataset dropped by half after 1990, leading to inhomogeneity and spurious trends (Section 10.6.3). Over the southern part of the Mediterranean, which is an area sparsely covered by meteorological stations, data scarcity can lead to large uncertainties in the different gridded datasets and strongly affect model evaluation (Section 10.6.4). Satellite observations can compensate the ground-based precipitation radar data sparsity to prevent an oversight of significant climate change signals (Yokoyama et al., 2019).

There are techniques for estimating and reconstructing missing data. The methods depend on the variable of interest, the temporal resolution (e.g., daily or monthly), and the type of climate (wet or dry), among others. There has been very little evaluation of the performance of classical and data mining methods (e.g., Sattari et al., 2017). The classical methods include the arithmetic mean, inverse distance weighting method, multiple regression analysis, multiple imputation, and single best estimator, while the data-mining methods include multilayer perceptron artificial neural network, support vector machine, adaptive neuro-fuzzy inference system, gene expression programming method, and K-nearest neighbour. Crowd-sourced data (individuals contribute their own data points to create a dataset for others to use) could play a role in minimizing data scarcity (Section 10.2.4).

Derived gridded datasets require merging data from different sources of observations and/or reanalysis data on a regular grid (Section 10.2.1.2; e.g., Xie and Arkin, 1997). However, in situ observations are distributed irregularly, especially over sparsely populated areas. This leads to an interpolation challenge. Gridded products of climate variables, including temperature and precipitation, are strongly affected (high confidence) by the interpolation method over complex orography and data scarce regions (Hofstra et al., 2008; Herrera et al., 2016).

There are two main approaches to produce gridded datasets: (i) based on in situ observations only, and (ii) combining in situ observations with remote-sensing data and/or reanalysis data. The first approach has been widely employed in regions with high station density using interpolation techniques, such as inverse-distance weighting, optimal interpolation, and kriging (Chen et al., 2008; Haylock et al., 2008; Frei, 2014; Isotta et al., 2014; Masson and Frei, 2014; Hiebl and Frei, 2016; Nguyen-Xuan et al., 2016). The second approach has been mainly applied in data-sparse regions with low station density, using simple bias adjustment, quantile mapping, and kriging techniques with in situ observations, remote-sensing and reanalysis data (Cheema and Bastiaanssen, 2012; Erdin et al., 2012; Dinku et al., 2014; Abera et al., 2016; Krähenmann et al., 2018).

Gridding of station data is affected by uncertainties stemming from measurement errors, inhomogeneities, the distribution of the underlying stations and the interpolation error, with station density being the dominant factor (Herrera et al., 2019). Uncertainty due to interpolation is typically small for temperature but substantial for precipitation and its derivatives, such as drought indices (Chubb et al., 2015; Hellwig et al., 2018). The largest uncertainties typically occur in sparsely sampled mountain areas (Section 10.2.2.5). Interpolation generally give rise to smoothing effects, such as low variability of the derived dataset with respect to the in situ observations (Chen et al., 2019). As a result, the effective resolution of gridded data is typically much lower than its nominal resolution. For instance, a 5 km gridded precipitation dataset for the European Alps has an effective resolution of about 10 to 25 km (Isotta et al., 2014). In an example for precipitation in Spain, the effective resolution converged to the nominal resolution only when at least 6 to 7 stations were inside the gridcell (Herrera et al., 2019). To account for the smoothing errors, new stochastic ensemble observation datasets have been introduced (Von Clarmann, 2014).

Spatiotemporal variability of meteorological parameters observed over mountainous areas is often large, indicating strong control exerted by local topography on meteorological parameters (Gultepe et al., 2014). Difficult access, harsh climatic conditions as well as instrumental issues make meteorological measurements extremely challenging at higher elevations (Azam et al., 2018; Beniston et al., 2018). Measurements of wind speed, temperature, relative humidity and radiative fluxes are critical for climate model evaluation, but difficult to handle due to their point-scale representativeness and small-scale spatiotemporal variability over mountainous terrain, and often need adjustment (Gultepe, 2015). High-altitude (>3000 metres) permanent meteorological stations are limited and current knowledge is mainly based on valley-bottom or low-elevation meteorological stations (Qin et al., 2009; Lawrimore et al., 2011; Gultepe, 2015; Condom et al., 2020), which, generally do not represent the higher elevation climate (Immerzeel et al., 2015; Shea et al., 2015).

Measuring precipitation amounts, especially of solid precipitation, in mountainous areas is particularly challenging due to the presence of orographic barriers, strong vertical and horizontal precipitation rate variability, and the difficulty in finding representative sites for precipitation measurements (Barry, 2012). However, the precipitation amounts can be indirectly estimated by the observed point mass balances at glacier accumulation areas representing net snow accumulation (Haimberger et al., 2012; Immerzeel et al., 2015; Sakai et al., 2015; Azam et al., 2018). There is very high confidence that precipitation measurements, especially solid precipitation, in mountainous areas are strongly affected by the gauge location and setup. Precipitation measurements are also affected by the type of measurement method, presence/absence of shielding, presence/absence of a heating system and operating meteorological conditions (Nitu et al., 2018). Solid precipitation measurements may have errors ranging from 20% to 50%, largely due to under-catch in windy, icing and riming conditions (Rasmussen et al., 2012), and therefore require corrections by applying transfer functions developed mainly from collected wind speed and temperature data (Kochendorfer et al., 2017). The latest Solid Precipitation Intercomparison Experiment (SPICE) report recommends measurements of wind speed, wind direction and temperature as the minimum standard ancillary data for solid precipitation monitoring (Nitu et al., 2018).

Recent advances in remote-sensing methods provide an alternative, but they also have limitations over mountainous areas. Different versions of the Tropical Rainfall Measuring Mission (TRMM) products were found to perform differently over mountainous areas (Zulkafli et al., 2014). Orographic heavy rainfall associated with Typhoon Morakot in 2009 was severely underestimated in all microwave products including TRMM 3B42 (Shige et al., 2013). The underestimation has been mitigated in the Global Satellite Mapping of Precipitation (GSMaP) product by considering the orographic effects (Shige et al., 2013). Studies have suggested a high accuracy of passive optical satellite (e.g., MODIS, Landsat) snow products under clear skies when compared with the field observations. However, cloud masking and sub-pixel cloud heterogeneity in these snow-cover products considerably restrict their applications (Kahn et al., 2011; Brun et al., 2015; Tang et al., 2017; Stillinger et al., 2019). Gridded datasets (e.g., CRU, GPCC Full Data Product, GPCC Monitoring Product, ERA-Interim, ERA5, ERA5-land, MERRA-2, MERRA-2 bias adjusted, PERSIANN-CDR) are of paramount importance, yet they often lack enough in situ observations to improve the temporal and spatial distribution of meteorological parameters over complex mountain terrain (Zandler et al., 2019).

Beyond climate monitoring, the quality and availability of multiple observational reference datasets play a central role in model evaluation. In fact, when using observations for model evaluation, there are multiple examples where inter-observational uncertainty is as large as the inter-model variability. This has been shown for various aspects of the Indian monsoon (Section 10.6.3; Collins et al., 2013a) and for precipitation uncertainties over Africa (Section 10.6.4; Nikulin et al., 2012; Sylla et al., 2013; Dosio et al., 2015; Bador et al., 2020) and Europe (Prein and Gobiet, 2017). Kotlarski et al. (2019) compared three high-resolution observational temperature and precipitation datasets (E-OBS, a compilation of national/regional high-resolution gridded datasets, and the EURO4M-MESAN 0.22° reanalysis based on a high-resolution limited-area model) with five EURO-CORDEX RCMs driven by ERA-Interim. Generally, the differences between RCMs are larger than those between observation datasets, but for individual regions and performance metrics, observational uncertainty can dominate. They also showed that the choice of reference dataset can have an influence on the RCM performance score. Over the high mountain Asia region and East Asia, differences among gridded precipitation datasets can generate significant uncertainties in deriving precipitation characteristics (J. Kim et al., 2015; Kim and Park, 2016; Guo et al., 2017). Over western North America, observational uncertainty induces differences in multi-decadal precipitation trends (Lehner et al., 2018). Taking a very different perspective, the agreement between model simulations may be used to estimate the uncertainty and quality of observations (Massonnet et al., 2016). There is high confidence that an ensemble of multiple observational references at a regional scale is fundamental for model performance assessment. The uncertainties vary according to region, season, and statistical properties (Cross-Chapter Box 10.2).

Statistical downscaling, bias adjustment and weather generators are post-processing methods used to derive climate information from climate simulations. They all require observational data for calibration as well as evaluation (Section 10.3.3.1). Typically, the so-called perfect prognosis methods use quasi-observations for the predictors (i.e., reanalyses) and actual observations for the predictands (the surface variables of interest). By contrast, bias adjustment methods use observations only for the predictands. Weather generators typically require only observed predictands, although some are conditioned on observed predictors as well. Very often these methods are based on daily data, because of user needs, but also because of the limited availability of sub-daily observations and the limited ability of climate models to realistically simulate sub-daily weather (Iizumi et al., 2012). Some methods are calibrated on the monthly scale, but some of the generated time series are then further disaggregated to the daily scale (e.g., Thober et al., 2014). A few methods, mainly weather generators, represent sub-daily weather (Mezghani and Hingray, 2009; Kaczmarska et al., 2014). Many methods simulate temperature and precipitation only, although some also represent wind, radiation and other variables. The limited availability of high quality and long observational records typically restricts these applications to a few cases (Verfaillie et al., 2017; Pryor and Hahmann, 2019). Overall, there is high confidence that limited availability of station observations, including variables beyond temperature and precipitation as well as sub-daily data, limit the use of statistical modelling of regional climate.

All the limitations and challenges of observational data discussed in Section 10.2.2 also apply to its use for post-processing of climate model data. High quality and long observational data series are particularly relevant to quantify uncertainties. Different reanalyses present significant discrepancies when used as key predictor variables at the daily scale and may even affect the downscaled climate change signal (Brands et al., 2012; Dayon et al., 2015; Manzanas et al., 2015; Horton and Brönnimann, 2019). There is high confidence that reanalysis uncertainties limit the quality of statistical downscaling in some regions, although no assessment has been made for the most recent reanalysis products.

An important issue for bias adjustment is the correct representation of the required spatial scale. Ideally, bias adjustment is calibrated against area-averaged data of the same spatial scale as the climate model output. Hence, high-quality observed gridded datasets with an effective resolution close to the nominal model resolution are required. Driven by the need to also generate regional-scale information in station-sparse regions, researchers have considered derived datasets that blend in situ and remote-sensing data to produce high-resolution observations to be used as predictands (Sections 10.2.1.2 and 10.2.2.4; Haiden et al., 2011; Wilby and Yu, 2013).

Following some early concept studies, the first practical applications of paleoclimate data assimilation over past centuries used only selected data to reconstruct past climate changes for analysis of a specific process or case (Widmann et al., 2010). Recently, assimilation of multiple series from various data sources, including tree rings, ice cores, lake cores, corals, and bivalves, has allowed production of reconstructions that can be widely shared and applied to multiple purposes, as with modern reanalyses (Hakim et al., 2016; Franke et al., 2017; Steiger et al., 2018; Tardif et al., 2019). Most of these paleo-reanalyses are global but there are products using regional models or targeted at specific regions such as Europe, East Africa and the Indian Ocean (Fallah et al., 2018; Klein and Goosse, 2018).

Paleo-reanalyses are enabling a new range of applications and have already provided useful information on seasonal-to-multi-decadal climate variability over past millennia. They are useful tools to study the co-variance between variables at interannual-to-centennial time scales and at regional to global spatial scales. In particular, they have highlighted the processes that can be responsible for changes in continental hydrology at multi-decadal time scales (Franke et al., 2017; Klein and Goosse, 2018; Steiger et al., 2018). Paleo-reanalyses have confirmed a large contribution of internal variability in past changes at regional scale during the pre-industrial period, superimposed on a weak common signal due to forcing changes (Goosse et al., 2012) and the absence of a globallycoherent warm period in the common era before the recent warming (Neukom et al., 2019). Reconstructions of the atmospheric state obtained in the reanalysis also provide robust evidence of a local enhancement of warming or cooling conditions due to changes in atmospheric circulation, such as for the warm conditions in some European regions around 950–1250 CE, the cooling observed in 1809/1810, or the cold and rainy 1816 summer in Europe (Cross-Chapter Box 4.1; Goosse et al., 2012; Hakim et al., 2016; Franke et al., 2017; Schurer et al., 2019).

Model-based regional climate projections are all based upon some type of global model, including state-of-the-art Earth system models (ESMs), coupled atmosphere–ocean general circulation models (GCMs) or atmosphere-only general circulation models (AGCMs) (see Section 1.5.3.1). They are collectively referred to as global models.

State-of-the-art global models are generally used to derive climate information at continental to global scales both for past and future climates (e.g., Chapters 3 and 4). The nominal horizontal resolution in CMIP5 global models is typically 100–200 km. The effective resolution, for which the shape of the kinetic energy spectrum is simulated correctly, is about three to five times larger (Klaver et al., 2020), and a similar relationship also applies to RCMs (Skamarock, 2004). This strongly limits their ability to resolve local details. Since AR5 the progress in reducing biases and providing more credible regional projections by global models has been moderate in spite of the more realistic representation of a number of processes and the increase in resolution of some models. For AR6, several of the new CMIP6 (Eyring et al., 2016a) model intercomparison projects (MIPs) address some of these limitations. The list of MIPs is provided in Chapter 1 (Table 1.3). High-Resolution MIP (HighResMIP; Haarsma et al., 2016) and Global Monsoons MIP (GMMIP; Zhou et al., 2016) specifically address the regional climate challenge using global models. HighResMIP focuses on producing global climate projections at a horizontal resolution of around 50 km grid spacing or finer while GMMIP aims at better understanding and predicting the monsoons.

An alternative to increasing resolution everywhere is offered by variable resolution global models, that is, with regionally finer resolution. They have been developed since the 1970s (Li, 1999), resulting in a first coordinated effort (SGMIP) by Fox-Rabinovitz et al. (2006, 2008). They are expected to offer the finest resolution possible in the region of interest, while still resolving the climate processes at the global scale (although at lower resolution). An overview of recent developments is in McGregor (2015). This is a rapidly developing field (Krinner et al., 2014; Ferguson et al., 2016; Huang et al., 2016) that will possibly contribute to improved future regional projections.

Regional climate models (RCMs) are dynamical models similar to global models that are applied over a limited area, but with a horizontal resolution higher than that of standard global models. They are the basis for dynamical downscaling to produce sub-continental climate information (e.g., Chapters 11, 12 and Atlas) but are also often used for process understanding. At lateral and, if applicable, lower boundaries, RCMs take their values from a driving dataset, which could be a global model or a reanalysis. RCMs are typically one-way nested: they do not feed back into the driving model, although two-way nested global model-RCM simulations have been performed that examine regional influence on large-scale climate, potentially improving it (Lorenz and Jacob, 2005; Harris and Lin, 2013; Junquas et al., 2016). Spectral nudging (Kida et al., 1991; Waldron et al., 1996; von Storch et al., 2000; Kanamaru and Kanamitsu, 2007) can increase consistency with the driving model, whereby selected variables, such as the wind field, are forced to closely follow a prescribed large-scale field over a specified range of spatial scales. RCMs can inherit biases from the driving global model in addition to producing biases themselves (Hall, 2014; Hong and Kanamitsu, 2014; Dosio et al., 2015; Takayabu et al., 2016). The consistency between the circulation features simulated by the RCM and those inherited through the boundary conditions depends on (i) the relative importance of the large-scale forcing compared to local-scale phenomena, and (ii) the size of the RCM domain (e.g., Diaconescu and Laprise, 2013). Large domains also allow the RCM to generate much of its own internally generated unforced variability (Nikiema et al., 2017, and references therein; Sanchez-Gomez and Somot, 2018).

The Coordinated Regional Climate Downscaling Experiment (CORDEX) initiative (Giorgi et al., 2009; Giorgi and Gutowski, 2015; Gutowski Jr. et al., 2016) provides ensembles of high-resolution historical (starting as early as 1950) and future climate projections for various regions. RCMs in CORDEX typically have a horizontal resolution between 10 and 50 km. But much finer spatial resolution is required to fully resolve deep convection, an important cause of precipitation in much of the world. Therefore, an emerging strand in dynamical downscaling employs simulations at convection permitting scales, at horizontal resolutions of a few kilometres, where deep-convection parametrizations can be switched off, approximately simulating deep convection (Prein et al., 2015; Stratton et al., 2018; Coppola et al., 2020). A recent study indicates that switching off the deep-convection parametrization may be beneficial also in simulations performed at coarser resolutions (Vergara-Temprado et al., 2020). Alternatively, some RCMs make use of scale-aware parametrizations that are able to adapt to increasing resolution without switching off the convection scheme (Hamdi et al., 2012; De Troch et al., 2013; Plant and Yano, 2015; Giot et al., 2016; Termonia et al., 2018; Yano et al., 2018).

RCMs have often consisted of atmospheric and land components that do not include all possible Earth system processes and therefore neglect important processes such as air-sea coupling (in standard RCMs sea surface temperatures, SSTs, are prescribed from global model simulations or reanalyses) or the chemistry of aerosol–cloud interaction (aerosols prescribed with a climatology), which may influence regional climate projections. Therefore, some RCMs have been extended by coupling to additional components like interactive oceans, sometimes with sea ice (Kjellström et al., 2005; Somot et al., 2008; Van Pham et al., 2014; Sein et al., 2015; Ruti et al., 2016; Zou and Zhou, 2016a; Zou et al., 2017; Samanta et al., 2018), rivers (Sevault et al., 2014; Lee et al., 2015; Di Sante et al., 2019), glaciers (Kotlarski et al., 2010), and aerosols (Zakey et al., 2006; Zubler et al., 2011; Nabat et al., 2015). The coupling of these components allows for the investigation of additional climate processes such as regional sea level change (Adloff et al., 2018), ocean–land interactions (Lima et al., 2019; Soares et al., 2019a), or the impact of high-frequency ocean–atmosphere coupling on the climatology of Mediterranean cyclones (Flaounas et al., 2018).

An alternative or addition to dynamical downscaling is the use of statistical approaches to generate regional projections. In AR5 these methods were collectively referred to as statistical downscaling, but their performance assessment has received little attention. A major conclusion was that a wide range of different methods exist and a general assessment of their performance is difficult (Flato et al., 2014). Since AR5, several initiatives have been launched to improve the understanding of statistical approaches such as VALUE (Validating and Integrating Downscaling Methods for Climate Change Research, now merged into the EURO-CORDEX activities; Maraun et al., 2015), STaRMIP (Statistical Regionalization Models Intercomparisons and Hydrological Impacts Project; Vaittinada Ayar et al., 2016) and BADJAM (Bias ADJustment of climate scenarios for Agricultural Model applications; Galmarini et al., 2019). The performance of different implementations of these approaches will be assessed in Section 10.3.3.7.

Transient simulations intend to represent the evolving climate state of the Earth system (Chapter 4). They are typically based on coupled global model simulations, such as those in the Diagnostic, Evaluation and Characterization of Klima (DECK) and ScenarioMIP part of CMIP6 covering the period 1850–2100 (Eyring et al., 2016a), and HighResMIP (1950–2050; Haarsma et al., 2016). Global transient climate simulations may be further downscaled by either dynamical or statistical downscaling. Currently available CORDEX RCM simulations (1950–2100) are based on CMIP5 (Gutowski Jr. et al., 2016).

In contrast, time-slice experiments are designed to represent only a specific period of time (typically 30 years). They are often run using global and regional models in atmosphere-only mode, forced by SSTs derived either from observations, as AMIP experiments, or from historical simulations and future projections of coupled global models. Compared to transient simulations, they offer advantages in being computationally cheaper (due to the lack of coupled ocean and short duration), which allows for the number of ensemble members(T. Zhang et al., 2016), and/or the resolution (Haarsma et al., 2013b; Davini et al., 2017) to be increased. Convection-permitting simulations, both covering the globe or particular regions, are currently conducted for short time slices only (Kendon et al., 2017; Hewitt and Lowe, 2018; Coppola et al., 2020; Pichelli et al., 2021). Another high-resolution time-slice data base is d4PDF (Mizuta et al., 2017; Ishii and Mori, 2020). Experiments covering a limited integration period have been carried out for coupled ocean–atmosphere RCMs (Sein et al., 2015; Zou and Zhou, 2016b, 2017). However, long spin-up periods are required to reach a stable stationary state in the deep ocean that otherwise might lead to invalid projections (Planton et al., 2012; Soto-Navarro et al., 2020).

Results from downscaling experiments often suffer from large-scale circulation biases in the driving global models such as misplaced storm tracks (Section 10.3.3.4), while changes in atmospheric circulation are often uncertain owing to both climate response uncertainty (Section 10.3.4.2) and internal variability (Section 10.3.4.3). In a given application, if one can assume that changes in the regional climate are dominated by thermodynamic rather than by circulation changes, so-called pseudo-global warming (PGW) experiments (Schär et al., 1996) may be helpful in mitigating the effects of circulation biases, and to fix the large-scale circulation to present climate. In classical PGW experiments, boundary conditions for the downscaling are taken from reanalysis data, but modified according to the thermodynamic signals of climate change. The boundary conditions thus represent the sequence of observed weather, but with adjusted temperatures, humidity and atmospheric stability. Recent applications of PGW experiments include assessments of climate change in Japan (Adachi et al., 2012; Kawase et al., 2012, 2013), the Los Angeles area (Walton et al., 2015), Hawaii (C. Zhang et al., 2016), and the Alps (Keller et al., 2018). Recently, PGW studies have been generalized to modify global model simulations with the objective of separating the drivers of regional climate change, such as the Mediterranean amplification (e.g., Brogli et al., 2019b; Section 10.3.2.3).

Equivalent simulations can be conducted for individual events, thereby allowing for very high resolution. With counterfactual past climate conditions, such simulations can be used for conditional event attribution (Trenberth et al., 2015; Chapter 11), using hypothetical future conditions to generate physical climate storylines of how specific events may manifest in a warmer climate. The approach has been employed to study extreme events that require very high resolution simulations such as tropical cyclones (Lackmann, 2015; Takayabu et al., 2015; Lau et al., 2016; Kanada et al., 2017a; Gutmann et al., 2018; Patricola and Wehner, 2018; J. Chen et al., 2020) or convective precipitation events (Pall et al., 2017; Hibino et al., 2018). The range of possible events is broader and has included Korean heatwaves (Kim et al., 2018) and monsoon onset in West Africa (Lawal et al., 2016). However, if only individual events are simulated, no immediate conclusions can be derived for changes to the occurrence probability of these events (F.E.L. Otto et al., 2016; Shepherd, 2016a).

Sensitivity studies are used to identify the impact of a specific forcing, driver or process on regional climate phenomena and changes and improve the process understanding. The influence of a single external forcing can be assessed with transient historical simulations within two different frameworks (Bindoff et al., 2013; Gillett et al., 2016). The first entails simulations taking prescribed (often observed) changes only in the external forcing of interest, the others being fixed at a constant value (often pre-industrial). The second framework is based on simulations in which all external forcings are applied other than the one of interest. Both approaches may not give the same results since the climate response to a range of forcings is not necessarily equal to the sum of climate responses to individual forcings (Ming and Ramaswamy, 2011; Jones et al., 2013; Schaller et al., 2013; Shiogama et al., 2013; Marvel et al., 2015; Deng et al., 2020).

To study the influence of internal variability, new approaches such as partial coupling simulations are now routinely used since AR5. These are coupled ocean–atmosphere simulations in which the interaction between atmosphere and ocean is only one-way over a specified ocean basin or sub-basin and two-way everywhere else. Different implementations have been used such as SST anomaly Newtonian relaxation at the air–sea interface or prescription of wind-stress anomalies from reanalysis (Kosaka and Xie, 2013, 2016; England et al., 2014; McGregor et al., 2014; Douville et al., 2015; Deser et al., 2017a). Such simulations have been applied to identify the regional impacts of the Pacific Decadal Variability (PDV) and Atlantic Multi-decadal Variability (AMV) (Kosaka and Xie, 2013; Watanabe et al., 2014; Delworth et al., 2015; Boer et al., 2016; Ruprich-Robert et al., 2017, 2018).

Nudging experiments have been used to identify the relative roles of dynamic and thermodynamic processes in climate model biases and specific extreme events (Wehrli et al., 2018, 2019). Another related framework is used to evaluate the impact land conditions have on a climate phenomenon in a pair of experiments with one simulation serving as control run, and a perturbed simulation with prescribed land conditions (i.e., soil moisture, leaf area index, or surface albedo) characterizing a specific state of the land surface (i.e., afforestation or deforestation). The difference between the perturbed and control simulations enables a robust assessment of the possible impact of land conditions on events like droughts and heatwaves (Seneviratne et al., 2013; Stegehuis et al., 2015; Hauser et al., 2016, 2017; van den Hurk et al., 2016; Vogel et al., 2017; Rasmijn et al., 2018; Strandberg and Kjellström, 2019).

RCM sensitivity simulations have been used in a similar way to assess the contribution of external forcings and large-scale drivers to projected regional climate change (Nabat et al., 2014; Brogli et al., 2019a, b) and the influence of selected drivers on observed extreme events (Meredith et al., 2015b; J. Wang et al., 2017; Ardilouze et al., 2019).

In summary, there is robust evidence that sensitivity experiments are key to assessing the influence of different forcings and drivers on regional climate change.

In recent years, the role of internal variability in the interpretation of climate projections has become clearer, particularly at the regional scale (Section 10.3.4.3). A considerable fraction of CMIP5 and CMIP6 resources has been invested in generating an ensemble of centennial or multi-centennial control simulations with constant external forcings (Pedro et al., 2016; Rackow et al., 2018). As part ofthe CMIP6 DECK (Eyring et al., 2016a) pre-industrial control (piControl) simulations have been conducted (Menary et al., 2018). Similarly, control simulations with present-day conditions (pdControl) have been performed to represent internal variability under more recent forcing conditions (Pedro et al., 2016; Williams et al., 2018). Control simulations have been used to study the role of internal variability, teleconnections and many other fundamental aspects of climate models (Z. Wang et al., 2015; Krishnamurthy and Krishnamurthy, 2016). Control simulations are also used along with large ensembles of historical or scenario simulations to assess the characteristics of the regional internal climate variability (Olonscheck and Notz, 2017).

Experiments driven by quasi-perfect boundary conditions or predictors (observations or reanalysis) can be useful to evaluate downscaling performance (Frei et al., 2003; Laprise et al., 2013), including the simulation of observed past trends (Lorenz and Jacob, 2010; Zubler et al., 2011; Nabat et al., 2014; Gutiérrez et al., 2018; Drugé et al., 2019; Bozkurt et al., 2020) and the added value of downscaling compared to the reanalysis fields (Section 10.3.3.2). Although the reanalysis model itself can introduce biases especially for non-assimilated variables (such as precipitation) it is assumed that in such a setting, discrepancies between the modelled and observed climate arise mostly from errors in the downscaling method (Laprise et al., 2013) or internal climate variability generated by the downscaling method (Böhnisch et al., 2020; Ehmele et al., 2020). Since AR5, reanalysis-driven RCMs have been extensively evaluated for many regions, especially in the CORDEX framework (see region specific examples in the Atlas).

Over Europe, the VALUE initiative assessed statistical downscaling for marginal, temporal, and spatial aspects of temperature and precipitation including extremes, and performed a process-based evaluation of specific climatic phenomena (Gutiérrezet al., 2019; Maraun et al., 2019a). Alternatively, statistical downscaling can be evaluated in so-called perfect model or pseudo-reality simulations (Charles et al., 1999), where a high-resolution climate model simulation is used as a proxy for a hypothetical present and future realities. A statistical downscaling model is first calibrated with this pseudo present-day climate and, subsequently, assessed whether it correctly reproduces the pseudo-future conditions (Dixon et al., 2016).

Since AR5, model evaluation has made use of a broad combination of diagnostics (Colette et al., 2012; Kotlarski et al., 2014; Eyring et al., 2016b; Gleckler et al., 2016; Ivanov et al., 2017, 2018; Vautard et al., 2021), ranging from long-term means to indices of extreme events (Zhang et al., 2011; Sillmann et al., 2013) or a combination of these (Dittus et al., 2016). This evaluation has shown that global models have pervasive biases in some aspects of their large-scale behaviour (Section 1.5.3.1, Chapter 3). More complex diagnostics are used to characterize specific meteorological phenomena (Sprenger et al., 2017), such as feedbacks in the El Niño–Southern Oscillation (ENSO; Bellenger et al., 2014), Madden-Julian Oscillation (MJO) characteristics (Benedict et al., 2014; Jiang et al., 2015; D. Kim et al., 2015; Ahn et al., 2017), extratropical modes of variability (Lee et al., 2019), cyclone tracking (Neu et al., 2013; Flaounas et al., 2018), front detection (Hope et al., 2014; Schemm et al., 2015), thunderstorm environment parameters (Bukovsky et al., 2017), African easterly waves (McCrary et al., 2014; Martin and Thorncroft, 2015), land–atmosphere coupling (Spennemann and Saulo, 2015; Santanello et al., 2018), and sea–atmosphere coupling (Bellenger et al., 2014; Mayer et al., 2017).

New diagnostics for multivariate dependencies are needed to characterize compound events (Section 11.8; Hobaek Haff et al., 2015; Wahl et al., 2015; Sippel et al., 2016, 2017; Tencer et al., 2016; Bevacqua et al., 2017; Careto et al., 2018; Zscheischler et al., 2018). However, their success depends on the availability of adequate observational data (Section 10.2.2). Multivariate dependencies discovered in compound events can also be used for designing and evaluating multivariate bias adjustment and statistical downscaling. Process-based diagnostics are useful for identifying the cause of model errors, although it is not always possible to associate a systematic error with a specific cause (Eyring et al., 2019). AR5 discussed two approaches of process-based evaluation: (i) the isolation of physical components or parametrizations by dedicated experiments (Section 10.3.2.4) and (ii) diagnostics conditioned on relevant regimes, usually synoptic-scale weather patterns. The regime-based approach has been used with both global models (e.g., Barton et al., 2012; Catto et al., 2015; Taylor et al., 2019) and RCMs (Endris et al., 2016; Bukovsky et al., 2017; Whan and Zwiers, 2017; Pinto et al., 2018), but also with perfect prognosis and bias adjustment methods (Marteau et al., 2015; Addor et al., 2016; Beranová and Kyselý, 2016; Soares and Cardoso, 2018; Soares et al., 2019b).

Recent studies highlight the importance of user-defined or user-relevant diagnostics for model evaluation (Maraun et al., 2015; Rhoades et al., 2018; Rössler et al., 2019b; Nissan et al., 2020). Diagnostics have been used to assess the performance of climate models to produce useful input data for impact models as in the comparison between RCMs and convection-permitting models to capture flood-generating precipitation events in the Alps (Reszler et al., 2018). Alternatively, the observed impact can be compared to that simulated by an impact model that uses input from both observations and climate models. This approach has been used to evaluate the influence of statistical downscaling and bias adjustment on hydrological (Rojas et al., 2011; H. Chen et al., 2012; Gutiérrez et al., 2019; Rössler et al., 2019b), agricultural (Ruiz-Ramos et al., 2016; Galmarini et al., 2019), forest and wildfire (Abatzoglou and Brown, 2012; Migliavacca et al., 2013) (Bedia et al., 2013), snow depth (Verfaillie et al., 2017), and regional ocean modelling (e.g., Macias et al., 2018).

There is high confidence that to assess whether a climate model realistically simulates required aspects of present-day regional climate, and to increase confidence of future projections of these aspects, evaluation needs to be based on diagnostics taking into account multiple variables and process understanding.

Obtaining regional information from global simulations may involve a range of different methods (Section 10.3.1). An approach with higher complexity or resolution is useful if it adds further, useful information to that of a reference model. Section 10.5 discusses the set of considerations that determine if the information is useful. This further useful information is often referred to as added value and is a function of variables, processes, and the temporal and spatial scales targeted taking into account the needs of specific users (Di Luca et al., 2012; Ekström et al., 2015; Giorgi and Gutowski, 2015; Torma et al., 2015; Rummukainen, 2016; Falco et al., 2019). There is no common definition of added value, but here it is considered a characteristic that arises when one methodology gives further value to what another methodology yields.

Downscaling is expected to improve the representation of a region’s climate compared to the driving global model (Di Luca et al., 2015). Arguably, there should be a clear physical reason for the improvement, which is applicable to the evaluation of added value in downscaled projections (Giorgi et al., 2016). The added value depends on the region, season, and governing physical processes (Lenz et al., 2017; Schaaf and Feser, 2018). Thus, added value of downscaling global model simulations is most likely where regional- and local-scale processes play an important role in a region’s climate, for example in complex or heterogeneous terrain such as mountains (Lee and Hong, 2014; Prein et al., 2016b), urban areas (Argüeso et al., 2014), along coastlines (Feser et al., 2011; Herrmann et al., 2011; Bozkurt et al., 2019), or where convective processes are important (Prein et al., 2015). Examples of model improvements and added value are given in the following subsections and the Atlas.

A first step in determining added value in downscaling is to analyse whether the downscaling procedure gives detail on spatial or temporal scales not well-resolved by a global model, thus potentially representing climatic features missing in the GCM. This added detail, referred to as potential added value (PAV; Di Luca et al., 2012), is insufficient for demonstrating added value in downscaling (Takayabu et al., 2016), but lack of PAV indicates that the downscaling method lacks usefulness. Added value is not guaranteed simply by producing model output at finer resolution. It depends on several factors, such as the simulation setup and the specific climatic variables analysed (Di Luca et al., 2012; Hong and Kanamitsu, 2014; Xue et al., 2014). A variety of performance measures are needed to assess added value (Section 10.3.3.1; Di Luca et al., 2016; Wilks, 2016; Ivanov et al., 2017, 2018; Soares and Cardoso, 2018).

A further challenge, especially at increasingly higher resolutions, is that adequate observational data may not be available to assess added value (Section 10.2, e.g., Di Luca et al., 2016; Zittis et al., 2017; Bozkurt et al., 2019). This implies a need for additional efforts to obtain, catalogue and quality-control higher resolution observational (or observation-based) datasets (Thorne et al., 2017; Section 10.2). Univariate demonstration of added value is necessary, but may be insufficient, as better agreement with observations in the downscaled variable may be a consequence of compensating errors that are not guaranteed to compensate similarly as climate changes. Multivariate analysis of added value is better able to demonstrate physical consistency between observed and simulated behaviour (Prein et al., 2013a; Meredith et al., 2015a; Reboita et al., 2018).

Regional climate is often controlled by large-scale weather phenomena, modes of variability and teleconnections (e.g., Sections 2.3 and 2.4, Annex IV). In particular, extreme events are often caused by specific, in some cases persistent, circulation patterns (Sections 11.3–11.7). It is therefore important for climate models to reasonably represent not only continental, but also regional climate and its variability for such extremes. As explained in Section 3.3.3, standard resolution global models can suffer biases in the location, occurrence frequency or intensity of large-scale phenomena, such that statements about a specific regional climate and its change can be highly uncertain (Hall, 2014). RCMs have difficulties improving especially large-scale circulation biases, although some successful examples exist. But due to their enhanced representation of complex topography and coastlines, RCMs may add value to simulating the regional expression of teleconnections. Bias adjustment cannot mitigate fundamental misrepresentations of the large-scale atmospheric circulation (Maraun et al., 2017, Cross-Chapter Box 10.2). This subsection illustrates the relevance of large-scale circulation biases for regional climate assessments with selected examples from the mid- to high latitudes and tropics.

10.3.3.3.1 Mid- to high-latitude atmospheric variability phenomena: Blocking and extratropical cyclones

Major large-scale meteorological phenomena for mid- to high latitude mean and extreme climate include atmospheric blocking and extratropical cyclones (Section 2.3.1.4). Atmospheric blocking is characterized by a quasi-stationary, long-lasting, high-pressure system that blocks and diverts the movement of synoptic cyclones (Woollings et al., 2018). In regions where blocking occurs, it is known to lead to cold conditions in winter and warmth and drought during summer, defining the seasonal regional climate in certain years (Sousa et al., 2017, 2018b). Extratropical cyclones are storm systems that propagate preferentially in confined storm-track regions, characterized by large eddy kinetic energy, heat and momentum transports that shape regional weather at mid- to high latitudes (Shaw et al., 2016). Given their importance in shaping mean and extreme regional climate (Sections 3.3.3.3, 11.3 and 11.4), an accurate representation of blocking and extratropical cyclones in global and regional climate models is needed to better understand regional climate variability and extremes as well as to project future changes (Section 11.7.2; Grotjahn et al., 2016; Mitchell et al., 2017; Rohrer et al., 2018; Huguenin et al., 2020). An overview of CMIP5 and CMIP6 model performance in simulating blocking and extratropical cyclones is given in Section 3.3.3.3. CMIP6 models still suffer from long-standing blocking biases identified in previous generations of models. However, blocking location has improved compared to CMIP5, while comparable performance is seen for blocking frequency and persistence (Figure 10.7). Increasing horizontal model resolution to about 20 km in the HighResMIP experiments improves the representation of blocking frequency and its spatial pattern in most models, but no clear effect could be shown for blocking persistence. Biases associated with these two phenomena are highly region- and season-dependent and their amplitudes vary among CMIP models (Drouard and Woollings, 2018; Schaller et al., 2018; Woollings et al., 2018; Harvey et al., 2020; Schiemann et al., 2020).

Figure 10.7 | Northern Hemisphere blocking performance in historical coupled simulations for different multi-model ensembles. Coupled Model Intercomparison Project Phases 5 and 6 (CMIP5/6): CMIP5 and CMIP6 Diagnostic, Evaluation and Characterization of Klima (DECK) historical simulations, 1950–2005, LC/HC: Low- (LC)/high- (HC) resolution coupled simulations from the PRIMAVERA project, 1950–2014 following the hist-1950 experiment of the CMIP6 HighResMIP Protocol (Haarsma et al., 2016). (Top) blocking frequency, for example, fraction of blocked days; (middle) root-mean-squared error in blocking frequency; (bottom) 90th percentile of blocking persistence, aggregated over an Atlantic domain (left, ATL: 90°W–90°E, 50°–75°N) and a Pacific domain (right, PAC: 90°E–270°E, 50°–75°N). Results are for boreal winter (December–January–February, DJF) and summer (June–July–August, JJA). Box-and-whisker plots for CMIP5/6 follow the methodology used in Figure 10.6 and show median (line), mean (triangle), and interquartile range (box) across 29 models for each ensemble. The reference estimate (ERA, asterisk) is from a 50-year reanalysis dataset that merged ERA-40 (1962–1978) and ERA-Interim (1979–2011) reanalyses. An estimate of internal variability for each metric (IV) is shown as a box-and-whisker plot over the asterisk and is obtained from a single-model ensemble (ECMWF-IFS high-resolution hist-1950 experiment, 6 × 65 years). For details on the methodology see (Schiemann et al., 2020). Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

RCMs have a very limited ability to reduce large-scale circulation errors of the driving GCM (Hall, 2014). In a study of five ERA-Interim-driven RCMs, Jury et al. (2018) showed that RCMs typically simulate fewer blocking events over Europe than are present in the driving data, irrespective of the RCM horizontal resolution. Based on a simple blocking bias-decomposition method, they suggest that blocking frequency biases can contribute to the RCM mean surface biases. Over some large domains, reanalysis-driven RCMs can significantly improve the representation of storm characteristics compared to the driving reanalysis near regions with complex orography and/or large water masses (Poan et al., 2018). However, this is not necessarily true if the domain is large enough because the RCM and its biases will then control the circulation leading to a biased performance with regard to storm characteristics (Pontoppidan et al., 2019). An ensemble of 12 RCMs with and without air-sea coupling reasonably reproduced the climatology of Mediterranean cyclones, and air-sea coupling had a rather weak impact (Flaounas et al., 2018). Over the Gulf Stream, however, air-sea coupling played an important role in representing cyclone development (Vries et al., 2019). Sanchez-Gomez and Somot (2018) showed that the effect of RCM internal variability on density of cyclone tracks is very significant and larger than for other variables such as precipitation. It is larger in summer than in winter, in particular over the Iberian Peninsula, northern Africa and the eastern Mediterranean, which are regions of enhanced cyclogenesis during the warm season.

Biases in the representation of large-scale atmospheric circulation can result in biased representation of regional climate. While, in principle, the connection between large-scale and regional biases is obvious, given the strong control of regional climate by large-scale phenomena, research on this connection is still limited. Munday and Washington (2018) relate CMIP5 model rainfall biases over South Africa to anomalous low-level moisture transport across high topography due to upstream wind biases and inaccurate representation of unresolved orographic drag effects. Addor et al. (2016) show that the overestimated frequency of westerly synoptic situations was a significant contributor to the wet bias in several RCMs in winter over Switzerland. Pepler et al. (2014, 2016) suggest that better capturing westerly-driven synoptic systems such as cold fronts and cut-off lows in climate models could be key in simulating the observed pattern correlation between rainfall and zonal wind in southern south-east Australia. Cannon (2020) shows global improvement in performance going from CMIP5 to CMIP6 for both frequency and persistence of circulation types.

The robust quantification of the influence of atmospheric circulation errors on regional climate remains a challenge as many parametrized processes such as cloud radiative effects and soil moisture or snow feedbacks can also contribute and interact with the circulation errors. Atmospheric nudging experiments where the simulated circulation is constrained to be close to that observed have been used to separate the circulation effect from other contributions to regional climate biases (Wehrli et al., 2018). The nudging approach requires detailed and careful implementation in order to limit detrimental effects due to the added tendency term in the model equations (Zhang et al., 2014; Lin et al., 2016). Based on single-model experiments, Wehrli et al. (2018) show that the circulation induced biases are often not the main contributors to mean and extreme temperature and precipitation biases for many regions and seasons.

There is high confidence that atmospheric circulation biases can deteriorate the model representation of regional land surface climate. Assessing the relative contributions of atmospheric circulation and other sources of bias remains a challenge due to the strong coupling between the atmosphere and other components of the climate system, including the land surface.

Regional climate is shaped by a wide range of weather phenomena occurring at scales from about 2000 km to 2 km (Figure 10.3). These modulate the influence of large-scale atmospheric phenomena and create the characteristic and potentially severe weather conditions. The climate in different regions will be affected by different mesoscale phenomena, of which several may be relevant. A skilful representation of these phenomena is a necessary condition for providing credible and relevant climate information for a given region and application. Therefore, it is important to understand the strengths and weaknesses of different model types in simulating these phenomena. The performance of different dynamical climate model types to simulate a selection of relevant mesoscale weather phenomena is assessed here.

10.3.3.4.1 Convection including tropical cyclones

Convection is the process of vertical mixing due to atmospheric instability. Deep moist convection is associated with thunderstorms and severe weather such as heavy precipitation and strong wind gusts. Convection may occur in single locations, in spatially extended severe events such as supercells, and organized into larger mesoscale convective systems such as squall lines or tropical cyclones, and embedded in fronts (see below). Shallow and deep convection are not explicitly simulated but parametrized in standard global and regional models. In consequence, these models suffer from several biases. AR5 has stated that many CMIP3 and CMIP5 models simulate the peak in the diurnal cycle of precipitation too early, but increasing resolution and better parametrizations help to mitigate this problem (Flato et al., 2014). Similar issues arise for RCMs with parametrized deep convection (Prein et al., 2015), which also tend to overestimate high cloud cover (Langhans et al., 2013; Keller et al., 2016).

Non-hydrostatic RCMs at convection-permitting resolution (4 km and finer) improve features such as the initiation and diurnal cycle of convection (Zhu et al., 2012; Prein et al., 2013a, b; Fosser et al., 2015; Stratton et al., 2018; Sugimoto et al., 2018; Finney et al., 2019; Berthou et al., 2020; Ban et al., 2021; Pichelli et al., 2021), the triggering of convection by orographic lifting (Langhans et al., 2013; Fosser et al., 2015), and maximum vertical wind speeds in convective cells (Meredith et al., 2015a). Also spatial patterns of precipitation (Prein et al., 2013a, b; Stratton et al., 2018), precipitation intensities (Prein et al., 2015; Fumière et al., 2020; Ban et al., 2021; Pichelli et al., 2021), the scaling of precipitation with temperature (Ban et al., 2014), cloud cover (Böhme et al., 2011; Langhans et al., 2013) and its resultant radiative effects (Stratton et al., 2018), as well as the annual cycle of tropical convection (Hart et al., 2018) are improved. Phenomena such as supercells, mesoscale convective systems, or the local weather associated with squall lines are not captured by global models and standard RCMs. Convection-permitting RCM simulations, however, have been shown to realistically simulate supercells (Trapp et al., 2011), mesoscale convective systems, their life cycle and motion (Prein et al., 2017; Crook et al., 2019), and heavy precipitation associated with a squall line (Kendon et al., 2014). There is high confidence that simulations at convection-permitting resolution add value to the representation of deep convection and related phenomena.

Convection is the key ingredient of tropical cyclones. An intercomparison of high-resolution AGCM simulations (Shaevitz et al., 2014) showed that tropical cyclone intensities appeared to be better represented with increasing model resolution. Takayabu et al. (2015) have compared simulations of typhoon Haiyan at different resolutions ranging from 20 km to 1 km (Figure 10.8). While the eyewall structure in the precipitation pattern was strongly smoothed in the coarse resolution simulations, it was well-resolved at the highest resolution. Gentry and Lackmann (2010) found similar improvements in simulating hurricane Ivan for horizontal resolutions between 8 km and 1 km. High-resolution coupled ocean–atmosphere simulations improve the representation of the radial structure of core convection and thereby the rapid intensification of the cyclone (Kanada et al., 2017b). There is high confidence that convection-permitting resolution is required to realistically simulate the three-dimensional structure of tropical cyclones.

Figure 10.8 | Hourly accumulated precipitation profiles (mm hour–1) around the eye of Typhoon Haiyan. Represented by (a) Global Satellite Mapping of Precipitation (GSMaP) data (multi-satellite observation), (b) Guiuan radar (PAGASA), (c) Weekly Ensemble Prediction System (WEPS) data (JMA; 60 km), (d) NHRCM (20 km), (e) NHRCM (5 km), and (f) WRF (1 km) models. Panels (b), (d–f) are adapted from Takayabu et al. (2015), CCBY3.0 https://creativecommons.org/licenses/by/3.0. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

Initial studies with convection-permitting global models suggests that improvements in representing convection, as described for RCMs above, have a positive impact on the tropical and extratropical atmospheric circulation and, thus, regional climate (Satoh et al., 2019; Stevens et al., 2019; see also Section 8.5.1.2 and Chapter 7). Computational constraints currently limit these simulations to a length of few months only, such that they cannot yet be used for routine climate change studies.

Both SRCCL (Jia et al., 2019) and SROCC (Hock et al., 2019) highlight the weaknesses of climate models at simulating atmosphere–surface feedbacks. The performance at simulating some of these feedbacks is assessed below (climate feedbacks in urban areas are discussed in Box 10.3).

The snow-albedo feedback contributes to enhanced warming at high elevations (Section 8.5; Pepin et al., 2015). Global models often do not simulate it realistically due to their misrepresentation of orography in complex terrain (Hall, 2014; Walton et al., 2015). The elevation dependence of historical warming, which is partly caused by the snow-albedo effect, is realistically represented across Europe by the ENSEMBLES RCMs (Kotlarski et al., 2015). Some EURO-CORDEX RCMs simulate a spring snow–albedo feedback close to that observed, whereas others considerably overestimate it (Winter et al., 2017). In a multi-physics ensemble RCM experiment, the cold bias in north-eastern Europe is amplified by the albedo feedback (García-Díez et al., 2015). For the Rocky Mountains, RCM simulations generally reproduce the observed spatial and seasonal variability in snow cover, but strongly overestimate the snow albedo (Minder et al., 2016). There is high confidence (medium evidence and high agreement) that RCMs considerably improve the representation of the snow-albedo effect in complex terrain.

Soil-moisture feedbacks influence changes in both temperature and precipitation. More than 30% of CMIP5 models overestimate the influence of preceding precipitation (a proxy for soil moisture) on temperature extremes in Europe and the USA (Donat et al., 2018), and many CMIP5 models simulate an unrealistic influence of evaporation on temperature extremes for wet regions in Europe and the US (Ukkola et al., 2018). RCMs were found to realistically simulate the correlation between latent and sensible heat fluxes and temperature (coupling strength) over Africa (Knist et al., 2017; Careto et al., 2018) and in northern and southern Europe, but to overestimate it in central Europe (Knist et al., 2017). Land surface models driven by global reanalysis agreed relatively well with observations. However, the coupling strength varied strongly across models at the regional scale, and a realistic partitioning of the incoming radiation into latent and sensible heat fluxes did not necessarily result in a realistic soil moisture-temperature coupling (Gevaert et al., 2018; Boé et al., 2020a).

Evaluating the representation of soil-moisture–precipitation feedbacks in climate models is challenging as different processes may induce feedbacks including moisture recycling, boundary-layer dynamics and mesoscale circulation. Moreover, the effects of soil moisture on precipitation may be region and scale dependent and may even change sign depending on the strength of the background flow (Taylor et al., 2013; Froidevaux et al., 2014; Guillod et al., 2015; Larsen et al., 2016; Tuttle and Salvucci, 2016). On seasonal-to-interannual time scales, CMIP5 models showed a stronger soil-moisture–precipitation feedback than estimated by satellite data (Levine et al., 2016). Taylor et al. (2013) found that convection-permitting RCMs perform well at simulating surface-induced mesoscale circulations in daytime convection and the observed negative soil moisture feedback, whereas an RCM with parametrized convection, even when run at the same resolution, simulated an unrealistic positive feedback. There is medium evidence and high agreement that simulations at convection-permitting resolution are required to realistically represent soil-moisture–precipitation feedbacks.

Ocean–atmosphere RCMs have successfully been used to understand and simulate phenomena involving strong regional feedbacks like tropical cyclones in the Indian Ocean (Samson et al., 2014), Indian summer monsoon (Samanta et al., 2018), East Asian summer monsoon (Zou et al., 2016), near coastline intense precipitation in the Mediterranean (Berthou et al., 2015, 2018), air-sea fluxes influencing heat and humidity advection over land (Sevault et al., 2014; Lebeaupin Brossier et al., 2015; Akhtar et al., 2018) or snow bands in the Baltic region (Pham et al., 2017). The positive impact of ocean-coupling on the simulation of strongly convective phenomena such as Medicanes, a class of severe cyclones in the Mediterranean, can only be diagnosed when using relatively fine atmospheric resolution of about 10 km (Akhtar et al., 2014; Flaounas et al., 2018; Gaertner et al., 2018). A positive impact of ocean coupling has been quantified in marginal sea regions with reduced large-scale influence (e.g., in the Baltic Sea area during weak phases of the NAO and thus weak influence of Atlantic westerlies (Kjellström et al., 2005; Pham et al., 2018). There is some evidence that coupled ocean components also positively impact RCM simulations of inland climates such as precipitation extremes in central Europe (Ho-Hagemann et al., 2017; Akhtar et al., 2019). There is high confidence that coupled ocean–atmosphere RCMs improve the representation of ocean–atmosphere feedbacks and related phenomena.

The influence of ice-sheet mass balance on regional climate, explored with global and regional models by (Noël et al., 2018; Fettweis et al., 2020), is discussed in Section 9.4.

Dust, with its regional character in both emissions and climatic influences, has traditionally been specified in climate simulations with a climatological estimate. In CMIP5 models, the influence of vegetation changes on mineral dust is largely underestimated while the influence of surface wind and precipitation are overestimated, resulting in a low bias of dust load (Pu and Ginoux, 2018). Interactive dust emission modules that simulate the dust optical depth in most of the key emission regions have only been recently introduced (Pu and Ginoux, 2018). However, coarse dust is underestimated in global models (Adebiyi and Kok, 2020). Simulations of future changes in dust are hindered by the uncertainties in future regional wind and precipitation as the climate warms (Evan et al., 2016), in the effect of CO₂ fertilization on source extent (Huang et al., 2017), in the dust feedbacks (Evans et al., 2019), and in the effect of human activities that change land use and disturb the soil, including cropping and livestock grazing, recreation and urbanization, and water diversion for irrigation (Ginoux et al., 2012).

Volcanoes also provide forcings with a marked regional impact (Cross-Chapter Box 4.1). This implies that models are expected to capture these effects (Bethke et al., 2017). Both proxy analyses and simulations have demonstrated reduced Asian monsoon precipitation after tropical and Northern Hemisphere (NH) volcanic eruptions due to reduced humidity and divergent circulation (Man and Zhou, 2014; Zhuo et al., 2014; F. Liu et al., 2016; Stevenson et al., 2016). Global model experiments (Zanchettin et al., 2013; Ortega et al., 2015; Sjolte et al., 2018; Michel et al., 2020) have suggested that tropical volcanic eruptions (larger than the one from Mount Pinatubo in 1991) may lead to a positive phase of the winter NAO in the following few years (with an uncertainty on the exact years affected), but this influence is not well-reproduced in climate models and requires very large ensembles (Driscoll et al., 2012; Toohey et al., 2014; Swingedouw et al., 2017; Ménégoz et al., 2018b). The ability to simulate the effect of volcanic aerosol in global models is evaluated in VolMIP (Zanchettin et al., 2016). Given the relevance of volcanic aerosol, a good knowledge of the initial conditions is important because the response has proven to be sensitive to them (Ménégoz et al., 2018a; Zanchettin et al., 2019). A few decadal prediction systems have illustrated that current systems can predict some aspects of regional climate a few years in advance (Swingedouw et al., 2017; Illing et al., 2018; Ménégoz et al., 2018a; Hermanson et al., 2020). However, a better performance requires information about volcanic location (Haywood et al., 2013; Pausata et al., 2015; Stevenson et al., 2016; F. Liu et al., 2018a), strength (Emile-Geay et al., 2008; H.-G. Lim et al., 2016; F. Liu et al., 2018b), and seasonality (Stevenson et al., 2017; Sun et al., 2019a, b).

Some recent regional climate changes can only be simulated by climate models if anthropogenic aerosols are correctly included (Sections 10.4.2.1, 10.6.3 and 10.6.4; Chapters 6 and 8). Examples of the importance of correctly representing anthropogenic aerosols are the recent enhanced warming over Europe (Nabat et al., 2014; Dong et al., 2017), the cooling over the East Asian monsoon region, leading to a weakening of the monsoon (Section 8.3.2.4; Song et al., 2014; Q. Wang et al., 2017), as well as changes in the monsoons of West Africa (Sections 8.3.2.4 and 10.4.2.1) and South Asia (Sections 8.3.2.4 and 10.6.3; Undorf et al., 2018). The relevance of appropriately representing anthropogenic aerosols has been widely studied in regional models (Boé et al., 2020a; Gutiérrez et al., 2020), with an advantage for models with interactive aerosol schemes (Drugé et al., 2019; Nabat et al., 2020). Without a fully coupled chemistry module, radiative forcing can be simulated by including simple models of sulphate chemistry or specifying the optical properties from observations and prescribing the effect of aerosols on the cloud-droplet number (Fiedler et al., 2017, 2019; Stevens et al., 2017). In all cases, the specification of the aerosol load limits the trustworthiness of the simulations at the regional scale when enough detail is not provided (Samset et al., 2019; Shonk et al., 2020; Z. Wang et al., 2021).

The inclusion of irrigation in global and regional models over the South Asian monsoon region (Section 10.6.3) has been found to be important to represent the monsoon circulation and rainfall correctly (Lucas-Picher et al., 2011; Guimberteau et al., 2012; Shukla et al., 2014; Tuinenburg et al., 2014; Cook et al., 2015a; Devanand et al., 2019). Similarly, the inclusion of irrigation over northern India and western Pakistan could be important for the correct simulation of precipitation over the Upper Indus Basin in northern Pakistan (Saeed et al., 2013). Irrigation in the East African Sahel inhibits rainfall over the irrigated region and instead enhances rainfall to the east, coherent with both observations and theoretical understanding of the local circulation anomalies induced by the lower surface air temperatures over the irrigated region (Alter et al., 2015). Although several studies show how modelled irrigation reduces daytime temperature extremes, few compare modelled results with observations. Global model studies have found improvements in simulated surface temperature when including irrigation (Thiery et al., 2017), in particular in areas where the model used has a strong land-atmosphere coupling (Chen and Dirmeyer, 2019). An RCM study over the North China Plain showed that the inclusion of irrigation led to a better representation of the observed nighttime warming (Chen and Jeong, 2018).

There is medium confidence that representing irrigation is important for a realistic simulation of South Asian monsoon precipitation. There is limited evidence that including irrigation in climate models improves the simulation of maximum and minimum daily temperatures as well as precipitation for other regions.

Regional land-radiation management, including modifying the albedo through, for instance, no-tillage practices, has been suggested as a measure to decrease regional maximum daily temperatures (see review in Seneviratne et al., 2018), but although modelled results and theoretical understanding are coherent, few studies have verified the results with observations. Hirsch et al. (2018) is an exception, showing that implementing minimal tillage, crop residue management and crop rotation in a global model over regions where it is practiced, improves the simulation of surface heat fluxes.

The performance of statistical downscaling models, bias adjustment and weather generators is determined by the chosen model structure (e.g., to represent variability and extremes or spatial dependence) and, if applicable, the predictors selected (Maraun et al., 2019a). The VALUE initiative has assessed a range of such methods in a perfect-predictor experiment where the predictors are taken from reanalysis data (Maraun et al., 2015, 2019a; Gutiérrez et al., 2019). Table 10.2 shows an overview comprising performance results from VALUE and other studies. These results isolate the performance of the statistical method in the present climate. The overall performance in a climate change application also depends on the performance of the driving climate model (Sections 10.3.3.3–10.3.3.6) and the fitness of both the driving model and the statistical method for projecting the climatic aspects of interest (Section 10.3.3.9).

This section assesses how well climate models perform at realistically simulating historical regional climatic trends. Current global model ensembles reproduce global to continental-scale surface temperature trends at multi-decadal to centennial time scales (CMIP5, CMIP6), but underestimate precipitation trends (CMIP5) (Sections 3.3.1 and 3.3.2). For regional trends, AR5 concluded that the CMIP5 ensemble cannot be taken as a reliable representation of reality and that the true uncertainty can be larger than the simulated model spread (Kirtman et al., 2014). Case studies of regional trend simulations by global models can be found in Sections 10.4.1 and 10.6, and region-by-region assessments in the Atlas. A key limitation for assessing the representation of regional observed trends by single transient simulations of global models (or downscaled versions thereof) is the strong amplitude of internal variability compared to the forced signal at the regional scale (Section 10.3.4.3). Even on multi-decadal time scales, an agreement between observed and individual simulated trends would be expected to occur only by chance (Laprise, 2014).

In the context of downscaling, the ability of downscaling methods to reproduce observed trends when driven with boundary conditions or predictors taken from reanalysis data (which reproduce the observed internal variability on long time scales) can be assessed. For temperature in the continental USA, reanalysis-driven RCMs skilfully simulated recent spring and winter trends, but did not reproduce summer and autumn trends, (Bukovsky, 2012). Over Central America, observed warming trends were reproduced (Cavazos et al., 2020). In contrast, a reanalysis-driven coupled atmosphere–ocean RCM covering the Mediterranean could not reproduce the observed SST trend (Sevault et al., 2014).

Similar studies have been carried out for statistical downscaling and bias adjustment using predictors from reanalyses (or in case of bias adjustment, dynamically downscaled reanalyses). For a range of different perfect prognosis methods, Huth et al. (2015) found that simulated temperature trends were too strong for winter and too weak for summer. The performance was similar for the different methods, indicating the importance of choosing informative predictors. Similarly, Maraun et al. (2019b) found that the performance of perfect prognosis methods depends mostly on the predictor and domain choice (for instance, temperature trends were only captured by those methods including surface temperature as predictor). Bias adjustment methods reproduced the trends of the driving reanalysis, apart from quantile mapping methods, which deteriorated these trends.

RCM experiments are often set up such that changes in forcing agents are included only via the boundary conditions, but not explicitly included inside the domain. Jerez et al. (2018) demonstrated that not including time-varying GHG concentrations within the RCM domain may misrepresent temperature trends by 1–2°C per century. Including the past trend in anthropogenic sulphate aerosols in reanalysis-driven RCM simulations substantially improved the representation of recent brightening and warming trends in Europe (Nabat et al., 2014; see Sections 10.3.3.6 and 10.6.4, and Atlas.8.4). Similarly, Bukovsky (2012) argued that RCMs may not capture observed summer temperature trends in the USA because changes in land cover are not taken into account. Barlage et al. (2015) have revealed that including the behaviour of groundwater in land schemes increases the performance of an RCM model to represent climate variability in the central USA. Hamdi et al. (2014) found that an RCM that did not incorporate the historical urbanization in the land-use, land-cover scheme is not able to reproduce the warming trend observed in urban stations, with a larger bias for the minimum temperature trend.

Overall, there is high confidence that including all relevant forcings is a prerequisite for reproducing historical trends.

AR5 stated that confidence in climate model projections is based on the physical understanding of the climate system and its representation in climate models. A climate model’s credibility for future projections may be increased if the model is able to simulate past variations in climate (Sections 10.3.3.8, 10.4.1 and 10.6; Flato et al., 2014). In particular, the credibility of downscaled information depends on the quality of both the downscaling method and of the global model providing the large-scale boundary conditions (Flato et al., 2014). Credibility is closely linked to the concept of adequacy or fitness-for-purpose (Section 1.5.4.1; Parker, 2009). From a regional perspective, one may ask about the fitness of a climate model for simulating future changes of specific aspects of a specific regional climate. The required level of model fitness may depend on the user context (Section 10.5). A key challenge is to link performance at representing present and past climate (Sections 10.3.3.3–10.3.3.8) to the confidence in future projections (Section 1.3.5; Baumberger et al., 2017) and it is addressed in this subsection.

A general idea of model fitness for a given application may be obtained by checking whether relevant large- (Section 10.3.3.4) and regional-scale (Sections 10.3.3.5 and 10.3.3.6) processes are explicitly resolved (Figure 10.3). The basis for confidence in climate projection is a solid process understanding (Flato et al., 2014; Baumberger et al., 2017). Thus, the key to assessing the fitness-for-purpose of a model is the evaluation of how relevant processes controlling regional climate are represented (Collins et al., 2018). A process-based evaluation may be more appropriate than an evaluation of the variables of interest (e.g., temperature, precipitation), because biases in the latter may in principle be reduced if the underlying processes are realistically simulated (Cross-Chapter Box 10.2), while individual variables may appear as well-represented because of compensating errors (Flato et al., 2014; Baumberger et al., 2017). Combining a process-based evaluation with a mechanistic explanation of projected changes further increases confidence in projections (Bukovsky et al., 2017). Fitness-for-purpose can also be assessed by comparing the simulated response of a model with simulations of higher resolution models that better represent relevant processes (Baumberger et al., 2017). For instance, Giorgi et al. (2016) have corroborated their findings on precipitation changes comparing standard RCM simulations with convection-permitting simulations.

The evaluation of model performance at historical variability and long-term changes provides further relevant information (Flato et al., 2014). Trend evaluation may provide very useful insight, but has limitations, in particular at the regional scale, mainly due to multi-decadal internal climate variability (Section 10.3.3.8), observational uncertainty (in both driving reanalysis and local trends; Section 10.2), and the fact that often not all regional forcings are known, and that past trends may be driven by forcings other than those driving future trends (Sections 10.4.1 and 10.6.3).

Increasing resolution (Haarsma et al., 2016) or performing downscaling may be particularly important when it modifies the climate change signal of a lower resolution model in a physically plausible way (Hall, 2014). Improvements may result from a better representation of regional processes, upscale effects, as well as the possibility of a region-specific model tuning (Sørland et al., 2018). For instance, Gula and Peltier (2012) showed that a higher resolution allows for a more realistic simulation of lake-induced precipitation, resulting in a more credible projection of changes in the snow belts of the North American Great Lakes. Similarly, Giorgi et al. (2016) demonstrated that an ensemble of RCMs better represents high-elevation surface heating and in turn increased convective instability. As a result, the summer convective precipitation response was opposite to that simulated by the driving global models (Figure 10.9). Similarly, Walton et al. (2015) showed that a kilometre-scale RCM enables a more realistic representation of the snow-albedo feedback in mountainous terrain compared to standard resolution global models, leading to a more plausible simulation of elevation-dependent warming. Bukovsky et al. (2017) argue that strong seasonal changes in warm-season precipitation in the Southern Great Plains of the USA, projected by RCMs, are more credible than the weaker global model changes because precipitation is better simulated in the RCMs.

Figure 10. 9 | Projected changes in summer (June to August) precipitation (in percent with respect to the mean precipitation) over the Alps between the periods 2070–2099 and 1975–2004. (a) Mean of four global climate models (GCMs) regridded to a common 1.32° × 1.32° grid resolution; (b) mean of six regional climate models (RCMs) driven with these GCMs. The grey isolines show elevation at 200 m intervals of the underlying model data. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11). Figure adapted from Giorgi et al. (2016).

Open figure

Including additional components, feedbacks and drivers can substantially modify the simulated future climate. For example, Kjellström et al. (2005) and Somot et al. (2008) have shown that a regional ESM can significantly modify the SST response to climate change of its driving global model with implications for the climate change signal over both the sea and land. In particular, coupled ocean–atmosphere RCMs may increase the credibility of projections in regions of strong air-sea coupling such as the East Asia–western North Pacific domain (Zou and Zhou, 2016b, 2017). Recent studies demonstrate the importance of including regional patterns of evolving aerosols in RCMs for simulating regional climate change (Boé et al., 2020a; Gutiérrez et al., 2020). RCMs not including the plant physiological response to increasing CO₂ concentrations have been shown to substantially underestimate projected increases in extreme temperatures across Europe compared to global models that explicitly model this effect (Schwingshackl et al., 2019).

A difference between the climate changes simulated by two models does not automatically imply the more complex or higher resolution model is superior (e.g., Dosio et al., 2019). Studies comparing convection-permitting RCM simulations to simulations of climate models with parametrized convection find, depending on the considered models, regions and seasons, either similar or qualitatively different projected changes in short duration extreme precipitation (Chan et al., 2014a, b, 2020; Ban et al., 2015; Tabari et al., 2016; Fosser et al., 2017; Kendon et al., 2017, 2019; Vanden Broucke et al., 2018). Process studies provide evidence that convection-permitting simulations better represent crucial local and mesoscale features of convective storms and thus simulate more plausible changes (Meredith et al., 2015a; Prein et al., 2017; Fitzpatrick et al., 2020), but further research is required to confirm and reconcile the different findings.

Studies assessing the fitness of statistical approaches for regional climate projections are still very limited in number. For statistical downscaling, a key issue is to include predictors that control long-term changes in regional climate. Models differing only in the choice of predictors may perform similarly in the present climate, but may project opposite precipitation changes (Fu et al., 2018; Manzanas et al., 2020). In addition to trend-evaluation studies (Section 10.3.3.8), perfect-model experiments (Section 10.3.2.5) have been used to assess whether a given model structure with a chosen set of predictors is capable of reproducing the simulated future climates (Gutiérrez et al., 2013; Räty et al., 2014; Dayon et al., 2015; Dixon et al., 2016; San-Martín et al., 2017). Importantly, it is found that standard analogue methods inherently underestimate future warming trends because of missing analogues for a warmer climate (Gutiérrez et al., 2013).

Bias adjustment assumes that model biases are time invariant (or more precisely, independent of the climate state), such that the adjustment made to present climate simulations is still applicable to future climate simulations. Many findings challenge the validity of this assumption, as already assessed in AR5 (Flato et al., 2014). Further research has addressed this issue by means of perfect model experiments (Section 10.3.2.5) and process understanding. Perfect-model studies with GCMs found that circulation, energy, and water-cycle biases are roughly state-independent (Krinner and Flanner, 2018), whereas temperature biases depend linearly on temperature (Kerkhoff et al., 2014). Others show that regional temperature biases may depend on soil moisture and albedo, and may thus be state-dependent (Maraun, 2012; Bellprat et al., 2013; Maraun et al., 2017; see Cross-Chapter Box 10.2 for further limitations of bias adjustment). The fitness of weather generators for future projections depends on whether they account for all relevant changes in their parameters, either by predictors or change factors (Maraun and Widmann, 2018b).

In any case, the fitness of regional climate projections based on dynamical downscaling or statistical approaches depends on the fitness of the driving models in projecting boundary conditions, predictors and change factors (Hall, 2014; Maraun and Widmann, 2018b).

Overall, there is high confidence that an assessment of model fitness for projections applying process-based evaluation, process-based plausibility checks of projections and a comparison of different model types, increases the confidence in climate projections. There is high confidence that increasing model resolution, dynamical downscaling, statistical downscaling with well-simulated predictors controlling regional climate change, and adding relevant model components can increase the fitness for projecting some aspects of regional climate when accompanied by a process-understanding analysis.

Global models reproduce many of the features of observed climate and its variability at regional scales. However, global models can show a variety of biases in, for instance, precipitation and temperature at scales ranging from continental (Prasanna, 2016) to sub-continental scales (Lovino et al., 2018), both in the mean and in higher order moments of the climatological distribution of the variable (Figure 10.6; Ren et al., 2019; Xin et al., 2020). Regional biases could occur even if all the relevant large-scale processes are correctly represented, but not their interaction with regional features such as orography or land–sea contrasts (Section 10.3.3.4). These biases have been considered an important limiting factor in model usability, especially at the regional scale (Palmer, 2016). In spite of this, global model simulations have been extensively used to create regional estimates of climate change (Chapters 11, 12 and Atlas), taking into account the result of a performance assessment (Chapter 11, Sections 10.3.3.3–10.3.3.8, and Atlas; Jiang et al., 2020). However, their application is limited in part by the effective resolution of these models (Klaver et al., 2020).

Global model performance at the regional scale is assessed in terms of the time or spatial averages of key variables (see Atlas; Brunner et al., 2019), the ability to reproduce their seasonal cycle (Hasson et al., 2013) or a set of extreme climate indicators (Chapter 11; Luo et al., 2020) and the representation of regional processes and phenomena, feedbacks, drivers and forcing impacts (Sections 10.3.3.4–10.3.3.6). In many cases, the performance estimates have been used to select models for either an application or a more in-depth study (Lovino et al., 2018), to select the models that provide boundary conditions to perform RCM simulations (McSweeney et al., 2015) or to weight the results of the global model simulations (Sanderson et al., 2015; Brunner et al., 2020). While some large-scale metrics are improved between the CMIP5 and CMIP6 experiments (Chapter 3; Cannon, 2020), there is not yet concluding evidence of a systematic improvement for surface variables at the regional scale.

The special class of high-resolution global models (Sections 1.5.3.1 and 10.3.3.1, Chapter 3; Haarsma et al., 2016; Prodhomme et al., 2016) is expected to improve some of the regional processes that are not appropriately represented in standard global models (Roberts et al., 2018). There is general consensus that increasing global model resolution improves some long-standing biases (Chapter 3, Section 10.3.3.3, and Figures 10.6 and 10.7; Demory et al., 2014, 2020; Schiemann et al., 2014; Dawson and Palmer, 2015; van Haren et al., 2015; Feng et al., 2017; Fabiano et al., 2020), although the resolution increase is not a guarantee of overall improvement (Section 8.5.1; Fabiano et al., 2020; Hertwig et al., 2021). For instance, increasing resolution in global models has been shown to improve Asian monsoon rainfall anchored to orography and the monsoon circulation (Johnson et al., 2016), but fails to solve the major dry bias. It is also difficult to disentangle the role of resolution increase and model tuning on the performance of the GCM (Anand et al., 2018). Some efforts have been undertaken to complement the performance improvements of resolution by using stochastic parametrizations (Palmer, 2019), which explicitly acknowledge the multi-scale nature of the climate system, in standard resolution global models with some success (Dawson and Palmer, 2015; MacLeod et al., 2016; Zanna et al., 2017, 2019). The expectation is to achieve a similar performance to the increase in resolution at a reduced computational cost.

Despite their known errors that affect model performance, there is high confidence that global models provide useful information for the production of regional climate information. There is robust evidence and high agreement that the increase of global model resolution helps in reducing the biases limiting performance at the regional scale, although resolution per se does not automatically solve all performance limitations shown by global models. There is robust evidence that stochastic parametrizations can help to improve some aspects of the global model performance that are relevant to regional climate information.

Global models tend to have difficulties in simulating climate over regions where unresolved local scale processes, feedbacks and non-linear scale interactions result in a degradation of the model performance compared to models with higher resolution. In this case, RCMs and variable resolution global models can resolve part of these processes in the regions of interest at an acceptable computational cost (Rummukainen, 2016; Giorgi, 2019; Gutowski Jr. et al., 2020).

The assessment of RCM performance needs to focus not only on mean climatology (Atlas), but also trends (Section 10.3.3.8) and extremes (Chapter 11), and the RCM’s ability at correctly reproducing relevant processes, forcings and feedbacks including aerosols, plant responses to increasing CO₂, and so on, (Schwingshackl et al., 2019; Boé et al., 2020a; Sections 11.2. and 10.3.3.3 to 10.3.3.8) to be fit for future projections (Section 10.3.3.9).

When RCMs are driven by global models, part of the uncertainty in the RCM simulation is introduced by the global model biases (Kjellström et al., 2018; Sørland et al., 2018; Christensen and Kjellström, 2020). As RCMs are typically not able to mitigate global model biases in large-scale dynamical processes, if such biases are substantial, and if the corresponding large-scale processes are important drivers of regional climate, downscaling is questionable (Section 10.3.3.3). However, when global models have weak circulation biases and regional climate change is controlled mainly by regional-scale processes and feedbacks, dynamical downscaling has the potential to add substantial value to global model simulations (Section 10.3.3.4 and Atlas; Hall, 2014; Rummukainen, 2016; Giorgi, 2019; Schwingshackl et al., 2019; Boé et al., 2020a; Lloyd et al., 2021).

There is very high confidence (robust evidence and high agreement) that RCMs add value to global simulations in representing many regional weather and climate phenomena, especially over regions of complex orography or with heterogeneous surface characteristics and for local-scale phenomena. Realistically representing local-scale phenomena such as land–sea breezes requires simulations at a resolution of the order of 10 km (high confidence). Simulations at kilometre-scale resolution add value in particular to the representation of convection, sub-daily summer precipitation extremes (high confidence) and soil-moisture–precipitation feedbacks (medium confidence). Resolving regional processes may be required to correctly represent the sign of regional climate change (medium confidence). However, the performance of RCMs and their fitness for future projections depend on their representation of relevant processes, forcings and drivers in the specific context (Sections 10.3.3.4–10.3.3.8).

Statistical downscaling, bias adjustment and weather generators outperform uncorrected output of global and regional models for a range of statistical aspects at single locations due to their calibration (Casanueva et al., 2016), but RCMs are superior when spatial fields are relevant (Mehrotra et al., 2014; Vaittinada Ayar et al., 2016; Maraun et al., 2019a). Similarly, there is some evidence that bias adjustment is comparable in performance when applied to global models and dynamically downscaled global models only for single locations, but dynamical downscaling prior to bias adjustment clearly adds value once spatial dependence is relevant (Maraun et al., 2019a). These results may explain why dynamical downscaling does not add value to global model simulations for (single-site) agricultural modelling, when both global and regional models are bias adjusted (Glotter et al., 2014), but dynamical downscaling adds value compared to bias-adjusted global model output for spatially distributed hydrological models (Qiao et al., 2014).

Overall, statistical downscaling methods with carefully chosen predictors and an appropriate model structure for a given application realistically represent many statistical aspects of present-day daily temperature and precipitation (high confidence, Section 10.3.3.7). Bias adjustment has proven beneficial as an interface between climate model projections and impact modelling in many different contexts (high confidence) (Section 10.3.3.7). Weather generators realistically simulate many statistical aspects of present-day daily temperature and precipitation (high confidence) (Section 10.3.3.7). The performance of these approaches and their fitness for future projections also depends on predictors and change factors taken from the driving dynamical models (high confidence) (Section 10.3.3.9).

Modelling chains for generating regional climate information range from the definition of forcing scenarios to the global modelling, and potentially to dynamical or statistical downscaling and bias adjustment (Section 10.3.1). The propagation and potential accumulation of uncertainties along the chain has been termed the cascade of uncertainty (Wilby and Dessai, 2010). Even within one model, like a global model, uncertainty propagates across scales. From a process point of view, these uncertainties are related to forcings and global climate sensitivity, and errors in the representation of the large-scale circulation (Section 10.3.3.3; McNeall et al., 2016) and regional processes (Section 10.3.3.4), feedbacks (Section 10.3.3.5) and drivers (Section 10.3.3.6). From a modelling point of view, these uncertainties are related to the choice of dynamical and statistical models (Section 10.3.1) and experimental design (Section 10.3.2). The overall uncertainty can be statistically decomposed into the individual sources (Evin et al., 2019; Christensen and Kjellström, 2020), although there might be non-linear dependencies between them.

Uncertainty propagation often increases the spread in regional climate projections when comparing global model and downscaled results, which has been used as an argument against top-down approaches to climate information (Prudhomme et al., 2010). Increased spread in the modelling chain may also arise from a more comprehensive representation of previously unknown or underrepresented uncertainties (Maraun and Widmann, 2018b). The increased spread in this case goes together with a better representation of processes and thus an increased model fitness-for-purpose (Section 10.3.3.9).

Climate response uncertainties (Chapter 1) can be represented by multi-model ensembles, although the sampled uncertainty typically underestimates the full range of uncertainty (Collins et al., 2013b; Shepherd et al., 2018; Almazroui et al., 2021). Traditionally, climate response uncertainty has been characterized by the ensemble spread around the multi-model mean change. The change has then further been qualified in terms of the agreement across models and compared to estimates of internal climate variability (Collins et al., 2013b). Since AR5, several limitations of this approach have been identified (Madsen et al., 2017) such as the failure to address physically plausible, but low-likelihood, high-impact scenarios (Chapters 1, 4, 8 and 9; Sutton, 2018) or that qualitatively different or even opposite changes may be equally plausible at the regional scale (Shepherd, 2014). In a multi-model mean these different responses would be lumped together, strongly dampened, and qualified as non-robust, whereas in fact high impacts might occur. Further, the multi-model mean itself is often implausible because it is a statistical construct (Zappa and Shepherd, 2017). Overall, there is high confidence that some regional future climate changes are not well-characterized by multi-model mean and spread.

Since AR5, physical climate storyline approaches (see also Chapter 1, Section 10.5.3, Box 10.2, and Atlas.2.5.2) have been developed to better characterize and communicate uncertainties in regional climate projections (Shepherd, 2019). A special class of such storylines attempts to attribute regional uncertainties to uncertainties in remote drivers. For instance, the Dutch Meteorological Service has presented climate projections for the Netherlands for different plausible changes of the mid-latitude atmospheric circulation and different levels of European warming (van den Hurk et al., 2014). Manzini et al. (2014) have quantified the impact of uncertainties in tropical upper troposphere warming, polar amplification, and stratospheric wind change on Northern Hemisphere winter climate change. Based on these results, Zappa and Shepherd (2017) separated the multi-model ensemble into physically consistent sub-groups or storylines of qualitatively different projections in relevant remote drivers of the atmospheric circulation. In a similar vein, (Ose et al., 2020) trace uncertainties in projections of the East Asian summer monsoon and Mindlin et al. (2020) conditioned the response of Southern Hemisphere mid-latitude circulation and precipitation to greenhouse gas forcing on large-scale climate indicators (Section 8.4.2.9.2).

These physical climate storylines help to physically explain contradicting regional projections and thus make the conveyed information a better representation of the true uncertainty (Hewitson et al., 2014a). Additionally, the attribution of regional uncertainties to drivers may in principle help reduce uncertainty in the case where some storylines can be ruled out because the projected changes in the driving processes appear to be physically implausible (Zappa and Shepherd, 2017). There is thus high confidence that storylines attributing uncertainties in regional projections to uncertainties in changes of remote drivers aid the interpretation of uncertainties in climate projections.

Another approach that has continued to develop for characterising and reducing projection uncertainties is the use of emergent constraints (Chapters 1, 4, 5 and 7; Hall et al., 2019). The idea is to link the spread in climate model projections via regression to the spread in present climate model biases for relevant driving processes. Models with lower biases are assigned higher weight in the projections, which in turn reduces the spread of the projections in a physical way and may additionally reduce projection uncertainty. For instance, Simpson et al. (2016) have reduced the spread in projections of North American winter hydroclimate by linking this spread to model biases in the representation of relevant stationary wave patterns. Other examples of using emergent constraints in a regional context are Brown et al. (2016), G. Li et al. (2017), Giannini and Kaplan (2019), Ose (2019) and Zhou et al. (2019).

A regional climate projection based on a single simulation from a single global model or driving a single RCM alone will inevitably be affected by not considering the internal variability (Figure 10.10). This is mainly due to the dominant influence of the chaotic atmospheric circulation on regional climate variability, in particular at mid- to high latitudes. Internal variability is an irreducible source of uncertainty for mid- to long-term projections with an amplitude that typically decreases with increasing spatial scale and lead time (Sections 1.4.3 and 4.2.1). However, regional-scale studies show that both large- and local-scale internal variability together can still represent a substantial fraction of the total uncertainty related to hydrological cycle variables, even at the end of the 21st century (Lafaysse et al., 2014; Vidal et al., 2016; Aalbers et al., 2018; Gu et al., 2018).

Figure 10.10 | Observed and projected changes in austral summer (December to February) mean precipitation in Global Precipitation Climatoloy Centre (GPCC), Climatic Research Unit Time Series (CRU TS) and 100 members of the Max Planck Institute for Meteorology Earth System Model (MPI-ESM. (a) 55-year trends (2015–2070) from the ensemble members with the lowest (left) and highest (right) trend (% per decade, baseline 1995–2014). (b) Time series (%, baseline 1995–2014) for different spatial scales (from top to bottom: global averages; South-Eastern South America; grid boxes close to São Paulo and Buenos Aires) with a five-point weighted running mean applied (a variant on the binomial filter with weights [1-3-4-3-1]). The brown (green) lines correspond to the ensemble member with weakest (strongest) 55-year trend and the grey lines to all remaining ensemble members. Box-and-whisker plots show the distribution of 55-year linear trends across all ensemble members, and follow the methodology used in Figure 10.6. Trends are estimated using ordinary least squares. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

Analysis of multi-model archives such as CMIP or CORDEX simulation results cannot easily disentangle model uncertainty and uncertainty related to internal variability. Since AR5, the development of single-model (global model and/or RCM) initial-condition large ensembles (SMILEs) has emerged as a promising way to robustly assess the regional-scale forced response to external forcings and the respective contribution of internal variability and model uncertainty to future regional climate changes (Section 4.2.5; Deser et al., 2014, 2020; Kay et al., 2015; Sigmond and Fyfe, 2016; Aalbers et al., 2018; Bengtsson and Hodges, 2019; Dai and Bloecker, 2019; Leduc et al., 2019; Maher et al., 2019; von Trentini et al., 2019; Lehner et al., 2020). The recent development of a multi-model archive of SMILE simulations facilitates the quantification and comparison of the influence of internal variability on global model-based regional climate projections between different models (Deser et al., 2020; Lehner et al., 2020). Another related development is the more frequent use of observation-based statistical models to assess the influence of internal variability on regional-scale global and regional model projections (Thompson et al., 2015; Salazar et al., 2016). However, these methods often implicitly assume that regional-scale internal variability does not change under anthropogenic forcing, which is a strong assumption that does not seem to hold at regional and local scales (LaJoie and DelSole, 2016; Pendergrass et al., 2017; W. Cai et al., 2018; Dai and Bloecker, 2019; Mankin et al., 2020; Milinski et al., 2020).

The appropriate ensemble size for a robust use of SMILEs depends on the model and physical variable being investigated, the spatial and time aggregation being performed, the magnitude of the acceptable error and the type of questions one seeks to answer (Deser et al., 2012, 2017b; Kang et al., 2013; Wettstein and Deser, 2014; Dai and Bloecker, 2019; Maher et al., 2019). It is noteworthy that the recent development of ensembles with a very large ensemble size (greater than 100) have led to new insights and methodologies to robustly assess the required ensemble size for questions such as the estimation of the forced response to external forcing or a forced change in modes of internal variability, such as ENSO, and its associated teleconnections (Herein et al., 2017; Maher et al., 2018; Haszpra et al., 2020; Milinski et al., 2020).

The use of SMILEs assumes that they have a realistic representation of internal variability and its evolution under anthropogenic climate change (Eade et al., 2014; McKinnon et al., 2017; McKinnon and Deser, 2018; Chen and Brissette, 2019). Assessing the realism of simulated internal variability for past and current climates remains an active research field with a number of issues such as the shortness and uncertainties of the observed record, in particular in data-scarce regions (Section 10.2.2.3), the signal-to-noise paradox (Section 4.4.3.1; Scaife and Smith, 2018), uncertainty in past observed external forcing estimates (Chapters 2, 6 and 7) and the limitations of assumptions underlying the statistical methods used to derive observational large ensembles (McKinnon et al., 2017; McKinnon and Deser, 2018; Castruccio et al., 2019). Calibration methods inspired by weather and seasonal forecasts can be used to improve the reliability of regional-scale climate projections from large ensembles (Brunner et al., 2019; O’Reilly et al., 2020). Interestingly, reliability is improved when the calibration is performed separately for the dynamical and residual components of the ensemble resulting from dynamical adjustment (Section 10.4.1; O’Reilly et al., 2020).

Importantly, accurately partitioning uncertainty in regional climate projections can provide an incentive for immediate action, accepting a large range of possible outcomes due to internal variability, while confounding model uncertainty with internal variability may be understood as a lack of knowledge and lead to delayed action in adaptation decision-making (Section 10.5.3; Maraun, 2013b; Mankin et al., 2020).

There is high confidence that the availability of SMILEs allows a robust assessment of the relative contributions of model uncertainty and internal variability in regional-scale projection uncertainty. There is high confidence that the use of SMILEs with appropriate ensemble size leads to an improved estimate of regional-scale forced response to an external forcing as well as of the full spectrum of possible changes associated with internal variability. There is high confidence that these improved estimates are beneficial for characterizing the full distribution of outcomes that is a key ingredient of climate information for robust decision-making and risk-analysis frameworks.

Ensembles of climate simulations play an important role in quantifying uncertainties in the simulation output (Sections 10.3.4.2 and 10.3.4.3). In addition to providing information on internal variability, ensembles of simulations can estimate scenario uncertainty and model (structural) uncertainty. Chapter 4, especially Box 4.1, discusses issues involved with evaluating ensembles of global model simulations and their uncertainties. In a downscaling context, further considerations are necessary, such as the selection of global model–RCM combinations when performing dynamical downscaling. This is a relevant issue when resources are limited. The structural uncertainty of both the global model and the downscaling method can be important (e.g., Mearnset al., 2012; Dosio, 2017), as well as further potential uncertainty created by inconsistencies between the global model and the downscaling method (e.g., Dosio et al., 2019), which could include, for example, differences in topography or the way to model precipitation processes (Mearns et al., 2013).

An important consideration is which set of global models should be used for global model–RCM combinations. If adequate resources exist, then large numbers of global model–RCM combinations are possible (Déqué et al., 2012; Coppola et al., 2021; Vautard et al., 2021). However, coordinated downscaling programmes can be limited by the human and computational resources available, for producing ensembles of downscaled output, which limits the number of feasible global model–RCM combinations. With this limitation in mind, a small set of GCMs may be chosen that span the range of equilibrium climate sensitivity in available global models (e.g., Mearns et al., 2012, 2013; Inatsu et al., 2015), though this range may be inconsistent with the likely range (Chapter 4), or some other relevant measure of sensitivity, such as the projected range of tropical SSTs (Suzuki-Parker et al., 2018). A further choice is to emphasize models that do not have the same origins or that do not use similar parametrizations and thus might be viewed as independent, a criterion that could be applied to both global models (Chapter 4) and RCMs (Evans et al., 2014). Global models and RCMs could also be discarded that unrealistically represent processes controlling the regional climate of interest (McSweeney et al., 2015; Maraun et al., 2017; Bukovsky et al., 2019; Eyring et al., 2019). Box 4.1 offers a more detailed discussion of the issues surrounding these approaches. Finally, global models may be selected to represent different physically self-consistent changes in regional climate (Zappa and Shepherd, 2017). Statistical methods can provide estimates of outcomes from missing global model–RCM combinations in a large matrix (Déqué et al., 2012; Heinrich et al., 2014; Evin et al., 2019).

However, even using a relatively small set of global models can still involve substantial computation that strains available resources, both for performing the simulations and for using all simulations in the ensemble for further impacts assessment. The NARCCAP programme (Mearns et al., 2012) used only a subset of its possible global model–RCM combinations that balanced comprehensiveness of sampling the matrix with economy of computation demand, while still allowing discrimination, via ANOVA methods, of global model and RCM influences on regional climate change (Mearns et al., 2013). An advantage of the sparse, but balanced matrix for those using the downscaling output for further studies, is that they have a smaller, yet comprehensive set of global model–RCM combinations to work with. Alternatively, data-clustering methods can clump together downscaling simulations featuring similar climate-change characteristics, so that only one representative simulation from each cluster may be needed for further impacts analysis, again systematically reducing the necessary number of simulations to work with (Mendlik andGobiet, 2016; Wilcke and Bärring, 2016).

Independently of the resources, participation of multiple models in a simulation programme such as CORDEX for RCMs or CMIP for global models creates ensembles of opportunity, which are ensembles populated by models that participants chose to use without there necessarily being an overarching guiding principle for an optimum choice. As discussed in Chapter 4, these ensembles are likely suboptimal for assessing sources of uncertainty. An important contributor to the suboptimal character of such an ensemble is that the models are not independent. Some may also have larger biases than others. Yet often, the output from models in these ensembles has received equal weight when viewed collectively, as was the case in much of the AR5 assessment (e.g., Collins et al., 2013b; Knutti et al., 2013; Flato et al., 2014; Kirtman et al., 2014). A number of emerging methodologies aim at optimizing the ensembles available by weighting the simulation results according to a number of criteria relevant at the regional scale that aim at obtaining more realistic estimates of the uncertainty (Sanderson et al., 2015; Brunner et al., 2020).

There is high confidence that ensembles for regional climate projections should be selected such that models unrealistically simulating processes relevant for a given application are discarded, but at the same time, the chosen ensemble spans an appropriate range of projection uncertainties.

Optimal fingerprint regression-based methods have been applied to detection and attribution of mean temperature anthropogenic signal in several regions of the world such as Canada, India, central Asia, northern and western China, Australia, and North Africa (Xu et al., 2015; C. Li et al., 2017; Dileepkumar et al., 2018; Y. Wang et al., 2018; Peng et al., 2019; Wan et al., 2019). The influence of anthropogenic forcing, and in particular that of greenhouse gases (GHGs), is robustly detected in annual and seasonal mean temperatures for all considered regions. Most of the observed regional temperature changes since the mid-twentieth century can only be explained by external forcings, with anthropogenic influence being the dominant factor. GHG increase is found to be the primary factor of the anthropogenic-induced warming while the aerosol forcing leads to a cooling offsetting a fraction of the GHG change (C. Li et al., 2016, 2017). While the influence of external natural forcing can often be detected as well, its contribution to observed changes is usually much smaller (C. Li et al., 2017; Wan et al., 2019). Temperature detection results are found to be robust to the use of different observational datasets and detection methodologies (Dileepkumar et al., 2018).

Detection of mean precipitation changes caused by human influence is much more difficult, due to a larger role of internal variability at regional to local scales, as well as substantial modelling and observational uncertainty (Wan et al., 2015; Sarojini et al., 2016; C. Li et al., 2017). However, multi-decadal precipitation changes due to anthropogenic forcing have been detected for several regions. Ma et al. (2017b) show that anthropogenic forcing has strongly contributed to the observed shift of China daily precipitation towards heavy precipitation. The observed weakening of the East Asia summer monsoon, also known as the southern flooding and northern drought pattern has been partially linked to anthropogenic forcing (Section 8.3.2.4.2; Song et al., 2014; Zhou et al., 2017; Tian et al., 2018). Changes in GHGs lead to increasing precipitation over southern China, while changes in anthropogenic aerosols over East Asia are the dominant factors determining drought conditions over northern China (Song et al., 2014; Tian et al., 2018). Based on all-forcing and single-forcing simulation ensembles with a high-resolution model, Delworth and Zeng (2014) found that the observed long-term regional austral autumn and winter rainfall decline over southern and particularly south-west Australia is partially reproduced in response to anthropogenic changes in GHGs and ozone in the atmosphere, whereas anthropogenic aerosols do not contribute to the simulated precipitation decline. In contrast, the observed increase of north-west Australian summer rainfall since 1950 has been partially attributed to anthropogenic aerosol based on CMIP5 detection and attribution single-forcing simulations (Section 8.3.2.4.6; Dey et al., 2019a, b).

It is noteworthy that these methods require a very significant reduction of spatial and temporal dimensions in order to reliably estimate the co-variance matrix of internal variability (an entire region is thus often considered as being only one or a few spatial points that represent the spatial average of the whole region or a few sub-regions; time samples are often 5- or 10-year averages). Finally, model bias is rarely considered in statistical models used in detection and attribution regional studies, while it has been shown to have a strong impact on the stability of detection results and their associated confidence intervals when increasing the spatial dimension (Ribes and Terray, 2013). New statistical methods are emerging to provide some alternative to standard optimal fingerprinting but they have not yet been evaluated and applied at regional scales (Section 3.2.2).

The primary objective of any attribution method is to optimally separate the influences of external forcing and internal variability on a global or regional climate record. In a multi-model ensemble context, the estimation of the externally-forced climate response has been typically performed by ensemble averaging of linear trends or regional domain spatial average, thus not taking into account the available and complete space and time co-variance information. Since AR5, methods using spatiotemporal information have been further developed and used to improve the separation between external and internal drivers in multiple or single historical climate realizations performed by a given global model.

The typical ensemble size of CMIP historical climate simulations for a given model traditionally range between one and ten members, with three often being the default choice. At the regional scale, a simple ensemble average with such sample sizes does not provide robust estimates of the response patterns to external forcing (Maher et al., 2019; Deser et al., 2020). Since AR5, pattern filtering methods such as signal-to-noise maximizing empirical orthogonal functions (Ting et al., 2009) have been shown to improve the identification of forced response patterns when few model members are available (Wills et al., 2020). Using SMILEs as a test bed, it has been shown that pattern filtering strongly reduces the number of ensemble members needed to estimate the forced response pattern compared to simple ensemble averaging. Pattern filtering allows the identification of low signal-to-noise signals such as the El Niño-like response to volcanic eruptions (Khodri et al., 2017; Wills et al., 2020).

Methods to extract the response to external forcing in an observed or simulated single realization include dynamical adjustment (Smoliak et al., 2015; Deser et al., 2016; Sippel et al., 2019) and time scale separation methods (DelSole et al., 2011; Wills et al., 2018, 2020). Dynamical adjustment seeks to isolate changes in surface air temperature or precipitation that are due purely to atmospheric circulation changes. The residual can then be analysed and attributed to internal changes in both land or ocean surface conditions and the thermodynamical response to external forcing. Smoliak et al. (2015) performed their dynamical adjustment using partial least squares regression of temperature to remove variations arising from sea level pressure changes. Deser et al. (2016) used constructed atmospheric circulation analogues and resampling to estimate the dynamical contribution to changes in temperature. Sippel et al. (2019) used machine learning techniques known as regularized linear regression to provide estimates of circulation-induced components of precipitation and temperature variability from global to local scales. It is noteworthy that the dynamical adjustment method by itself cannot account for the component of the forced response associated with circulation changes that project onto atmospheric internal variability. However, this component can be estimated within a model framework by averaging the dynamical contribution across multiple members of a SMILE (Deser et al., 2016).

Dynamical adjustment methods have been used by, for instance, Deser et al. (2016), Saffioti et al. (2016), O’Reilly et al. (2017), Gong et al. (2019), and R. Guo et al. (2019). Deser et al. (2016) focused on the causes of observed and simulated multi-decadal trends in North American temperature. They demonstrated that the main advantage of this technique is to narrow the spread of temperature trends found by the model ensemble and to bring the dynamically-adjusted observational trend much closer to the forced response estimated by the model ensemble mean. Similar results were obtained by Saffioti et al. (2016) regarding recent observed winter temperature and precipitation trends over Europe. Similarly, O’Reilly et al. (2017) applied dynamical adjustment techniques to more carefully determine the influence of the Atlantic Multi-decadal Variability (AMV; Annex IV.2.7) on continental climates. Over Europe, summer temperature anomalies induced thermodynamically by the warm phase of the AMV are further reinforced by circulation anomalies; meanwhile, precipitation signals are largely controlled by dynamical responses to the AMV. Based on a partial least-squares approach, Gong et al. (2019) showed that recent winter temperature 30-year trends over northern East Asia are strongly influenced by internal variability linked to decadal changes of the Arctic Oscillation. Using dynamical adjustment purely on precipitation observations, R. Guo et al. (2019) showed that human influence has led to increased winter precipitation across north-eastern North America, as well as a small region of north-western North America, and to an increase in precipitation across much of north-western and north central Eurasia. The latter results confirm previous findings obtained by standard optimal fingerprinting methods (Wan et al., 2015).

Time scale separation methods such as the low-frequency component analysis and ensemble empirical mode decomposition methods take advantage of the longer time scale associated with anthropogenic external forcing compared to that of most internal modes of variability. The low-frequency component analysis method tries to find low-frequency variability patterns by searching for linear combinations of a moderate number of empirical orthogonal functions that maximize the ratio of low-frequency to total variance. It has first been used to separate internal modes of interannual and decadal variability from slowly varying and externally-forced variability in the Pacific and Atlantic oceans (Wills et al., 2018, 2019). The methodology has also been applied to patterns of observed surface air temperature to isolate the slow components of observed changes that are consistent with the expected response to anthropogenic greenhouse gas and aerosol forcing (Wills et al., 2020).

The ensemble empirical mode decomposition method (Wu and Huang, 2009; Wilcox et al., 2013; Ji et al., 2014; Qian and Zhou, 2014) decomposes data, such as time series of historical temperature and precipitation, into independent oscillatory modes of decreasing frequency. The last step of the method leaves behind a smooth and low-frequency residual time series. Typically, the non-linear anthropogenic trend (e.g., of 20th-century temperature) can be reconstructed by summing the long-term mean, the residual, and eventually the lowest-frequency mode to account for a multi-decadal forced signal, for instance associated with anthropogenic aerosol forcing. The ensemble empirical mode decomposition method is an example of a data-driven, non-parametric approach that can be used to directly provide an estimate of the forced response without the need for model data (Qian, 2016).

The univariate detection method does not use spatial pattern information, but compares observed trends in gridded datasets with distributions of trends from ensembles of simulations during the historical period (Knutson et al., 2013; Knutson and Zeng, 2018). The trends arising from simulations constrained by natural forcing-only and all-forcing are compared with distributions of trends purely due to internal variability and derived from long simulations with constant pre-industrial external forcing. Consistency between observed and simulated historical trends is also assessed with statistical tests that can be applied independently over a large number of grid points. The fraction of area over a given region where the change is classified as detectable, attributable, or consistent/inconsistent, is then finally estimated. The method can be viewed as a simple consistency test for both amplitude and pattern of observed versus simulated trends. Its application to CMIP3 and CMIP5 models suggests that 80% of the Earth’s surface has a detectable anthropogenic warming signal (Knutson et al., 2013). Regarding regional land precipitation changes over the 1901–2010 and 1951–2010 periods, application of the univariate detection method based on CMIP5 models suggests attributable anthropogenic changes at several locations such as increases over regions of the north-central USA, southern Canada, Europe, and southern South America and decreases over parts of the Mediterranean region, northern tropical Africa and south-western Australia (Delworth and Zeng, 2014; Knutson and Zeng, 2018).

Another regional attribution technique is based on the similarity of past changes between observations and one or several simulations of a large ensemble that share the same time evolution for a suggested driver of these changes. Huang et al. (2020b) used a perturbed physics ensemble to attribute the drying trend of the Indian monsoon over the latter half of the 20th century to decadal forcing from the Pacific Decadal Variability (PDV; Annex IV.2.6). The ensemble members predicted different trends in PDV behaviour across the 20th century and the negative precipitation trend was only replicated in those members with a strong negative-to-positive PDV transition across the 1970s, consistent with the observed PDV behaviour (see also the detailed case study in Section 10.6.3). In a similar manner, Cvijanovic et al. (2017) addressed the possible influence of Arctic sea ice loss on the North Pacific pressure ridge and, consequently, on south-western USA precipitation. They sampled the uncertainties in selected sea ice physics parameters to achieve a ‘low Arctic sea ice’ state in their perturbed simulations. They then compared the latter with control simulations representative of sea ice conditions at the end of the 20th century to assess changes purely due to sea ice loss.

New methods aiming to remove underlying model biases before performing detection and attribution, for instance related to precipitation changes, are emerging based on image transformation techniques such as warping (Levy et al., 2014a). By correcting location and seasonal precipitation biases in CMIP5 models, Levy et al. (2014b) showed that the agreement between observed and fingerprint patterns can be improved, further enhancing the ability to attribute observed precipitation changes to external forcings. The improvement mainly relies on the assumption that precipitation changes are tied to the underlying climatology, which has been shown to be a reasonable assumption in regions of the world where intensification of the hydrological cycle is expected (Held and Soden, 2006).

Importantly, evidence that the models employed in regional-scale attribution are fit for purpose is essential in order to estimate the degree of confidence in the attribution results (Section 10.3.3). For example, models need to be evaluated and assessed in their ability to simulate internal variability modes that are known to be important drivers of regional climate change (Sections 3.7 and 10.3.3.3 and Annexes IV.2 and IV.3). Models are likely to have different performance in different regions and therefore their evaluation needs to be performed in terms of key physical processes and mechanisms relevant to the climate of the region under consideration (Section 10.3.3).

To conclude, there is very high confidence (robust evidence and high agreement) that the use of diverse and independent attribution methods, multiple model ensemble types and observed datasets strengthens the robustness of results of regional-scale attribution studies. Since AR5, multiple SMILEs have provided an adequate testbed for new attribution methodologies aimed at separating forced signals from internal variability in observational records as well as small-size single-model ensembles.

The Sahel, fed by the West African monsoon, has experienced severe decadal rainfall variations (Figure 10.11a). Abundant rainfall in the 1950s–1960s was followed by a large negative trend (Figure 10.11b) until at least the 1980s, over which annual rainfall fell by 20–30% (Hulme, 2001). The subsequent partial recovery (B. Wang et al., 2021) is more uncertain: rain-gauge studies suggest a return to long-term positive anomalies in the western Sahel in the early 2000s (Panthou et al., 2018), while CHIRPS merged satellite/gauge data show a wetter western Sahel since 1981 (Bichet and Diedhiou, 2018a, b). The recovery has been more significant over the central rather than the western Sahel (Lebel and Ali, 2009; Maidment et al., 2015; Sanogo et al., 2015) and a multiple-gauge record supports a greater recovery to the eastern side (Nicholson et al., 2018). In this attribution example, drivers of the long-term drought and subsequent partial recovery are discussed, including anthropogenic GHG and aerosol emissions, and sea surface temperature (SST) variations that, in part, relate to internal variability. The reader is also referred to assessment in Section 8.3.2.4. We define the Sahel within 10°N–20°N across to 30°E, consistent with the eastern boundary used in Chapter 8, and the rainy season as spanning June to September.

Figure 10.11 | Attribution of historic precipitation change in the Sahelian West African monsoon during June to September. (a) Time series of CRU TS precipitation anomalies (mm day^–1, baseline 1955–1984) in the Sahel box (10°N–20°N, 20°W–30°E) indicated in panel (b) applying the same low-pass filter as that used in Figure 10.10. The two periods used for difference diagnostics are shown in grey columns. (b) Precipitation change (mm day^–1) in CRU TS data for 1980–1990 minus 1950–1960 periods. (c) Precipitation difference (mm day^–1) between 1.5× and 0.2× historical aerosol emissions scaling factors averaged over 1955–1984 and five ensemble members of HadGEM3 experiments after Shonk et al. (2020). (d) Sahel precipitation anomaly time series (mm day^–1, baseline 1955–1984) in Coupled Model Intercomparison Project Phase 6 (CMIP6) for 49 historical simulations with all forcings (red), and thirteen for each of greenhouse gas-only forcing (light blue) and aerosol-only forcing (grey), with a thirteen-point weighted running mean applied (a variant on the binomial filter with weights [1-6-19-42-71-96-106-96-71-42-19-6-1]). The CMIP6 subsample of all forcings matching the individual forcing simulations is also shown (pink). (e) Precipitation linear trend (% per decade) for (left) decline (1955–1984) and (right) recovery periods (1985–2014) for ensemble means and individual CMIP6 historical experiments (including single-forcing) as in panel (d) plus 34 CMIP5 models (dark blue). Box-and-whisker plots show the trend distribution of the three coupled and the d4PDF atmosphere-only single-model initial-condition large ensembles (SMILEs) used throughout (Chapter 10 and follow the methodology used in Figure 10.6. The two black crosses represent observational estimates from GPCC and CRU TS. Trends are estimated using ordinary least-squares regression. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

The role of SST forcing in the rainfall decline is assessed first. Competing mechanisms from equatorial Atlantic SSTs and inter-hemispheric SST gradients regulate decadal variability in the Sahel (Nicholson, 2013), alternatively explained by tropical warming leading to Sahel drought, while North Atlantic warming promotes increased rainfall (Rodríguez-Fonseca et al., 2015). The SST influence has been formalized in an AMV framework (Giannini et al., 2013; Martin and Thorncroft, 2014; Martin et al., 2014; Park et al., 2015), suggesting that relative North Atlantic SST warming increases the Northern Hemisphere differential warming, enhancing Sahel rainfall. The AMV influence is supported by CMIP5 initialized decadal hindcasts (Gaetani and Mohino, 2013; Mohino et al., 2016; Sheen et al., 2017), which outperform empirical predictions based on persistence. Some caution is needed since the full magnitude of internal variability is not captured in most CMIP5 models, as poor resolution prevents reproduction of AMV teleconnection responses (Vellinga et al., 2016), and the magnitude of AMV-related SST variation may be underestimated in CMIP5 (Section 3.7.7, which also assesses that the AMV may be partially forced). The influence of PDV has been studied to a lesser extent, with the PDV positive phase having a negative impact on Sahel rainfall in combined observational/CMIP5 analysis (Villamayor and Mohino, 2015). The closer match between the observed rainfall declining trend and those in an atmosphere-only SMILE, in which SSTs are matched to observations, compared to three coupled SMILEs in which they are not, suggests that the underlying ocean surface might be essential in driving the decline (Figure 10.11e).

In terms of anthropogenic emissions, regional aerosol emissions from Europe, and to a lesser extent from Asia, have been shown in a global model to weaken Sahel precipitation either through a weakened Saharan heat low or via the Walker circulation (Dong et al., 2014). Greenhouse gases (GHGs) and anthropogenic aerosol can be considered together to control ITCZ position based on temperature asymmetry at the hemispheric scale. GHGs increase Sahel precipitation, while aerosol reduces it (in coupled slab-ocean model experiments by Ackerley et al. (2011) following Biasutti and Giannini (2006)). This effect is stronger when models account for aerosol–cloud interactions (Allen et al., 2015). Perturbed physics GCM ensembles suggests that aerosol emissions were the main driver of observed drying over 1950–1980 (Ackerley et al., 2011), supported by CMIP5 single-forcing experiments (Polson et al., 2014). A coherent drying signal in CMIP5 over the extended 1901–2010 period has also been found, although smaller than the observed trend (Knutson and Zeng, 2018). By applying aerosol scaling factors to the historical period in order to sample the uncertainty in CMIP5 aerosol radiative forcing, Shonk et al. (2020) found differences of 0.5 mm day^–1 for Gulf of Guinea rainfall between strong and weak aerosol experiments as illustrated in Figure 10.11c, although the drying appears further south than observed due to model bias.

For the partial recovery in West African monsoon and Sahel rainfall since the late 1980s, a detection study using three reanalyses (Cook and Vizy, 2015) shows a connection to increasing Saharan temperatures at a rate two to four times greater than the tropical mean, also confirmed by multiple observational and satellite-based data (Zhou and Wang, 2016; Vizy and Cook, 2017) and the review of Cook and Vizy (2019). Reanalyses are also noted to significantly underestimate the Saharan warming (Zhou and Wang, 2016). Saharan warming causes a stronger thermal low and more intense monsoon flow, providing more moisture to the central and eastern Sahel, supported by CMIP5 models (Lavaysse et al., 2016), although not all models capture the observed rainfall–heat–low relationship. Sahel rainfall is also incorrectly located in prototype versions of a few CMIP6 models, related to tropospheric temperature biases (Martin et al., 2017). Amplified Saharan warming has increased the wind shear, leading to a tripling of extreme storms since 1982, which may partially explain the recovery (Taylor et al., 2017). Instead, observations, multiple models and SST-sensitivity experiments with AGCMs have suggested that stronger Mediterranean Sea evaporation enhances low-level moisture convergence to the Sahel, increasing rainfall (Park et al., 2016). Meanwhile, an AGCM study suggested that GHGs alone (in the absence of SST warming) could cause Sahel rainfall recovery, with an additional role for anthropogenic aerosol (Dong and Sutton, 2015); recent changes in North Atlantic SSTs, although substantial, did not exert a significant impact on the recovery. Large spread in the recovery in a five-member AGCM ensemble suggests that atmospheric internal variability cannot be discounted (Roehrig et al., 2013).

Consistent timing of the southward ITCZ shift during the decline period in CMIP3 and CMIP5 historical simulations supports the role of external forcing, chiefly anthropogenic aerosol (Hwang et al., 2013). The evolution of the observed decline and recovery is largely followed by the CMIP5 multi-model mean, further supporting the role of external drivers (Giannini and Kaplan, 2019). Updated results from CMIP6 for historical simulations with all and single forcings are represented in Figure 10.11d,e showing smaller trends than those observed. Giannini and Kaplan (2019) attempted to unify the driving mechanisms for decline and recovery based on singular-value decomposition of observed and modelled SSTs. Since the 1950s, tropical warming arising from GHGs and North Atlantic cooling from aerosol led to regional stabilization, suppressing Sahel rainfall. The subsequent reduction in aerosol emissions then led to North Atlantic warming and recovery of Sahel rainfall. Such mechanisms continue into the near-term future in idealized and modified RCP experiments, with scenarios featuring more aggressive reductions in aerosol emissions, or including aerosol–cloud interactions, favouring a greater northward shift of rainfall (Allen, 2015; Westervelt et al., 2017, 2018; Scannell et al., 2019). There is paleoclimate evidence of changes to Sahel rainfall in the past, in particular with enhancement of the West African monsoon during the mid-Holocene. However, the mechanisms governing such a change have been shown to be largely dynamical in nature (D’Agostino et al., 2019), suggesting that the mid-Holocene cannot be used to inform the credibility of changes due to greenhouse warming.

There is very high confidence (robust evidence and high agreement) that patterns of 20th-century ocean and land surface temperature variability have caused the Sahel drought and subsequent recovery by adjusting meridional gradients. There is high confidence (robust evidence and medium agreement) that the changing temperature gradients that perturb the West African monsoon and Sahel rainfall are themselves driven by anthropogenic emissions: warming by GHG emissions was initially restricted to the tropics but suppressed in the North Atlantic due to nearby emissions of sulphate aerosols, leading to a reduction in rainfall. The North Atlantic subsequently warmed following the reduction of aerosol emissions, leading to rainfall recovery.

A positive trend in summer (December to February) precipitation has been detected in multiple observational sources in south-eastern South America since the beginning of the 20th century (Gonzalez et al., 2013; Vera and Díaz, 2015; Wu et al., 2016; H. Zhang et al., 2016; Díaz and Vera, 2017; Saurral et al., 2017). Sedimentary records from the Mar Chiquita lake indicate that the last quarter of the 20th century was wetter than any period during the last 200 years (Piovano et al., 2004). In this attribution example the drivers contributing to the positive trend for the period 1951–2014 are discussed (Figure 10.12a). Precipitation anomalies of Climatic Research Unit Time Series (CRU TS) as well as for the two members of a SMILE with the most negative and positive trends for 1951–2014 are displayed in Figure 10.12b. The trend for 1951–2014 using CRU TS and GPCC is illustrated in Figure 10.12c, and for the region defined by the black quadrilateral, it amounts to 2.8 (CRU TS) – 3.5 (GPCC) mm per month and decade (see black crosses in Figure 10.12d) while the mean summer monthly precipitation for the same period is 104 (CRU TS) –109 (GPCC) mm. The trend is also detectable in daily and monthly extremes (Re and Barros, 2009; Marengo et al., 2010; Penalba and Robledo, 2010; Doyle et al., 2012; Donat et al., 2013; Lorenz et al., 2016).

Figure 10.12 | South-Eastern South America positive mean precipitation trend and its drivers during 1951–2014. (a) Mechanisms that have been suggested to contribute to South-Eastern South America summer wetting. (b) Time series of austral summer (December to February) precipitation anomalies (%, baseline 1995–2014) over the South-Eastern South American region (26.25°S–38.75°S, 56.25°W–66.25°W), black quadrilateral in the first map of panel (c). Black, brown and green lines show low-pass filtered time series for CRU TS), and the members with driest and wettest trends of the MPI-ESM single-model initial-condition large ensemble (SMILE; between 1951–2014), respectively. The filter is the same as the one used in Figure 10.10. (c) Mean austral summer precipitation spatial linear 1951–2014 trends (mm per month and decade) from CRU TS and GPCC. Trends are estimated using ordinary least squares regression. (d) Distribution of precipitation 1951–2014 trends over South-Eastern South America from GPCC and CRU TS (black crosses), CMIP6 all-forcing historical (red circles) and MIROC6, CSIRO-Mk3-6-0, MPI-ESM and d4PDF SMILEs (grey box-and-whisker plots). Grey squares refer to ensemble mean trends of their respective SMILE and the red circle refers to the CMIP6 multi-model mean. Box-and-whisker plots follow the methodology used in Figure 10.6. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

The influence of SST anomalies on south-eastern South America precipitation have been studied extensively on interannual to multi-decadal time scales (Paegle and Mo, 2002). The positive phase of El Niño–Southern Oscillation (ENSO; Annex IV.2.3) is related to stronger mean and extreme rainfall over south-eastern South America (Ropelewski and Halpert, 1987; Grimm and Tedeschi, 2009; Robledo et al., 2016). The ENSO influence may be modulated by the PDV (Kayano and Andreoli, 2007; Fernandes and Rodrigues, 2018) and the AMV (Kayano and Capistrano, 2014). PDV and AMV also influence the south-eastern South American climate independently of ENSO (Barreiro et al., 2014; Grimm and Saboia, 2015; Robledo et al., 2020). While Pacific SSTs dominate the overall influence of oceanic variability in the region, the Atlantic variability seems to dominate on multi-decadal time scales and has been proposed as a driver for the long-term positive trend (Seager et al., 2010; Barreiro et al., 2014). Based on experiments designed to test how south-eastern South America precipitation is modulated by tropical Atlantic SSTs, Seager et al. (2010) showed that cold anomalies in the tropical Atlantic favour wetter conditions by inducing an upper-tropospheric flow towards the equator, which, via advection of vorticity, leads to ascending motion over south-eastern South America (Figure 10.12a). Monerie et al. (2019) supported this argument showing a negative relationship between south-eastern South America precipitation and the AMV index (Huang et al., 2015) using an AGCM coupled to an ocean mixed-layer model with nudged SSTs.

The positive trend of precipitation has also been attributed to anthropogenic GHGemissions and stratospheric ozone depletion. CMIP5 models only show a positive trend when including anthropogenic forcings (Vera and Díaz, 2015). These results were supported by Knutson and Zeng (2018) based on univariate detection/attribution analysis of annual mean trends for the 1901–2010 and 1951–2010 periods. However, the main features of summer mean precipitation and variability of South America are still not well-represented in all CMIP5 and CMIP6 models (Gulizia and Camilloni, 2015; Díaz and Vera, 2017; Díaz et al., 2021). This motivates the construction of ensembles that exclude the worst performing models (Section 10.3.3.4). The construction of ensembles of CMIP5 historical simulations with realistic representation of precipitation anomalies with opposite sign over south-eastern South America and eastern Brazil showed that the trend since the 1950s could be related to changes in precipitation characteristics only when simulations included anthropogenic forcings (Díaz and Vera, 2017). GHG emissions have been related to increased precipitation in south-eastern South America through three different mechanisms (Figure 10.12a). First, GHG warming induces a non-zonally uniform pattern of SST warming that includes a warming pattern over the Indian and Pacific oceans that excites wave responses over South America (Junquas et al., 2013). Zonally uniform SST patterns of warming alone lead to precipitation signals opposite to those observed in an AGCM (Junquas et al., 2013). Second, GHG radiative forcing drives an expansion of the Hadley cell so that its descending branch moves poleward from the region, generating anomalous ascending motion and precipitation (H. Zhang et al., 2016; Saurral et al., 2019). The third mechanism by which increased GHG can contribute to increased precipitation in the region is through a delay of the stratospheric polar vortex breakdown. As depicted in Figure 10.12a, both stratospheric ozone depletion and increased GHGs have contributed to the later breakdown of the polar vortex in recent decades (McLandress et al., 2010; Wu and Polvani, 2017; Ceppi and Shepherd, 2019). Mindlin et al. (2020) developed future atmospheric circulation storylines (Section 10.3.4.2, Box 10.2) for Southern Hemisphere mid-latitudes with the CMIP5 models and found that for south-eastern South America summer precipitation, increases are related to the late-spring breakdown of the stratospheric polar vortex. The connecting mechanism is through a lagged southward shift of the jet stream (Saggioro and Shepherd, 2019), which enhances cyclonic activity over the region (Wu and Polvani, 2017).

A common feature among the above discussed studies is that even if global models simulate positive trends when forced with GHG and/or stratospheric ozone, these trends are in general smaller than those observed (e.g., CMIP6 trends in red open circles in Figure 10.12d). Díaz et al. (2021) showed that to capture the observed trend a multi-model ensemble of SMILEs is needed. Out of the 12 large ensembles examined (with ensemble size varying in the 16–100 range), only seven simulated the observed trend within their range. This could partly be explained by model biases in mean precipitation and its interannual variability. In the sub-ensemble of six models that reproduce reasonably well the observed spatial patterns of mean precipitation and interannual variability, the ensemble mean spread is lower, and the forced response, taken as the multi-model ensemble mean, is slightly more positive than that of the six poorly performing models. The signal-to-noise ratio, estimated as the ratio of the forced response to the spread due to internal variability, is also slightly higher for the best-performing models, suggesting that selecting the best-performing models may have an influence on both attribution of the observed trend and emergence of the forced response in future (Section 10.4.3).

There is high confidence that South-Eastern South America summer precipitation has increased since the beginning of the 20th century. Since AR5, science has advanced in the identification of the drivers of the precipitation increase in South-Eastern South America since 1950, including GHG through various mechanisms, stratospheric ozone depletion and Pacific and Atlantic variability. There is high confidence that anthropogenic forcing has contributed to the South-Eastern South America summer precipitation increase since 1950, butvery low confidence on the relative contribution of each driver to the precipitation increase.

Persistent hydroclimatic drought in south-western North America remains a much-studied event. Drought is a regular feature of the south-western North America’s climate regime, as can be seen in both the modern record, and through paleoclimate reconstructions (Cook et al., 2010; Woodhouse et al., 2010; Williams et al., 2020), as well as in future climate model projections (Cook et al., 2015a). Since the early 1980s, which were relatively wet in terms of precipitation and streamflow, the region has experienced major multi-year droughts such as the turn-of-the-century drought that lasted from 1999 to 2005, and the most recent and extreme 2012–2014 drought that in certain locations is perhaps unprecedented in the last millennium (Section 8.3.1.6; Griffin and Anchukaitis, 2014; Robeson, 2015). Shorter dry spells also happened between these multi-year droughts making 1980 to present a period with an exceptionally steep trend from wet to dry (Figure 10.13a), leading to strong declines in Rio Grande and Colorado river flows (Lehner et al., 2017b; Udall and Overpeck, 2017). While robust attribution of this trend is complicated by the large natural variability in this region, the 20th century warming has been suggested to increase the chances for hydrological drought periods by lowering runoff efficiency (Woodhouse et al., 2016; Lehner et al., 2017b; Woodhouse and Pederson, 2018) and affecting evapotranspiration (Williams et al., 2020). There is some evidence suggesting that the Last Glacial Maximum, a period of low atmospheric CO₂, about 21 ka ago, has a thermodynamically-driven zonal mean precipitation response similar to that of the current state with relatively high CO₂ levels when compared with the pre-industrial period. Pluvial conditions at that time and a reduction in precipitation from the Last Glacial Maximum to the pre-industrial period are consistent with drying trends for the region in models with GHG concentrations exceeding pre-industrial levels. However, the dominant large-scale drivers responsible for the precipitation changes observed during these two transitions are markedly different: mainly ice-sheet retreat and increasing insolation on one hand, increasing GHGs on the other hand. This suggests that the Last Glacial Maximum correspondence is fortuitous which strongly limits its use to capture future hydrological cycle changes (Section 8.3.2.4.4; Morrill et al., 2018; Lowry and Morrill, 2019). Furthermore, the conclusion of the Last Glacial Maximum drying versus wetting seems to strongly depend on the physical property of interest, hydrologic or vegetation indicators (Scheff et al., 2017). Droughts are characterized by deficits in total soil moisture content that can be caused by a combination of decreasing precipitation and warming temperature, which promotes greater evapotranspiration. Regional-scale attribution of the prevalence of south-western North America drought since 1980 then mostly focuses on the attribution of change in these two variables.

Figure 10.13 | Attribution of the south-western North America precipitation decline during the 1983–2014 period. (a) Water year (October to September) precipitation spatial linear trend (in percent per decade) over North America from 1983 to 2014. Trends are estimated using ordinary least squares. Top row: observed trends from CRU TS, REGEN, GPCC, and the Global Precipitation Climatology Project (GPCP). Middle row: driest, mean and wettest trends (relative to the region enclosed in the black quadrilateral, bottom row) from the 100 members of the MPI-ESM coupled SMILE. Bottom row: driest, mean and wettest trends relative to the above region from the 100 members of the d4PDF atmosphere-only SMILE. (b) Time series of water year precipitation anomalies (%, baseline 1971–2000) over the above south-western North America region for CRU TS (grey bar charts). Black, brown and green lines show low-pass filtered time series for CRU TS, driest and wettest members of the d4PDF SMILE, respectively. The filter is the same as the one used in Figure 10.10. (c) Distribution of south-western region-averaged water-year precipitation 1983–2014 trends (in percent per decade) for observations (CRU TS, REGEN, GPCC and GPCP, black crosses), CMIP6 all-forcing historical simulations (red circles), the MIROC6, CSIRO-Mk3-6-0, MPI-ESM and d4PDF SMILEs (grey box-and-whisker plots). Grey squares refer to ensemble mean trends of their respective SMILE and the red circle refers to the CMIP6 multi-model mean. Box-and-whisker plots follow the methodology used in Figure 10.6. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

The observed south-western North America drying fits the narrative of what might happen in response to increasing GHG concentrations due to a poleward expansion of the subtropics, that is conducive to drying trends over subtropical to mid-latitude regions (Hu et al., 2013b; Birner et al., 2014; Lucas et al., 2014). However, several studies based on modern reanalyses and CMIP5 models have recently shown that the current contribution of GHGs to Northern Hemisphere tropical expansion is much smaller than in the Southern Hemisphere and will remain difficult to detect due to large internal variability, even by the end of the 21st century (Section 3.3.3.1; Garfinkel et al., 2015; Allen and Kovilakam, 2017; Grise et al., 2018, 2019). In addition, the widening of the Northern Hemisphere tropical belt exhibits strong seasonality and zonal asymmetry, particularly in autumn and the North Atlantic (Amaya et al., 2018; Grise et al., 2018). Therefore, it seems that the recent Northern Hemisphere tropical expansion results from the interplay of internal and forced modes of tropical width variations and that the forced response has not robustly emerged from internal variability (Sections 3.3.3.1 and 10.4.3).

A second possible causal factor is the role for ocean-forced or internal atmospheric circulation change. Analysis of observed and CMIP5-simulated precipitation indicates that the drought prevalence since 1980 is linked to natural, internal variability in the climate system (Knutson and Zeng, 2018). Based on observations and ensembles of SST-driven atmospheric simulations, Seager and Hoerling (2014) suggested that robust tropical Pacific and tropical North Atlantic forcing drove an important fraction of annual mean precipitation and soil moisture changes and that early 21st century multi-year droughts could be attributed to natural decadal swings in tropical Pacific and North Atlantic SSTs. A cold state of the tropical Pacific would lead by well-established atmospheric teleconnections to anomalous high pressure across the North Pacific and southern North America, favouring a weaker jet stream and a diversion of the Pacific storm track away from the south-west (Delworth et al., 2015; Seager and Ting, 2017). The multi-year drought of 2012–2016 has been linked to the multi-year persistence of anomalously high atmospheric pressure over the north-eastern Pacific Ocean, which deflected the Pacific storm track northward and suppressed regional precipitation during California’s rainy season (Swain et al., 2017). Going into more detail, Prein et al. (2016a) used an assessment of changing occurrence of weather regimes to judge that changes in the frequency of certain regimes during 1979–2014 have led to a decline in precipitation by about 25%, chiefly related to the prevalence of anticyclonic circulation patterns in the north-east Pacific. Finally, the moderate model performance in representing Pacific SST decadal variability and its remote influence (Section 3.7.6) as well as its change under warming may affect attribution results of observed and future precipitation changes (Seager et al., 2019).

It has also been suggested that the ocean-controlled influence is limited and internal atmospheric variability has to be invoked to fully explain the observed history of drought on decadal time scales (Seager and Hoerling, 2014; Seager and Ting, 2017). From roughly 1980 to the present, the regional climate signals show an interesting mix between forced and internal variability. Lehner et al. (2018) used a dynamical adjustment method and large ensembles of coupled and SST-forced atmospheric experiments to suggest that the observed south-western North America rainfall decline mainly results from the effects of atmospheric internal variability, which is in part driven by a PDV-related phase shift in Pacific SST around 2000 (Figure 10.13b,c). Based upon four SMILEs (three using a GCM and another one an AGCM constrained by observed SSTs) and a CMIP6 multi-model suite constrained by observed external forcings, Figure 10.13 shows, in agreement with Lehner et al. (2018), that observed SSTs with their associated atmospheric response are the main drivers of the south-western North America precipitation decrease during the 1983–2014 period. Once aspects of the internal variability are removed by dynamical adjustment, the observed precipitation change signal and simulated anthropogenically-forced components look more similar (Lehner et al., 2018).

Importantly, as the AR6 assessment views the PDV as being mostly driven by internal variability (Section 3.7.6), the lines of evidence cited above suggest that the contribution of natural and anthropogenic forcings to the precipitation decline has a small amplitude. Unlike the precipitation deficit, the accompanying south-western North America warming is driven primarily by anthropogenic forcing from GHGs rather than atmospheric circulation variability and may help to enhance the drought through increased evapotranspiration (Knutson et al., 2013; Diffenbaugh et al., 2015; Williams et al., 2015, Williams et al., 2020; Lehner et al., 2018, 2020).

To conclude, there is high confidence (robust evidence and medium agreement) that most (>50%) of the anomalous atmospheric circulation that caused the south-western North America negative precipitation trend can be attributed to teleconnections arising from tropical Pacific SST variations related to PDV. There is high confidence (robust evidence and medium agreement) that anthropogenic forcing has made a substantial contribution (about 50%) to the south-western North America warming since 1980.

The robustness of regional-scale attribution differs strongly between temperature and precipitation changes. While the influence of anthropogenic forcing on regional temperature long-term change has been detected and attributed in almost all land regions, a robust detection and attribution of human influence on regional precipitation change has not yet fully occurred for many land regions (Section 10.4.3). Although the contribution of anthropogenic forcing to long-term regional precipitation change has been detected in some regions, a robust quantification of the contributions of different drivers remains elusive. The delayed emergence of the anthropogenic precipitation fingerprint with respect to temperature is likely due to the opposing sign of the fast and slow land precipitation forced responses and time-dependent SST change patterns (Sections 8.2.1 and Section 10.4.3), stronger internal variability (Section 10.3.4.3) as well as larger observational uncertainty (Section 10.2) and impact of model biases. The contribution of internal variability to the observed changes can also be very sensitive to the period length and level of spatial aggregation for the region under scrutiny (Section 4.4.1 and Cross-Chapter Box 3.1; Kumar et al., 2016). Finally, even in the case of temperature changes at multi-decadal time scale, internal variability can still be a substantial driver of regional changes due to cancellation between different external forcings (Nath et al., 2018).

To conclude, it is virtually certain (robust evidence and high agreement) that anthropogenic forcing has been a major driver of temperature change since 1950 in many sub-continental regions of the world. There is high confidence (robust evidence and medium agreement) that anthropogenic forcing has contributed to multi-decadal mean precipitation changes in several regions, for example western Africa, south-east South America, south-western Australia, northern central Eurasia, and South and East Asia. However, at regional scale, the role of internal variability is stronger while uncertainties in observations, models and external forcing are all larger than at the global scale, precluding a robust assessment of the magnitude of the relative contributions of greenhouse gases, including stratospheric ozone, and different aerosol species.

Standard methodologies to derive the regional forced response include pattern-scaling and the time-shift or epoch approach (Section 4.2.4; Tebaldi and Arblaster, 2014; Vautard et al., 2014; Herger et al., 2015; Tebaldi and Knutti, 2018; Christensen et al., 2019). Pattern-scaling assumes that the spatial patterns of regional change, often based on a time-averaged 20- or 30-year period at the end of the 21st century, are roughly constant in time, and simply scale linearly with global mean warming. The time-shift approach defines a target in terms of global warming level (GWL) and locates the time segment, usually 20 or 30 years, in historical or scenario simulations in which global mean warming matches the required GWL (Section 10.1.2 and Cross-Chapter Box 11.1). Physical consistency between multiple variables and space-time co-variance are fully preserved in the time-shift approach, which is not the case for pattern-scaling (Herger et al., 2015). Importantly, pattern scaling cannot account for the non-linearity arising from either interacting quasi-linear processes (Chadwick and Good, 2013) and purely non-linear mechanisms, which have been shown to be present in CMIP5 models for high GWL (4°C) and affect precipitation more than temperature at the regional-scale (Section 8.5.3.1; Good et al., 2015, 2016). The time-shift approach can also be used to test whether regional climate change patterns depend on the rate of global mean warming and external forcing pathways, in addition to global warming magnitude. A global evaluation of both approaches in projecting the forced temperature and precipitation response for a highly mitigated scenario based on a moderately mitigated one has been performed using a perfect-model framework (Tebaldi and Knutti, 2018). The amplitude of errors for both approaches appears to be substantially smaller than model uncertainty approximated by the CMIP5 multi-model spread.

Based on large and coordinated modelling exercises such as CMIP5 and CORDEX, the time-shift approach has been largely used to assess differences in regional climate impacts for different GWLs, with a strong focus on 1.5°C versus 2°C (Karmalkar and Bradley, 2017; Dosio and Fischer, 2018; Karnauskas et al., 2018; W. Liu et al., 2018; Taylor et al., 2018; Weber et al., 2018; Chapter 3, SR1.5, Hoegh-Guldberg et al., 2018). Comparisons between pattern-scaling and time-shift approaches allow assessment of the scalability of the regional climate change signal and the extent to which pattern-scaling assumptions still hold at regional scale for a wide range of GWL. This was the approach followed by Matte et al. (2019) in their assessment of the scalability of European regional climate projections. Based on EURO-CORDEX projections, they performed a detailed comparison between the pattern scaling and the GWL spatial patterns (GWL range: 1°C, 2°C and 3°C) for different seasons, regional model resolutions, and both temperature and precipitation. High pattern correlation values (greater than 0.9) are found between the scaled pattern and all GWL patterns for temperature. In the case of precipitation, the correspondence is slightly lower, especially in summer, for high GWLs (2°C and 3°C) and much lower for 1°C.

Figure 10.14 illustrates a similar comparison based on the CMIP6 multi-model ensemble forced with the scenario SSP5-8.5 and applied to two large-scale continental areas. The forced response to anthropogenic forcing is simply taken as the CMIP6 multi-model mean of future regional climate change relative to the 1850–1900 reference period. Robustness of the forced response is based on both significance of the change and model agreement about the sign (direction) of change (Cross-Chapter Box Atlas.1; Figure 10.14). Caution has to be exercised against a too literal interpretation of lack of robust change given that significance and sign agreement can be sensitive to spatial and temporal aggregation (Cross-Chapter Box Atlas.1, Figure 2) and lack of a robust change does not necessarily translate to lack of regional-scale climate change impacts (McSweeney and Jones, 2013; Hibino and Takayabu, 2016).

Figure 10.14 | Robustness and scalability of anthropogenic signals at regional scale. (a) Spatial patterns of European and African summer (June to August) surface air temperature change (in °C °C^–1) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) multi-model mean (45 models, one member per model, historical simulations and scenario SSP5-8.5) at different global warming levels (GWLs) and the end-21st century scaling pattern estimated from the multi-model mean difference between 2081–2100 and the pre-industrial period (1850–1900) divided by the corresponding global mean warming. The scale of all GWL patterns has been adjusted to a global mean warming of 1°C (for example, the resulting 3°C spatial pattern has been divided by three). The scales of the GWL patterns have to be multiplied by their threshold values to obtain the actual simulated warming. The metrics shown in the bottom left corner of the GWL pattern plots indicate the spatial pattern correlation and the root-mean-square difference between the GWL patterns and the scaling pattern. The number in bold just above the metrics gives the number of used CMIP6 models (out of 45) that have reached the GWL threshold. Areas with robust change (at least 66% of the models have a signal-to-noise ratio greater than one and 80% or more of the models agree on the sign of the change) are coloured with no pattern overlaid (Cross-Chapter Box Atlas.1). Areas with a significant change (at least 66% of the models have a signal-to-noise ratio greater than one) and lack of model agreement (meaning that less than 80% of the models agree on the sign of the change) are marked by cross-hatching. Areas with no change or no robust change (less than 66% of the models have a signal-to-noise ratio greater than one) are marked by negatively sloped hatching. (b) Same as (a) but for North, Central and South America annual mean precipitation relative change (percent °C^–1). The baseline for precipitation climatology is 1850–1900. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

If projected regional mean temperature (Figure 10.14a) and precipitation (Figure 10.14b) changes were to scale linearly with global mean warming, the adjusted spatial patterns would be congruent with each other at different GWLs. While pattern scaling seems to be a reasonable first-order approximation for both temperature and precipitation changes in tropical and high latitude regions (high pattern correlation values), there are a number of regions exhibiting substantial amplitude differences at different GWLs (northern Africa and Middle East, southern and eastern Europe for temperature; south-western North America, Chile and north-eastern Brazil for precipitation). These differences hint at the possible influence of non-linear mechanisms (Good et al., 2015), including soil-moisture feedbacks (Seneviratne et al., 2010; Vogel et al., 2017), a time-dependent balance between the different contributions of fast and slow response to greenhouse gas forcing as well as changing SST response patterns (Long et al., 2014; Good et al., 2016; Ceppi et al., 2018; Zappa et al., 2020). Decreasing spatial pattern amplitude with increasing GWL suggests that the initial transient regional response overshoots the long-term change in regions such as northern Africa for summer temperature and south-western South America for precipitation (Zappa et al., 2020). In the latter region, long simulations with stabilized GHG concentrations even suggest a change of sign when near-equilibrium is reached (Sniderman et al., 2019). The reverse behaviour, increasing pattern amplitude with increasing GWL, is seen for summer temperature in southern and eastern Europe and for precipitation in south-western North America (Sniderman et al., 2019; Zappa et al., 2020), suggesting that, in these regions, the initial transient response is lagging global mean warming and final regional climate change will be reached once GHG concentrations are stabilized.

There is high confidence that the time-evolving contribution of different mechanisms operating at different time scales can modify the amplitude of the regional-scale response of temperature, and both the amplitude and sign of the regional-scale response of precipitation, to anthropogenic forcing. These mechanisms include non-linear temperature, precipitation and soil-moisture feedbacks, and slow and fast response of SST patterns and atmospheric circulation changes to increasing GHGs.

This section provides an assessment of the different approaches used in emergence studies as well as sensitivities to methodological choices. The section then focuses on the possible influence of internal variability on future emergence of the simulated mean precipitation anthropogenic signal at regional scales with some illustrative examples.

In climate science, emergence or distinguishability of a signal refers to the appearance of a persistent change in the probability distribution and/or temporal properties of a climate variable compared with that of a reference period (Section 1.4.2; Giorgi and Bi, 2009; Mahlstein et al., 2011, 2012; Hawkins and Sutton, 2012). Similar to anthropogenic climate change detection (Cross-Working Group Box: Attribution in Chapter 1), signal emergence can be detected, at least initially, without identifying the physical causes of the emergence (Section 1.4.2). In the context of human influence on climate, the objective of emergence studies is the search for the appearance of a signal characterizing an anthropogenically-forced change relatively to the climate variability of a reference period, defined as the noise.

Precise definitions of signal and noise as well as a metric to measure the relative importance of the signal are key ingredients of the emergence framework and depend on the framing question. In particular, emergence study results can depend on the specific definitions of signal and noise such as the level of spatial and temporal aggregation (McSweeney and Jones, 2013). For instance, grid-point scale emergence will likely be delayed compared with region-average emergence (Section 11.2.4 and Cross-Chapter Box Atlas.1, Figure 2; Fischer et al., 2013; Maraun, 2013b; Lehner et al., 2017a). The signal is often estimated by a running mean multi-decadal average or probability distribution function of the physical variable under scrutiny in order to avoid false emergence due to manifestation of multi-decadal internal variability (King et al., 2015). In the case of extremes such as climate records, a notion of multi-year persistence or recurrence can also be used to fully characterize the anthropogenic signal and its emergence (Christiansen, 2013; Bador et al., 2016).

Emergence is also sensitive to the noise characteristics: assuming a common signal definition, larger signal-to-noise values and earlier emergence will arise if the noise is based on decadal mean variability rather than interannual variability (Kusunoki et al., 2020). Depending on the framing question, the noise can include or omit external natural forcing such as volcanic and solar forcing (Zhang and Delworth, 2018; Silvy et al., 2020). Furthermore, emergence results are very sensitive to the choice and length of the reference period (Section 1.4.1). The reference period can be the pre-industrial, the very recent past or even a time-evolving baseline, depending on both the framing and assumption that adaptation to the current climate has already occurred (King et al., 2015; Zhang and Delworth, 2018; Brouillet and Joussaume, 2020). These choices will then determine the type of simulations and periods that will be used to construct the noise distribution. Finally, the permanence of future emergence cannot be taken for granted when emergence occurs in the late-21st century based on simulations ending in 2100 (Hawkins et al., 2014; King et al., 2015; Lehner et al., 2017a).

Robust assessments and comparisons of past emergence between observations and models are strengthened by the use of consistent definitions of signal and noise (Lehner et al., 2017a; Hawkins et al., 2020). In the case of future emergence under increasing greenhouse gas emissions, two main approaches have been followed to assess emergence. The first is based on estimating the signal and noise (and sometimes the signal-to-noise ratio as well) in individual models before using the resulting distribution median or mean to construct the final emergence metric (Hawkins and Sutton, 2012; Maraun, 2013b; Sui et al., 2014; Barrow and Sauchyn, 2019). The second method first estimates the signal as a multi-model mean change and the noise variance as a combination of internal variability and model structural differences (Giorgi and Bi, 2009; Mariotti et al., 2015; Nguyen et al., 2018). The first approach allows the definition of emergence of the signal relative to internal variability only and treats model error as source of uncertainty (Maraun, 2013b; Lehner et al., 2017a). The second assumes that the multi-model mean is the optimal estimate of the signal and confounds internal variability and model structural differences in the noise estimate. It is noteworthy that most emergence studies implicitly assume model independence (Annan and Hargreaves, 2017; Boé, 2018; Box 4.1) and therefore sensitivity of emergence results to model selection or weighting is rarely performed (Akhter et al., 2018).

Metrics can vary from a simple signal-to-noise ratio to statistical distributional tests (King et al., 2015; Gaetani et al., 2020) and give median estimates and uncertainty bounds for the date (or time of emergence) corresponding to the exceedance of specific thresholds by the emergence metric. Reconciling future emergence results among different studies is challenging due to their many methodological differences including the choice of the reference period, the selected climate models and scenario, the precise definition of signal and noise and the choice of different signal-to-noise thresholds to characterize robust emergence. Contrasting with binary yes/no statements, emergence can also be viewed as a continuous process characterized by an amplitude or level, for example the value of the signal-to-noise ratio, that is a function of time or global warming level.

Since AR5, the development and production of SMILEs (Sections 4.2.5 and 10.3.4.3) has allowed the assessment of the influence of internal variability on anthropogenic signal emergence. The influence of internal variability, and specifically of the unforced atmospheric circulation, on temperature signal emergence can delay or advance the time of emergence by a decade or two in mid- to high-latitude regions (Lehner et al., 2017a; Koenigk et al., 2020). Internal variability can also result in small or decreasing decadal to multi-decadal heatwave frequency trends under the historical anthropogenic forcing over most regions, thereby delaying emergence of unprecedented heatwave frequency trends relative to the pre-industrial trend distribution (Sections 11.2–11.3; Perkins-Kirkpatrick et al., 2017).

Regional precipitation future changes are much more impacted by internal variability than their temperature counterpart (Monerie et al., 2017b; Dai and Bloecker, 2019; Singh and AchutaRao, 2019; von Trentini et al., 2019; Koenigk et al., 2020). Relative to mean temperature changes, this larger influence of internal variability on mean precipitation changes contributes, among other factors (Sarojini et al., 2016), to a much delayed emergence of the forced precipitation response in observations (Hawkins et al., 2020). Based on the CMIP6 multi-model ensemble forced with the scenario SSP5-8.5, we assess the future emergence of mean precipitation forced change as a function of GWLs for all AR6 land regions (Figure 10.15a). The methodology is a straightforward adaptation of the standard approach (Hawkins and Sutton, 2012). While the standard method is only based on the signal-to-noise ratio exceedance of a specified threshold (taken as one), the approach used here assumes that grid-point emergence occurs when the forced change is considered robust following the AR6 WGI definition of robustness for projected changes (Cross-Chapter Box Atlas.1). At a GWL of 1°C, emergence only occurs in high-latitude regions (Wan et al., 2015; R. Guo et al., 2019), albeit with only small (less than 30%) area fraction with robust change. Robust changes in tropical and subtropical regions only appear from GWLs of 1.5°C, for example in south-western South America (Boisier et al., 2016), western Africa (Hawkins et al., 2020; Section 10.4.2.1) and southern Australia (Delworth and Zeng, 2014). Substantial (taken here simply as area fraction greater than 50%) emergence only occurs in some tropical, subtropical and mid-latitude regions when high GWLs (3°C–4°C) are reached. Importantly, even at these high GWL values, there are still a large number of these regions with robust changes covering less than 50% of their area. In contrast, most high-latitude regions have an area fraction with robust changes greater than 80% at GWLs of 3°C and above.

Figure 10.15 | Future emergence of anthropogenic signal at regional scale. (a) Percentage area of land regions with robust annual mean precipitation change as a function of increasing global warming levels (GWLs). Robustness of the precipitation change is first estimated at each grid-point followed by the estimation of the AR6 region area with robust changes. For each Coupled Model Intercomparison Project Phase 6 (CMIP6) model considered (45 models, one member per model, historical simulations and scenario SSP5-8.5), the annual mean precipitation change is based on the difference between a 20-year average centred on the GWL crossing year and the mean precipitation during the pre-industrial period (1850–1900) taken as a reference. The change is considered to be robust when at least 66% of the models (30 out of 45) have a signal-to-noise ratio greater than one and at least 80% of them (36 out of 45) agree on the sign of change. The signal-to-noise ratio is estimated for each model from the ratio between the change and the standard deviation of non-overlapping 20-year means of the corresponding pre-industrial simulation (scaled by square root of 2 times 1.645). (b) Time evolution of the percentage area of land region with robust annual mean precipitation change for five AR6 land regions. Thick solid lines represent precipitation changes based on the same CMIP6 ensemble as in (a). Thin solid, dotted and dashed lines represent changes based on the three coupled single-model initial-condition large ensembles (SMILEs) used in Chapter 10, illustrating the influence of internal variability on the emergence of robust change. The change is estimated from the difference between all consecutive 20-year periods from 1900–1919 up to 2081–2100 and the pre-industrial period. The line colour indicates the sign of the robust change given by the multi-model mean (CMIP6) or ensemble mean (SMILE) change: brown (decreasing precipitation) and dark green (increasing precipitation). Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

We now illustrate the potential influence of internal variability on late or lack of emergence for a few AR6 land regions (Figure 10.15b). For each of these AR6 regions, the time evolution of the percentage area with robust annual mean precipitation change is estimated for both the CMIP6 multi-model ensemble and the three coupled SMILEs used throughout Chapter 10. Similarity in percentage area time evolution between CMIP6 and the three coupled SMILEs suggests that internal variability can substantially influence the timing of emergence. For example, internal variability could explain the mid-21st century emergence (percentage area greater than 50%) of the drying and wetting signal over the Mediterranean and South Asia (see also Section 10.6.3) regions, respectively. Internal variability can also contribute to the late and moderate emergence over South-Eastern South America (see also Section 10.4.2) and West South Africa (see also Section 10.6.2). In contrast, it cannot explain the lack of robust changes (percentage area less than 30%) over Western Africa at the end of the 21st century, suggesting that model differences are also contributing to the lack of emergence (Monerie et al., 2017a, b). In addition to different forced signals, the differences of time evolution between the three SMILEs, in particular for African regions, point to the issue of global model performance in accurately representing internal variability and its future changes. While overestimation and underestimation of internal variability in current models have been reported (Eade et al., 2014; Laepple and Huybers, 2014), methodological challenges to assess the magnitude and spatial pattern of model biases in simulating internal variability, still remain Section 10.3.4.3). Therefore, the existence of model biases and the limited knowledge of their characteristics lead to limitations about a precise quantification of internal variability influence on delayed regional-scale emergence.

There is high confidence that consistency in definitions of signal and noise, choice of the reference period and signal-to-noise threshold, is important to robustly assess the future emergence of anthropogenic signals across different types or generations of models, as well as comparing past emergence results between observations and models. There is high confidence that internal variability can delay the emergence of the regional-scale mean precipitation anthropogenic signal in many regions, mainly located in the tropics, subtropics and mid-latitudes. An accurate estimation of the delay in regional-scale emergence caused by internal variability remains challenging due to global model biases in their representation of internal variability as well as methodological difficulties to precisely estimate these biases (high confidence).

Without considering the specific context, the distillation of climate information relevant to users may poorly serve the goal of informing adaptation and policy (Cash et al., 2003; Lemos et al., 2012; Baztan et al., 2017). Section 10.1.4 identifies three implicit framing issues of constructing and delivering user-relevant climate information: practical issues arising from the climate information sources, issues with including the context in constructing the information, and difficulties presented by complex networks of practitioners. The social context strongly influences decisions about constructing information and requires a nuanced and holistic approach to recognize the complexity of a coupled social and physical system (Daron et al., 2014). For example, urban water managers must recognize the dependency of the city on different water resources and the interplay of both local and national government legislation that can involve a range of different constituencies and decision makers (Scott et al., 2018; Savelli et al., 2021).

Context plays a role in determining the risks that may affect human systems and ecosystems and consequently the climate information needs. The context may also limit access to such information. Hence, the context imposes inherent constraints on how climate information can be constructed and optimally aligned with its intended application. Although contexts are unlimited in variety, some key contextual elements include:

• Whether the problem formulation needs to be constructed through consultative activities that, for instance, help identify thresholds of vulnerability in complex urban or rural systems (Baztan et al., 2017; Willyard et al., 2018) or is more a matter of addressing a generic vulnerability already identified, such as the frequency of flood events or recurrence intervals of multi-year droughts (Hallegatte et al., 2013).
• Societal capacity, such as cultural or institutional flexibility and willingness to respond to different scientific information (e.g., Hart and Nisbet, 2012; Kahan, 2012, 2013).
• The technical capability and expertise of the different actors, including users, producers, and communicators (e.g., Sarewitz, 2004; Gorddard et al., 2016).
• Potential contrasts in value systems such as the different views of the Global North compared to those of economies in transition or under development (Henrich et al., 2010a, b; Sapiains et al., 2021).
• The relative importance of climate change in relation to non-climate stressors on the temporal and spatial scales of interest to the user, which at times are not the ones initially assumed by the producers (Otto et al., 2015).
• Availability, timing and accessibility of the required climate information, including the availability of sources such as observations, model simulations, literature and experts of the relevant regional climate (Mulwa et al., 2017). In developing countries, the availability of all or some of these sources may be limited (Dinku et al., 2014).

These and other contextual elements can frame subsequent decisions about the construction of regional climate information for applications. For example, an engineer typically seeks quantitative information, while the policy community may be more responsive to storylines and how information is positioned within a causal network describing regional climate risk (Section 1.4.4 and Box 10.2). Multiple contexts can coexist and potentially result in competing approaches (for example, when urban governance contends with regional water-resource management in the same area).

Developing climate information relevant to user needs can be influenced by the explicit and implicit values of all parties: those constructing the information, those communicating the information, those receiving the information, and, critically, those who construct the problem statement being addressed. A discussion of how values in the scientific community shape climate research appears in Section 1.2.3.2. The influence of values need not be a source of bias or distortion; it is sometimes appropriate and beneficial: critical scrutiny from a diverse range of value-governing perspectives may uncover and challenge biases and omissions in the information that might otherwise go unrecognized (Longino, 2004). Dialogue among all parties in a culturally, socially, and economically heterogeneous society is therefore important for recognizing and reconciling value differences to best yield information that is salient, relevant and avoids ambiguity, most notably when informing the complexity of risks and resilience for human systems and ecosystems in developing nations (e.g., Baztan et al., 2017).

Thus, a challenge with constructing climate information for users, especially about impactful change, is that producing the information may need to involve people with a variety of backgrounds, who have different sets of experiences, capabilities, and values. The information thus would need to accommodate and be relevant to a range of different ways of viewing the problem (Sarewitz, 2004; Rosenzweig and Neofotis, 2013; Gorddard et al., 2016). Failure to recognize the variety of people using the climate information can make it ineffective, even if the source data on which it is based is of the highest quality, and may create a danger of maladaptation.

A substantial body of evidence shows that the receptivity of individuals to climate information is strongly conditioned by motivated reasoning (Hart and Nisbet, 2012; Kahan, 2012, 2013), wherein a person’s reception of climate information is influenced by the values of the community with which the person identifies. Adherence to a community’s values forms part of an individual’s social identity (Hart and Nisbet, 2012). Individuals thus frame their analysis and understanding of climate information in the context of cultural values espoused by their community (Hart and Nisbet, 2012; Kahan, 2012, 2013; Campbell and Kay, 2014; Bessette et al., 2017; Tschakert et al., 2017; Vezér et al., 2018). Successful framing of climate information products thus seeks to identify common ground with users, taking account of their values and interests.

Given the relevance of both context and values, the effectiveness of climate information can increase if developed in partnership with the target communities (Figure 10.17; Tschakert et al., 2016). Such an approach can inspire trust among all parties and at the same time promote a co-production process (Cash et al., 2003). Recipients of information have the greatest trust when the communicator is perceived as understanding their context and sharing their values and identity (Corner et al., 2014). As a consequence, developing mental models informed by user values can help with understanding complex climate models and their outcomes (Bessette et al., 2017).

Figure 10.17 | Effective regional climate information requires shared development of actionable information that engages all parties involved and the values that guide their engagement. Participants in the development of climate information come from varying perspectives, based in part on their professions and communities. Each of the three broad categories shown in the Venn diagram (Users, Producers, Scientists) is not a homogenous group, and often has a diversity of perspectives, values and interests among its members. The subheadings in each category are illustrative and not all-inclusive. The arrows connecting those categories represent the distillation process of providing context and sharing climate relevant information. The arrows that point toward the centre represent the distillation of climate information that involves all three categories.

Open figure

The importance of a co-production process does not preclude the climate-research community from taking steps to develop and convey relevant information on its own. Indeed, communicating expert consensus about contested scientific issues is beneficial (Goldberg et al., 2019). Climate services (Section 10.5.4), in particular, can become an effective means for using sources from the climate community and crafting these to be consistent with the needs, interests and values of stakeholder communities. However, simply presenting more information without recognizing user values and the contextual elements listed in Section 10.5.2.1 may be ineffective (Kahan, 2013). An aversion to climate information discordant with one’s pre-existing beliefs can actually become stronger for people who are more scientifically literate: they feel more confident sifting through all sources of information to find support for their positions (Kahan, 2012). A challenge is that if climate information is not framed carefully, recognizing context and user values, it may make the sceptical person less receptive to further information about climate change (Corner et al., 2012; Hart and Nisbet, 2012; Shalev, 2015). A further complication is that audiences may view climate change as a problem distant in time and space (Spence et al., 2012), too threatening to acknowledge (Brügger et al., 2015; McDonald et al., 2015), or too economically challenging to accept (Bessette et al., 2017). Identifying positive outcomes that align with user values, instead of adaptation and mitigation efforts, appears to promote the interest in and the success of climate information (Bain et al., 2012).

Climate processes occur on a range of spatial and temporal scales, from global to local, from centuries and longer to days or less (Section 10.1.2 and Figure 10.3). Similarly, decisions by stakeholders cover a range of spatial and temporal scales that can vary with the size of their region of interest and scope of activity. However, the link between decision scales and the spatial and temporal resolution of climate and related non-climatic, natural-system information is not straightforward, and failure to recognize mismatches between the two can undermine the effectiveness and relevance of the information (Cumming et al., 2006; Sayles, 2018).

Nevertheless, the scale of regional climate information does not have to be the same as the decision scale. Physical-climate storylines (Box 10.2) valid at large scales can be used to develop understanding that is relevant to local decisions. For example, global climate change affecting Antarctic ice-mass loss is relevant to formulating Dutch responses to sea level rise (Haasnoot et al., 2020). On the other hand, extreme precipitation processes can occur on scales of tens of kilometres and smaller and thus require high resolution climate information when projecting future changes (e.g., Xie et al., 2015). An important factor for developing effective climate information using the distillation process is aligning the vulnerabilities of the social and economic systems under consideration ranging from, for example, those important to a farmer to those important to a national agricultural ministry (Andreassen et al., 2018; O’Higgins et al., 2019). Thus, more sophisticated matching of spatial and temporal resolution of climate information with decision scales requires engagement across a hierarchy of governance structures at national, regional and local level (e.g., Lagabrielle et al., 2018).

Data, either from observations or models, is in general not inherently information, but may contain relevant information if interpreted appropriately (Hewitson et al., 2017). The same applies to other sources of climate information. Relevance is controlled by the given user context (Section 10.5.2.1) and relates to the required temporal and spatial scales (Section 10.5.2.3), the characteristics of required variables (often referred to as indicators), and the meteorological and climatic phenomena driving these variables (Section 10.1.3). For example, if climate information for driving impact models is sought (e.g., McSweeney et al., 2015), the impact modelling analysis in the target region is the specific user context.

Climate risk assessment considers all plausible outcomes (Weaver et al., 2017; Marchau et al., 2019; Sutton, 2019). Thus, a key element of information construction is the exploration and reconciliation of different sources of information (Barsugli et al., 2013; Hewitson et al., 2014b; Maraun and Widmann, 2018b) and involves mainly two issues: first, assessing the fitness of different sources in the given context and thereby potentially omitting (or down-weighting) selected sources (Sections 10.3.3), and, second, integrating different sources into a broader picture within a context (Sections 10.3.4).

A non-comprehensive selection of approaches that may contribute to the construction of information includes:

• Overall assessment and intercomparison of different sources of information, including hierarchies of models and identification of potentially conflicting results (Figure 10.16), where observational availability plays a critical role (Section 10.2.3).
• Assessing the emergence of forced trends from internal variability (Section 10.4.3), and testing whether differences in simulations can be explained by internal variability, ideally using initial-condition large ensembles (Sections 10.3.4.3 and 10.4.3).
• Assessing the interdependence of chosen models to identify the amount of independent information (Section 10.3.4.4).
• Process-based evaluation with focus on those processes that are relevant for the specific application (Sections 10.3.3.4–10.3.3.10).
• Weighting or sub-selecting ensembles based on a priori knowledge or the outcome of a process-based evaluation, while sampling as much uncertainty as possible (Section 10.3.4.4).
• Tracing back differences in projections to the representation of fundamental processes, for example, by using physical climate storylines (Sections 10.3.4.2 and Box 10.2) or sensitivity simulations (Section 10.3.2.3).
• Producing physical-climate storylines (Box 10.2) to explore uncertainties not sampled by available model ensembles (Shepherd et al., 2018), for example in pseudo-global warming experiments (Section 10.3.2.2), or to simulate events that have never happened before but are nevertheless plausible (Lin and Emanuel, 2016).
• Attributing observed changes to different external forcings and internal drivers (Section 10.4.1).
• Comparing observed trends with past simulated trends in order to constrain projections with, for instance, the Allen–Stott–Kettleborough method (Allen et al., 2000; Stott and Kettleborough, 2002; Stott et al., 2013) to explain drivers of past observed trends (Section 10.4.2) for understanding future trends.
• Integrating present-day performance via emergent constraints to reduce projection uncertainty (Section 10.3.2).
• Complementing the observational and model-based sources with expert judgement (e.g., integrating knowledge from theory or experience that is available from experts or the literature; Section 10.5.1).

These approaches often can be used in combination to increase confidence in conclusions drawn (Hewitson et al., 2017).

Awareness and understanding of the users’ decision-making context is a central and key aspect of developing tailored, context-appropriate information (Briley et al., 2015), as clearly evidenced by the climate services’ experiences (e.g., Vincent et al., 2018). Understanding the context, however, is not trivial and requires understanding of both the user and provider (Guido et al., 2020) if the information is to be robust, reliable and relevant (Giorgi, 2020). Translating the information into context requires consideration of terminology and expectations (Briley et al., 2015), issues of user interpretation (Daron et al., 2015), and hence necessitating engagement in co-production with all attendant challenges (Vincent et al., 2021). The actual provision of climate information may be conducted at different levels of sophistication, ranging from generic data provision via web portals (Hewitson et al., 2017), potentially including impact-relevant climate indicators, region-specific factsheets and stakeholder reports, social media (Pearce et al., 2019), to a close engagement with specific stakeholders in co-exploring the research (Steynor et al., 2016).

Climate information products may often lack explanations of their potential use and misuse (Street, 2016; Lamb, 2017; Chimani et al., 2020). This is particularly important if the information is provided as a generic, publicly accessible product without a specific context (Hewitson et al., 2017). Context-specific collaboration, especially if organized in workshop, enables a close transdisciplinary co-exploration of the results as in the form of climate risk narratives (Jack et al., 2020, Box 10.2). Such approaches explicitly account for the user context, values and non-climatic stressors (Steynor and Pasquini, 2019).

The transdisciplinary interaction with stakeholders has been categorized into top-down, bottom-up and interactive approaches (Berkhout et al., 2013). Traditional top-down approaches frame the research from the perspective of global climate change as a driver of regional climate risk. Bottom-up approaches, also referred to as scenario-neutral impact studies (Prudhomme et al., 2010; A. Brown et al., 2012; C. Brown et al., 2012; Culley et al., 2016) begin with the user’s articulation of vulnerability in the context of climatic and non-climatic stressors, follow with the definition of key system thresholds of climatic variables, and only incorporate climate data to assess the likelihood of threshold exceedances. Bottom-up approaches are special cases of robust decision-making (Lempert et al., 2006; Lempert and Collins, 2007; Walker et al., 2013; Weaver et al., 2013), which are designed to account for uncertainties not represented by climate models as well as non-climatic stressors. Interactive approaches combine aspects of top-down and bottom-up approaches. The choice of approach depends on the context. While bottom-up approaches might be optimal in a local context, where case-specific risks are addressed, top-down approaches provide generic information that may serve a range of different purposes, for example, at the national scale (Berkhout et al., 2013). All these approaches benefit from the integration of fully distilled climate information (Berkhout et al., 2013; Maraun and Widmann, 2018b).

As implied by (Section 10.5.2, meeting the needs of users can be a substantial challenge for climate scientists if they misunderstand or have limited understanding of user needs and context (Porter and Dessai, 2017). Several barriers in user communities can trigger and sustain this challenge. This can include an institutional aversion to incorporating new tools into decision-making (Callahan et al., 1999). Coincident with this factor, there may be limited staff capacity, lack of management support and lack of a mandate to plan for climate change (Lee and Whitely Binder, 2010).

Following from those challenges, constructing and communicating regional climate information often occurs under the overarching assumption that uncertainty is a problem and reducing uncertainty is the priority (Eisenack et al., 2014; J. Otto et al., 2016). This is both a psychological (Morton et al., 2011) as well as a pragmatic barrier in cases where uncertainty appears to limit the ability to make decisions (Mukheibir and Ziervogel, 2007). However, where in-depth engagements with decision contexts are undertaken, these initial barriers are often dismantled to reveal a more complex, nuanced and potentially more productive intersection with climate information producers that can efficiently handle uncertainty (e.g., Rice et al., 2009; Lemos et al., 2012; Moss, 2016). Specifically, disclosure of all uncertainties in the climate information, transparency about the sources of these uncertainties, and tailoring the uncertainty information to specific decision frameworks have the potential for reducing problems of distilling and communicating uncertain climate information (J. Otto et al., 2016).

There is high confidence that distilling climate information for a specific purpose benefits from a co-production process that involves users of the information, considers the specific user context and the values of relevant actors such as users and scientists, and translates the resultant information into the broader user context. This process allows users to take ownership of the information, builds relationships and trust between information users and producers and helps to overcome barriers in the information construction. This process enhances trust in the information as well its usefulness, relevance, and uptake, especially when the communication involves complex, contextual details (high confidence). The optimal approach for the transdisciplinary collaboration with users depends on the specific context conditioned by the sources available and the actors involved, which together are dependent on the regions considered and the framing by the question being addressed.

Drawing upon multiple lines of evidence in the construction of climate information increases the fitness of this information and creates a stronger foundation (high confidence). The lines of evidence can include multiple observational datasets, ensembles of different model types, process understanding, expert judgement, and indigenous knowledge, among others. Attribution studies, the characterization of possible outcomes associated with internal variability and a comprehensive assessment of observational, model and forcing uncertainties and possible contradictions using different analysis methods are important elements of distillation. To make the most appropriate decisions and responses to changing climate it is necessary to consider all physically plausible outcomes from multiple lines of evidence, especially in the case when they are contrasting such as in the examples of Cross-Chapter Box 10.1 and Section 10.6.2.

Cape Town’s ‘Day Zero’ water crisis in 2018 threatened a shut-down of water supply to 3.4 million inhabitants of the city and resulted in domestic water use restriction of 50 litres per person per day lasting for nine months (pre-drought unconstrained water use was about 170 litres per person per day, DWA, 2013), punitive water tariffs, and temporary closure of irrigation systems. Problems with water supply in many large cities in developing countries are endemic and rarely reported internationally. The water crisis in Cape Town attracted considerable international attention to a city with functional government structures, well-developed services (compared to other urban centres in Africa), a centre of international tourism, and an economic hub with GDP of 22 billion USD (about 7,500 USD per capita, Gallie et al., 2018). Economic and social impacts of the crisis were significant. Loss of revenue for companies of all sizes resulted not only from the scaling down of water-dependent activities, but also from the need to invest in water-efficient technologies and processes. Tourism was affected through reduced arrivals and bookings, although only temporarily (CTT, 2018). In the agricultural sector, 30,000 people were laid-off and production dropped by 20% (Piennaar and Boonzaaier, 2018). The crisis initially polarized society, with conflict emerging between various water users and erosion of trust in the government, but eventually social cohesion and an acute awareness of limited water resources emerged (Robins, 2019).

Cape Town’s crisis resulted from a combination of a strong, rare multi-year meteorological drought (Figure 10.18), estimated at 1 in 300 years (Wolski, 2018), and factors related to the nature of the water supply system, operational water management and water resource policies. Cape Town was very successful in implementing water-saving actions after the previous drought of 2000–2003, reducing water losses from over 22% to 15% (Frame and Killick, 2007; DWA, 2013), breaking the previous coupling of growth in water demand with growth in population. As a consequence, Cape Town won a Water Smart City award from the C40 Cities program only three years prior to the crisis. However, the water-saving actions, together with changing priorities in water resource provision from infrastructure-oriented towards resource and demand management, may well have led to delays in implementation of the expansion of water supply infrastructure (Muller, 2018). The expansion plan, formulated a decade prior to the crisis, included an expectation of long-term climate-change drying in the region (DWAF, 2007). The crisis also exposed structural deficiencies of water management and inadequacy of a policy process in which decisions about local water resources are taken at a national level, particularly in a situation of political tension (Visser, 2018). The crisis was widely seen as a harbinger of future problems to be faced by the city, and a highlight of vulnerability of many cities in the world resulting from the interplay of three factors: (i) the fast urban-population growth, (ii) the economic, policy, infrastructural and water resource paradigms and constraints, and (iii) anthropogenic climate change.

Figure 10.18 | Historical and projected rainfall and Southern Annular Mode (SAM) over the Cape Town region. (a) Yearly accumulation of rainfall (in mm) obtained by summing monthly totals between January and December, with the drought years 2015 (orange), 2016 (red), and 2017 (purple) highlighted in colour. (b) Monthly rainfall for the drought years (in colour) compared with the 1981–2014 climatology (grey line). Rainfall in (a) and (b) is the average of 20 quality controlled and gap-filled series from stations within the Cape Town region (31°S–35°S, 18°W–20.5°W). (c) Time series of the SAM index and of historical and projected rainfall anomalies (%, baseline 1980–2010) over the Cape Town region. Observed data presented as 30-year running means of relative total annual rainfall over the Cape Town region for station-based data (black line, average of 20 stations as in (a) and (b), and gridded data (average of all gridcells falling within 31°S–35°S, 18°W–20.5°W), GPCC (green line) and CRU TS (olive line). Model ensemble results presented as the 90th-percentile range of relative 30-year running means of rainfall and the SAM index from 35 CMIP5 (blue shading) and 35 CMIP6 (red shading) simulations, 6 CORDEX simulations driven by 1 to 10 GCMs (cyan shading), 6 CCAM (purple shading) simulations from individual ensemble members, and 50 members from the MIROC6 SMILE simulations (orange shading). The light blue, dark red and yellow lines correspond to NCEP/NCAR, ERA20C and 20CR, respectively. The SAM index is calculated from sea level pressure reanalysis and GCM data as per Gong and Wang (1999) and averaged over the aforementioned bounding box. CMIP5, CORDEX and CCAM projections use RCP8.5, and CMIP6 and MIROC6 SMILE projections use SSP5-8.5. (d) Historical and projected trends in rainfall over the Cape Town region and in the SAM index. Observations and gridded data processed as in (c). Trends calculated as Theil-Sen trend with block-bootstrap confidence interval estimate. Markers show median trend, bars 95% confidence interval. Global models in each CMIP group were ordered according to the magnitude of trend in rainfall, and the same order is maintained in panels showing trends in the SAM. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

An evaluation of the relative role of rainfall and temperature signal in the 2015–2017 hydrological drought gives a strong indication that lack of rainfall was the primary driver (Otto et al., 2018) leading to the 2018 water crisis. Thus, the remainder of this section focuses on rainfall. Section 11.6 offers a discussion of African drought over broader areas, including mechanisms relevant to them.

Cape Town is located at the south-western tip of Africa, within an approximately 100 km × 300 km region that receives 80% of its rainfall during the austral winter (March to October), with the largest portion in June to August. In the vicinity of Cape Town, rainfall is strongly heterogeneous, ranging from about 300 mm/year on coastal plains to >2,000 mm/year in mountain ranges. The Cape Town water supply relies on surface water reservoirs located in a few small mountain catchments (about 800 km² in total). The Cape Town region receives 85% of its rainfall from a series of cold fronts forming within mid-latitude cyclones. The remainder is brought in by infrequent cut-off lows that occur throughout the year (Favre et al., 2013). This creates a very strong water resource dependency on a single rainfall delivery mechanism that may be strongly affected by anthropogenic climate change (Chapter 4 and Section 10.6.2.6).

The 2015–2017 drought had strong low-rainfall anomalies in shoulder seasons (March to May and September to November, though weaker in the latter), and average rainfall in June and July (Sousa et al., 2018a; Mahlalela et al., 2019). The anomaly resulted from fewer rainfall events and lower average intensity of events. The anomaly was strongest in the mountainous region where the water supply system’s catchments are located (Wolski et al., 2021).

Although the 2015–2017 drought was unprecedented in the historical record, the Cape Town region has experienced other droughts of substantial magnitude, notably in the 1930s, 1970s and more recently in 2000–2003. Long-term (>90 years) rainfall trends are mixed in sign, location-dependent, and weak (Kruger and Nxumalo, 2017; Wolski et al., 2021); mid-term (about 50 years) trends are similarly mixed in sign (MacKellar et al., 2014). In the south-western part of the region, rainfall is mostly decreasing in the post 1981 period, particularly in December–January–February and March–April–May, although there is no trend or a weak wetting in June–July–August (Sousa et al., 2018a; Wolski et al., 2021). Rainfall trends of similar magnitude and duration to the post-1981 trend accompanied previous strong droughts in the region (Wolski et al., 2021).

South Africa and the Cape Town region have good instrumental weather data. Records start in the late 1800s, with in excess of 10 gauges reporting since the 1920s, expanding to about 80 gauges in the 1980s, but the number of stations has declined since. The mountains have only a few stations, which receive more than 1000 mm per year. In view of the strong heterogeneity of rainfall, changes in the number of stations contributing to datasets such as Climatic Research Unit (CRU) and Global Precipitation Climatology Project results in a lack of consistency between them, which limits their reliability in the region (Section 10.2; Wolski et al., 2021).

Because the primary rainfall mechanism is frontal rain, the most relevant large-scale drivers are those that affect cyclogenesis, frontogenesis and the mid-latitude westerlies’ latitudinal position and moisture supply. These drivers and, thus, the region’s rainfall are linked to the Antarctic Oscillation (AAO; Reason and Rouault, 2005) or Southern Annular Mode (SAM), the dominant monthly and interannual mode of Southern Hemisphere atmospheric variability, and a measure of the pressure gradient between mid- and high latitudes. (See Sections 3.3, 3.7, 4.3 and Annex IV.2.2 for more general discussion of the SAM.) While in the post-1930 period, the SAM displays a long-term positive trend, the Cape Town region’s rainfall does not, and only the post-1979 trends of rainfall and SAM are conceptually consistent. For example, a positive trend in the SAM is associated with a negative trend in rainfall (Section 10.6.2.5 and Figure 10.18). There is also good agreement between the seasonality of the SAM and rainfall trends in the post-1979 period: a drying trend appears strongly in December to February and March to May, but not in June to August and September to November (Wolski et al., 2021), and trends in the SAM have similar seasonal dependence (E.-P. Lim et al., 2016; Section 3.7.2). Additionally, there is a similar seasonal pattern in the post-1979 trends in indices capturing the southern edge of the Hadley circulation (Grise et al., 2018).

In the longer-term, Cape Town regional rainfall is characterized by a multi-decadal scale quasi-periodicity (Figure 10.18; Dieppois et al., 2019; Wolski et al., 2021), with the 2015–2017 drought and previous strong droughts (1930s and 1970s) occurring during the rainfall’s periodic low phases. However, the studies linking the Cape Town 2015–2017 drought to the hemispheric processes expressed by the SAM (Sousa et al., 2018a; Burls et al., 2019; Mahlalela et al., 2019) focused almost exclusively on the post-1979 period, when global reanalyses are available. Detailed understanding of the drivers of previous (1930s and 1970s) Cape Town region droughts and the role of hemispheric processes expressed by the SAM in the pre-1979 period is missing.

The Cape Town regional rainfall is also potentially linked to other hemispheric phenomena, such as the expansion of the tropics and, specifically, the South Atlantic high-pressure system and the position of the subtropical jet, which share some variability with the SAM. The relationships between these phenomena and Cape Town rainfall have not been thoroughly investigated outside of the context of the 2015–2017 drought, but the drought itself was associated with poleward expansion of the subtropical anticyclones in the South Atlantic and South Indian oceans and (a resulting) poleward displacement of the moisture corridor across the South Atlantic (Sousa et al., 2018a), as well as a weaker subtropical jet (Mahlalela et al., 2019). Burls et al. (2019) also link the decline in the number of rainy days to the increase in sea level pressure along the poleward flank of the South Atlantic high-pressure system and the intensity of the post-frontal ridging high. Additionally, there is a possible linkage between Cape Town rainfall and near-shore cold sea surface temperature (SST) anomalies arising from Ekman upwelling due to reduced westerly and increased south-easterly winds. These might lead to suppression of convection and reduction of rainfall over land (Rouault et al., 2010). All these phenomena are conceptually consistent with the poleward migration of the westerlies and expansion of the tropics.

Rainfall in the Cape Town region also responds to SST anomalies in the south-east Atlantic, including the Agulhas Current retroflection region, which may drive intensification of low-pressure systems, leading to the trailing front strengthening as it makes landfall over the Cape Town region (Reason and Jagadheesha, 2005). There are also linkages at the seasonal time scale between the Cape Town regional rainfall and Antarctic sea ice (Blamey and Reason, 2007).

In addition to mid-latitude controls, subtropical processes also play a role in the Cape Town region’s rainfall variability. The 10°S–30°S region of the subtropical Atlantic, parts of the South American continent and even parts of the African continent north of Cape Town are sources of moisture for atmospheric river events contributing to frontal rainfall (Blamey et al., 2018; Ramos et al., 2019), with implications for the 2015–2017 drought (Sousa et al., 2018a). Also, the second major rainfall contributing system, cut-off-lows, is conditional on moisture supply from the subtropics (Abba Omar and Abiodun, 2020).

Although El Niño–Southern Oscillation (ENSO) influences climate in southern Africa, any relationship between ENSO and Cape Town’s rainfall is weak and inconsistent, showing the strongest impact in May to June (Philippon et al., 2012). ENSO, however, does influence large-scale processes and phenomena relevant to the drought, though the relationship between ENSO and the SAM is complex, with each ENSO event influencing the SAM differently in different seasons (Ding et al., 2012). Similarly, ENSO affects meridional circulation and thus the subtropical anticyclone as well as the polar and subtropical jets (Seager et al., 2019), but only modifying, not controlling, their role in Cape Town’s rainfall.

Paleoclimate studies reveal that long-term variability in the winter rainfall region of South Africa (including Cape Town) is consistent with a general framework of warming/cooling-induced latitudinal migration of the westerlies and transformation of the subtropical high-pressure belt and associated hemispherical processes (see section 10.2.3.2 for assessment of paleoclimate analysis). The synchronicity of winter rainfall with Antarctic ice-core-derived polar temperature anomalies is consistently revealed in studies using different paleoclimate proxies and time scales of 1400 years (Stager et al., 2012), about 3000 years (Hahn et al., 2016) and 12,000 years (Weldeab et al., 2013). Changes in rainfall regimes at shorter (decadal) time scales appear to reflect influence of local processes such as the Agulhas current’s interaction with the Atlantic, resulting in changes in SST and coastal upwelling, as well as modification of the wind tracks by topography (Stager et al., 2012).

Due to the small scale of the Cape Town region, robust comparison of CMIP simulations to observations is difficult. However, in general, CMIP5 models capture the seasonality well, such as the dominance of austral winter rains, although they overestimate the peak and underestimate the shoulder season rainfall (Mahlalela et al., 2019). Trends in rainfall are particularly difficult to assess as they are generally weak and depend strongly on the time period and dataset adopted for the analyses (Section 10.6.2.3). A multi-method attribution study (Otto et al., 2018) estimates the probability of the 2015–2017 drought to have increased by a factor of three since pre-industrial times (with a wide 95% confidence interval of 1.5 to 6). However, throughout the 20th century, a substantial portion of the global models (about 36% of CMIP5 and 44% of CMIP6 models, as well as many of the MIROC SMILE members) simulate a statistically significant (95% level) decline in total annual rainfall, while there is no robust long-term trend in observations (Figure 10.18). Section 10.4 offers a more detailed assessment of attribution challenges.

Global models capture the overall behaviour of the observed main hemispherical processes, such as the expansion of the tropics, a positive trend in SAM and the poleward shift of the westerly jet. However, they fail to capture details of their observed climatology and variability (Simpson and Polvani, 2016), and the magnitudes of simulated trends vary, though the models typically underestimate observed trends in these processes (Purich et al., 2013; Staten et al., 2018). In general, CMIP5 models do capture the SAM-regional rainfall association, although not consistently across all seasons (Purich et al., 2013; E.-P. Lim et al., 2016).

Global models show strong consistency in a drying signal for the Cape Town region, with the reduction in total annual rainfall of up to 20% by the end of the 21st century in CMIP5 RCP8.5 and CMIP6 SSP5-8.5 simulations (Figure 10.18; Almazroui et al., 2020c). The consistency across the models is a robust signal compared to the rest of southern Africa, where the climate change signal varies spatially: stronger drying in the west and moderate drying or weak wetting in the east (DEA, 2013, 2018; see Atlas.4.4 for further discussion of southern Africa precipitation projections). Rainfall changes projected for the Cape Town region are consistent with projected changes in hemispheric-scale processes and regional-scale dynamics that point toward reduced frequency of frontal systems affecting that region. These changes include robust signals in CMIP5 models for the Southern Hemisphere for a poleward expansion of the tropics (Hu et al., 2013b), poleward displacement of mid-latitude storm tracks (Chang et al., 2012), increased strength and poleward shift of the westerly winds (Bracegirdle et al., 2018) and subtropical jet-streams (Chenoli et al., 2017), and a shift toward a more positive phase of the SAM (E.-P. Lim et al., 2016). However, despite the consistency in circulation changes, the emergence of anthropogenic rainfall change above unforced variability in West Southern Africa remains uncertain for annual rainfall throughout most of the 21st century, even under SSP5-8.5 (Figure 10.15).

There is also a substantial increase in the frequency of conditions supporting atmospheric rivers and water vapour transport towards the south-west coast of southern Africa in the projected climate (Espinoza et al., 2018). This behaviour has strong implications for the region, as most topographically high locations receive rainfall from persistent atmospheric rivers (Blamey et al., 2018). A thorough understanding of the role of atmospheric rivers in the Cape Town region under a changing climate is missing.

Dynamical downscaling studies implemented with a stretched-grid model (Engelbrecht et al., 2009) revealed a signal compatible with the driving CMIP5 ensemble, that is, consistent drying throughout the region, amplifying in time, irrespective of the considered emissions scenario and the generation of global models (DEA, 2013, 2018). A multi-model CORDEX ensemble indicates a robust signal of reduction of total annual rainfall in the future, although there is less agreement on how changes in rainfall occurrence may evolve in the region, such as through fewer consecutive rain days or longer dry spells (Abiodun et al., 2017; Maúre et al., 2018). For the end of the century under RCP8.5, Dosio et al. (2019) also found drying. Moreover, in their analysis, the drying is associated with an increase in the number of consecutive dry days and a reduction in number of rainy days. Their results are consistent with the driving global models for all the precipitation indices, and they are robust independent of the choice of the regional climate model (RCM) or global model. However, collectively, these analyses indicate that uncertainty remains in the characteristics of the precipitation decrease.

There is a consistency in rainfall projections with the projections of rainfall drivers and with the general understanding of the influence of global warming on the circulation dynamics and rainfall patterns in the region. Thus, the expansion of the South Atlantic high-pressure system, related to widespread warming of the tropics and poleward shift of the subsiding limb of the Hadley cell, is associated with the southward displacement of the subtropical jet, and southward migration of mid-latitude westerlies and storm tracks, in addition to changes in the SAM (Section 10.6.2.4). These effects are also relatively consistent with recent (post-1980s) declines in rainfall in the Cape Town region. The storyline of an extended drought is thus a set of events that can yield reduced rainfall in the Cape Town region: a poleward shift of the downward branch of the Hadley cell that produces a sustained southward shift in mid-latitude westerlies and storm tracks. The behaviour is potentially reinforced by changes in the SAM.

There is high agreement among observational data and reanalyses that the recent (post-1979) downward trend in the Cape Town region’s rainfall leading to the 2015–2017 drought is related to the hemispheric processes of poleward shift in the westerlies and expansion of the Hadley circulation. However, there is less support for the precipitation–circulation relationship in historical CMIP5 and CMIP6 simulations. As a consequence, there is only medium confidence that these process changes produced the 2015–2017 drought leading to the 2018 water crisis.

For the water-resource planner who has to deal with potential drought like the 2015–2017 event, several lines of evidence indicate future drying: the projected precipitation by global models and RCMs of different spatial resolutions, and the observed and projected changes of circulation patterns consistent with drier conditions, the paleoclimatic evidence confirming a millennial-scale circulation–rainfall link. However, the distillation is limited by a lack of information about whether or not a relationship between Cape Town precipitation and large-scale circulation processes adequately explains droughts in the twentieth century prior to 1979.

Thus, although a clear association appears in observations from 1979 onward between increasing GHG concentrations, drying in the Cape Town region and behaviour of a key circulation process, the SAM, further analysis suggests caution. Not all global models show the historical post-1979 association among these factors, and when the observational record is extended back further to times when the anthropogenic greenhouse forcing was weaker, there is no strong association between the SAM and Cape Town drought. Thus, there is only medium confidence in the expectation of a future drier climate for Cape Town.

The Indian summer monsoon provides 80% of the country’s annual rainfall from June to September, supplying the majority of water for agriculture, industry, drinking and sanitation to over a billion people. Any variations in the monsoon on time scales from days to decades can have large impacts (Challinor et al., 2006; Gadgil and Gadgil, 2006). Evidence from paleoclimate records (Sections 8.3.2.4.1) shows high confidence in a weakened Indian monsoon during cold epochs of the past such as the Younger Dryas (12,800–11,600 years ago) as measured by speleothem oxygen isotopes (Kathayat et al., 2016). There is a pressing need to understand if the monsoon will change in the future under anthropogenic forcing and to quantify such changes. Multiple datasets have shown robust negative trends since the 1950s until the turn of the century (Bollasina et al., 2011) followed by a recovery (Jin and Wang, 2017), yet repeated assessments project the monsoon to increase in strength under enhanced GHG forcing (Christensen et al., 2007, 2013; Sections 8.3.2.4.1 and 8.4.2.4.1). The apparent contradiction between future projections and observed historical trends makes the region an ideal choice for an in-depth assessment. The reader is also referred to the South Asia (SAS) regional assessment of precipitation extremes (Section 11.9), which is not discussed here for brevity.

Local geography gives rise to distinct differences in societal experience of the summer monsoon. The south-westerly monsoon winds are incident upon the Western Ghats mountains on the west coast, leading to orographic enhancement and heavy rains (Shige et al., 2017), which supply rivers with water for much of the southern peninsula, often the subject of inter-regional water disputes. The northern plains contain the Ganges river and also India’s most intensive agriculture, both rainfed and irrigated. Synoptic systems known as monsoon depressions cross the northern east coast, supplying much of the rain in central India (Hunt and Fletcher, 2019). Further north, the eastern Himalayas are dominated by the summer monsoon, while the western Himalayas receive most rainfall from western disturbances during winter (Palazzi et al., 2013). Meanwhile, south-eastern India sits under a rain shadow (the only region to receive more rainfall during the winter monsoon).

India has one of the oldest rain-gauge networks in the world, leading to the production of numerous observational products (reviewed in Khouider et al., 2020). Gridded gauge-based products dating back to the 19th century reveal pronounced decadal variability (Sontakke et al., 2008). Trends for India over the whole 20th century are inconclusive (Knutson and Zeng, 2018), although declining over central and northern areas (Roxy et al., 2015). Assessment of multiple observational datasets covering the Indian summer monsoon reveals significant declining rainfall over the second half of the 20th century (Section 8.3.2.4.1 and Figure 10.19c,d). A subsequent recovery has been noted since the early 2000s (Jin and Wang, 2017).

Figure 10. 19 | Changes in the Indian summer monsoon in the historical and future periods. Observational uncertainty demonstrated by a snapshot of rain-gauge density (% of 0.05° subgrid boxes containing at least one gauge) in the APHRO-MA 0.5° daily precipitation dataset for June to September 1956. (b) Multi-model ensemble (MME) mean bias of 34 CMIP6 models for June to September precipitation (mm day^–1) compared to CRU TS observations for the 1985–2010 period. (c) Maps of rainfall trends (mm day^–1 per decade) in CRU TS observations (1950–2000), the CMIP6 MME-mean of SSP5-8.5 future projections for 2015–2100 (34 models), the CMIP6 hist-GHG and hist-aer runs, both measured over 1950 to 2000. (d) Low-pass filtered time series of June to September precipitation anomalies (%, relative to 1995–2014 baseline) averaged over the central India box shown in panel (b). The averaging region (20°N–28°N, 76°E–87°E) follows other works (Bollasina et al., 2011; Jin and Wang, 2017; Huang et al., 2020b). Time series are shown for CRU TS (brown), GPCC (dark blue), REGEN (green), APHRO-MA (light brown) observational estimates and the IITM all-India rainfall product (light blue) in comparison with the CMIP6 mean of 13 models for the all-forcings historical (pink) the aerosol-only (hist-aer, grey) and greenhouse gas-only (hist-GHG, blue). Dark red and blue lines show low-pass filtered MME-mean change in the CMIP6 historical/SSP5-8.5 (34 models) and CMIP5 historical/RCP8.5 (41 models) experiments for future projections to 2100. The filter is the same as that used in Figure 10.11 (d). To the right, box-and-whisker plots show the 2081–2100 change averaged over the CMIP5 (blue) and CMIP6 (dark red) ensembles. Note that some models exceed the plotting range (CMIP5: GISS-E2-R-CC, GISS-E2-R, IPSL-CM5B-LRl and CMIP6: CanESM5-CanOE, CanESM5 and GISS-E2-1-G). (e) Precipitation linear trend (% per decade) over Central India for historical 1950–2000 (left) and future 2015–2100 (right) periods in Indian Monsoon rainfall in observed estimates (black crosses), the CMIP5 historical-RCP8.5 simulations (blue), the CMIP6 ensemble (dark red) for historical all-forcings experiment and SSP5-8.5 future projection, the CMIP6 hist-GHG (light blue triangles), hist-aer (grey triangles) and historical all-forcings (same sample as for hist-aer and hist-GHG, pink circles). Ensemble means are also shown. Box-and-whisker plots show the trend distribution of the three coupled and the d4PDF atmosphere-only (for past period only) SMILEs used throughout (Chapter 10 and follow the methodology used in Figure 10.6. (f) Example spread of trends (mm day^–1 per decade) for the period 2016–2045 in RCP8.5 SMILE experiments of the MPI-ESM model, showing the difference between the three driest and three wettest trends among ensemble members over central India. All trends are estimated using ordinary least-squares regression. Further details on data sources and processing are available in the chapter data table (Table 10.SM.11).

Open figure

Observational products containing critical inhomogeneities in gauge distribution and reporting over time are acknowledged as suitable for mesoscale analysis (Rajeevan and Bhate, 2009), while use for climate trends requires consistent reporting over time from quality-controlled gauges (e.g., about 2000 gauges since the 1950s in Rajeevan et al., 2006). A newer 0.25°-gridded product covering 1901 onwards (Pai et al., 2014, 2015), based on Shepard’s interpolation method for irregularly-spaced stations (Shepard, 1968), shows increased intensity of daily rainfall and extremes over some regions, especially in the late-20th century. However, changes to the inputted gauges may have introduced an artificial jump in extreme rainfall since 1975 over central India (Section 10.2.2.3; Lin and Huybers, 2019). They suggest that this method may have masked declines in mean rainfall and highlight the need for availability of raw gauge data to allow transparent assessments. Khouider et al. (2020) have successfully tested a probabilistic interpolation method for India to overcome problems inherent in algorithms based on inverse-distance weighting when applied to data-sparse regions. An example snapshot of the uneven distribution of rain gauges in a common observational product is shown in Figure 10.19a.

The uncertainty among local and international observational products for India can pose challenges when evaluating climate models (as in Section 10.2.2.6; Prakash et al., 2015). For the seasonal mean summer monsoon rainfall, Collins et al. (2013a) found large biases separating many CMIP5 models from the available observational products. However, for seasonal mean variability, the spread across observational products was larger than across the CMIP5 ensemble.

The relevant drivers for long-term change in the mean Indian summer monsoon are summarized briefly:

• Increased greenhouse gas (GHG) concentrations (chiefly CO₂) are a strong contributor to changes in the monsoon, with repercussions for the meridional temperature contrast driving the monsoon circulation (Ueda et al., 2006; Roxy et al., 2015), for the monsoon winds in the lower troposphere (Cherchi et al., 2011; Krishnan et al., 2013), or for the availability of moisture from the Indian Ocean (May, 2011).
• Industrial emissions of sulphate aerosol predominantly in the Northern Hemisphere could change inter-hemispheric energy transports and weaken the monsoon (Polson et al., 2014; Undorf et al., 2018). The effect of local anthropogenic emissions of black carbon (chiefly from cooking fires) is uncertain (Lau and Kim, 2006; Nigam and Bollasina, 2010).
• India’s green revolution over the late-20th century led to considerable land-use change, with massive expansion of agriculture at the expense of forest and shrublands. As a result, India’s northern plains feature widespread irrigation, suggested to be a cause of drying (Mathur and AchutaRao, 2020).
• Decadal modes of variability such as the Pacific Decadal Variability (PDV, Annex IV) and Atlantic Multi-decadal Variability (AMV, Annex IV), which may be partly forced (Section 3.7.7), are known to cause decadal modulation of the monsoon (Krishnamurthy and Krishnamurthy, 2014; Naidu et al., 2020).

The interplay of these external and internal drivers is key to understanding past and future monsoon change.

The robust decline of Indian summer monsoon rainfall averaged over India in the second half of the 20th century (Section 10.6.3.3) is not in line with expectations arising from thermodynamic constraints on the water cycle in a warming world (Section 8.2.2) and has been regarded as a puzzle (Goswami et al., 2006). Assessing the attribution of 20th-century changes to Indian rainfall is the subject of coordinated modelling under the Global Monsoon MIP (GMMIP; Zhou et al., 2016), but is complicated by long-standing dry biases in coupled CMIP3, CMIP5 (Sperber et al., 2013) and CMIP6 (Figure 10.19b) global models. These dry biases are connected to a lower tropospheric circulation that is too weak (Sperber et al., 2013) and wet biases in the equatorial Indian Ocean (Bollasina and Ming, 2013). Section 8.3.2.4.1 finds high confidence that anthropogenic aerosol emissions have dominated the observed declining trends of countrywide Indian summer monsoon rainfall, consistent with findings at the global-monsoon scale (Section 3.3.3.2).

Stronger Northern Hemisphere aerosol emissions cool it relative to the Southern Hemisphere, increasing northward energy transport at the expense of moisture transport towards India (Bollasina et al., 2011). The attribution to anthropogenic aerosols is supported in CMIP5 single-forcing experiments, including some testing the sensitivity to local and remote emissions (Guo et al., 2015, 2016; Shawki et al., 2018), comparing CMIP5 GCMs forced by both aerosol and GHG to GHG only (Salzmann et al., 2014) and reducing emissions to pre-industrial levels (Takahashi et al., 2018). The large spread between individual model realisations of comparable magnitude to the aerosol-induced signal suggested to Salzmann et al. (2014) that internal variability may also play a role over regions such as northern-central India. Further uncertainty surrounds the level of radiative forcing. Dittus et al. (2020) forced a GCM with historical aerosol emissions scaled between 0.2 and 1.5 times their observed values, representing the spread in CMIP5 effective radiative forcing. The strongest forcing led to around 0.5 mm day^–1less late-20th century Indian monsoon rainfall than the weakest (Shonk et al., 2020). Meanwhile, the uncertainty surrounding aerosol–cloud interactions could change the sign of long-term precipitation trends (Takahashi et al., 2018).

There is some evidence that declining Indian monsoon rainfall is due to regional SST warming patterns, themselves arising due to radiative forcing from GHG (e.g., in the Indian Ocean, Guemas et al., 2013). Roxy et al. (2015) artificially raised SST in a GCM in the equatorial Indian Ocean (the region of strongest observed SST warming), leading to a weakened monsoon. Annamalai et al. (2013) used a GCM to suggest instead that preferential warming of the western North Pacific may force a Rossby-wave response to its west that weakens the monsoon through dry advection and subsidence. These hypotheses are not borne out in GHG-forced future projections (Section 10.6.3.6).

A small anthropogenic contribution may be expected from local land-use/land-cover changes and land management. India is the world’s most irrigated region with around 0.5 mm/day in places, although peaking higher in summer (Cook et al., 2015b; McDermid et al., 2017). Including irrigation in GCMs and RCMs slows the monsoon circulation and diminishes rainfall (Lucas-Picher et al., 2011; Guimberteau et al., 2012; Shukla et al., 2014; Tuinenburg et al., 2014; Cook et al., 2015b) due to reduced surface temperature (Thiery et al., 2017), reducing the monsoon wind and moisture fluxes towards India (Mathur and AchutaRao, 2020). However, implementation methodologies for irrigation in climate models are simplified and often do not account for spatial heterogeneity while overestimating demand and supply (Section 10.3.3.6; Nazemi and Wheater, 2015; Pokhrel et al., 2016). Changing forest cover to agricultural land in an RCM (Paul et al., 2016) finds weakened summer monsoon rainfall especially in central and eastern India, due to decreased local evapotranspiration. Decreased evapotranspiration from a warmer surface since the 1950s in the CMIP5 ensemble may also feedback on the supply of moisture (Ramarao et al., 2015). Based on an AGCM study and literature review, Krishnan et al. (2016) support the role of land-use/land-cover change in adding to the effects of aerosol in weakening the monsoon, in addition to dynamic effects on the circulation caused by rapid warming of the Indian Ocean.

In addition to anthropogenic forcing, there is evidence that internal variability in the Pacific is a significant driver. Huang et al. (2020b) compared a large perturbed-physics ensemble (HadCM3C) with a SMILE for the historical period. Ensemble members replicating the negative Indian rainfall trend were accompanied by a strong phase change in the PDV from negative to positive, consistent with SST observations. Jin and Wang (2017) have demonstrated increasing Indian monsoon rainfall since 2002 in a variety of observed datasets, suggesting the increase is due either to a change in dominance of a particular forcing (for example from aerosol to GHG) or to a change in phase of internal variability such as the PDV. Huang et al. (2020b) partially attribute the rainfall recovery to a phase change in the PDV, supported by a SMILE study combined with reanalyses (Ha et al., 2020).

The drying trend of Indian summer monsoon rainfall since the mid-20th century can be attributed with high confidence to aerosol as the dominant anthropogenic forcing with a further contribution from internal variability, supported by the review of B. Wang et al. (2021) including CMIP6 results. Understanding the interplay between anthropogenic and internal drivers will be important for understanding future change.

The AR5 (Christensen et al., 2013) concluded that Indian summer monsoon rainfall will strengthen under all RCP future climate scenarios, while the circulation will weaken (medium confidence). SR1.5 (Hoegh-Guldberg et al., 2018) found onlylow confidence in projections of monsoon change at 1.5°C and 2°C, or any difference between them. The AR6 assessment of (Chapter 8 (Section 8.4.2.4.1) finds more precipitation in future projections (also depicted in Figure 10.19c,d,e), supported by reviews of CMIP3, CMIP5 and CMIP6 models (Turner and Annamalai, 2012; Kitoh, 2017; Z. Chen et al., 2020; B. Wang et al., 2021).

Given the assessment for a future wetter monsoon dominated by GHG emissions and attribution of the late-20th century decline to aerosol (Sections 8.3.2.4.1 and 10.6.3.5), the change between dominant forcings will lead, at some point, to a positive trend. For example, RCP4.5 experiments in an AGCM forced by coupled model-derived future SSTs showed continuation of 20th-century drying, before a rainfall recovery (Krishnan et al., 2016). By holding aerosol emissions at 2005 levels, lower monsoon rainfall is found throughout the 21st century than in a standard RCP8.5 scenario (Zhao et al., 2019), suggesting that the timing of the recovery will be partially controlled by the rate at which aerosol emissions decline. The spread in spatial distribution of aerosol emissions in SSPs may also play a role in near-term projections (Samset et al., 2019). Under divergent air-quality policies, SSP3 features a dipole of declining sulphate emissions for China but increases over India, leading to suppression of GHG-related precipitation increases there (Wilcox et al., 2020). For the near-term future around the mid-21st century, the interplay between internal variability and external forcing must be considered (Singh and AchutaRao, 2019). Huang et al. (2020a) used two SMILEs to show that internal variability related to PDV could potentially overcome the GHG-forced upward trend in Indian monsoon rainfall, consistent with assessments of the global monsoon for the near term (Section 4.4.1.4). Emergence of the anthropogenic signal for South Asian precipitation is shown from the 2050s onwards in CMIP6 (Figure 10.15b).

In long-term projections, robust signals consist of a weakened upper-tropospheric meridional temperature gradient, either due to upper-level heating over the tropical Pacific (Sooraj et al., 2015) or Indian oceans (Sabeerali and Ajayamohan, 2018) in CMIP5, and increased seasonal mean rainfall, including in CMIP6 (Almazroui et al., 2020b; B. Wang et al., 2021). The weakened temperature gradient combines with increased atmospheric stability to weaken the monsoon overturning circulation, with some findings showing northward movement of the lower-tropospheric monsoon winds in response to a stronger land–sea temperature contrast in CMIP3 and CMIP5 (Sandeep and Ajayamohan, 2015; Endo et al., 2018). The northward shift was also found in the genesis of synoptic systems (monsoon depressions) in a single high-resolution AGCM forced by an ensemble of SSTs derived from four GCMs under the RCP8.5 scenario (Sandeep et al., 2018).

Projections can also be expressed in terms of global-mean warming levels (GWLs) rather than time horizons (Cross-Chapter Box 11.1). Advancing on SR1.5, amplification of mean and extreme monsoon rainfall at 2.0°C compared to 1.5°C has been found both by an AGCM forced by future SST patterns (Chevuturi et al., 2018) and by using time slices in CMIP5 GCMs (Yaduvanshi et al., 2019; J. Zhang et al., 2020). These findings are consistent with the general scaling of Indian monsoon precipitation per degree of warming in CMIP5 (Zhang et al., 2019) and CMIP6 (B. Wang et al., 2021). Increasing GWLs also lead to emergence of the anthropogenic signal over larger proportions of the South Asian region (Figure 10.15a).

Decomposition of the increased rainfall signal showed that while the dynamic component led to a drying tendency, this was overcome by the thermodynamic contribution (Sooraj et al., 2015; Z. Chen et al., 2020). Alternative decomposition experiments using AGCMs and their coupled counterparts found increases in the lower-tropospheric temperature gradient and monsoon rainfall to be dominated by the fast radiative response to GHG increase rather than SST changes (Li and Ting, 2017; Endo et al., 2018). The response to SST forcing featured a large model spread, particularly arising from the dynamic component (Li and Ting, 2017). Chen and Zhou (2015) found that the Indo-Pacific SST warming pattern dominated the uncertainty in Indian monsoon rainfall change. Finally, in assessing the relative impact of CO₂ radiative forcing and plant physiological changes in quadrupled CO₂ experiments in four Earth system models, Cui et al. (2020) showed little impact of plant physiology on annual rainfall for the Indian region.

While several of the above studies selected model subsets to constrain future projections based on standard performance metrics of the historical period, such as pattern correlation and root-mean-square error, Latif et al. (2018) included a performance measure based on agreement with historical rainfall trends. This is an unproven constraint for regional projections (Section 10.3.3.9), since the 20th-century rainfall trend over India is assessed to have been driven chiefly by aerosol and other factors such as PDV (Sections 8.3.2.4.1 and 10.6.3.5), while the dominant late-21st century forcing is GHG emissions. Modern emergent-constraint techniques (Section 10.3.4.2) are being applied to the Indian monsoon such as G. Li et al. (2017), who found that models with excessive tropical western Pacific rainfall tend to project a greater Indian monsoon rainfall change in future, due to an exaggerated cloud-radiation feedback. Correcting for this bias reduces the future change.

In summary, long-term future scenarios dominated by GHG increases (such as the RCPs) suggest increases in Indian summer monsoon rainfall (high confidence), dominated by thermodynamic mechanisms leading to increases in the available moisture. In the near-term, there is high confidence (medium agreement, robust evidence) that increased rainfall trends due to GHGs could be overcome by aerosol forcing or internal variability.

Coordinated monsoon-relevant dynamical downscaling efforts such as CORDEX South Asia (Gutowski Jr. et al., 2016; Choudhary et al., 2018) are relevant to the Indian summer monsoon, first with assessment of their added value (Section 10.3.3.2 and Atlas.5.3.3). Singh et al. (2017) compared nine CORDEX-South Asia RCMs against their driving CMIP5 GCMs, for present-day rainfall patterns and processes related to intra-seasonal variability. They found no consistent improvement other than for spatial patterns (e.g., rainfall close to better-resolved orography); some characteristics were made worse. Both the rainfall pattern and its bias were worsened in CORDEX compared to CMIP5 in Mishra et al. (2018). In contrast, Varikoden et al. (2018) found improved representation of historical rainfall patterns, such as over the Western Ghats mountains (consistent with Singh et al., 2017), reducing the dry bias; improvements were not found over the northern plains, which are dominated by synoptic variability known as monsoon depressions. Similarly, Sabin et al. (2013) compared a uniform 1° resolution model ensemble with another zoomed to about 35 km over South Asia. Local zooming improved simulation of orographic precipitation and the monsoon trough. For the future, a surrogate approach (like pseudo-global warming, see Section 10.3.2.2) was used in an RCM to test the impacts of warming or moistening on monsoon depressions (Sørland and Sorteberg, 2016; Sørland et al., 2016). The depressions are found to give more rainfall in future, dominated by strengthened synoptic circulation from the warming perturbation. By forcing an RCM with a perturbed parameter ensemble of a GCM, Bal et al. (2016) made projections under SRES A1B for the 2020s, 2050s and 2080s. They noted increases in rainfall of 15–24% for India. Finally, evidence from a single CORDEX South Asia RCM showed a mixed signal for changes in peak season rainfall under RCP2.6 and RCP8.5 (Ashfaq et al., 2021).

Statistical downscaling and other post-processing require calibration in historical conditions (e.g., Akhter et al., 2019) and assessment of fitness-for-purpose (Section 10.3.3.9) before use for future projections. Given the noted biases in GCM monsoon simulation (Section 10.6.3.5), Vigaud et al. (2013) used a variant of quantile mapping to bias adjust (Section 10.3.1.3.2 and Cross-Chapter Box 10.2) GCM outputs. For the historical period, the pattern, mean and seasonal cycle of rainfall versus the input GCMs were improved. Increased future monsoon rain, albeit in older SRES A2 projections, was found for southern India. Salvi et al. (2013) used regression-based perfect prognosis (Section 10.3.1.3.1) for the whole country at 0.5° resolution based on five ensemble members of a GCM in SRES scenarios. They noted increases over rainy regions of west coast and north-east India, but decreases in the north, west and south-east. Madhusoodhanan et al. (2018) statistically downscaled 20 CMIP5 models to 0.05° resolution. While the global models projected increased rainfall, the downscaled ensemble depicted both increasing and decreasing trends, indicating uncertainty. However, key physical processes operating at below-GCM scale cannot be resolved nor calibrated for, such as aspects of the flow around topography. This is notably an issue given the resolution disparity between the driving global models and output, and the regional challenges in observational data used for calibration (Section 10.6.3.3).

There are mixed messages as to whether downscaling adds value to climate projections of the Indian summer monsoon; however, there is high confidence in projections of precipitation changes in orographic regions given the consistent improved representation in these regions among several dynamical downscaling studies.

Formal storyline approaches (see Box 10.2) have been used infrequently for the Indian summer monsoon, representing a knowledge gap. In an expert-elicitation approach (Dessai et al., 2018), physically plausible futures substantiated by climate processes were constructed, focusing on a river basin in southern India. Possible outcomes were framed based on changes in two drivers: availability of moisture from the Arabian Sea and strength of the low-level flow. The narratives identified were able to explain 70% of the variance in monsoon rainfall over 1979–2013, the implication being that climate uncertainties could be easily communicated to stakeholders in the context of present-day variability.

The storylines terminology could be used to loosely describe the interplay between internal variability and forced change (see Section 10.6.3.6), such as considering the difference between groups of wettest and driest ensemble members of a SMILE for the near-term future in Figure 10.19f. However, given the interest in low-likelihood high-impact scenarios (Sutton, 2018), we can also consider possible storylines for the Indian monsoon constructed from evidence in paleoclimate records and modelling. For example, a future AMOC collapse could cause reduced monsoon rainfall (Section 8.6.1; Liu et al., 2017), offsetting increases expected due to GHG. Large tropical volcanic eruptions are also known to weaken the Asian summer monsoon, in observations and model simulations over the last millennium (Section 8.5.2.3; Zambri et al., 2017), although a hemispheric dependence is found, with Southern Hemisphere eruptions even strengthening the monsoon around India (Zuo et al., 2019). Typically, future climate projections do not consider plausible eruption scenarios and their mitigating effects on greenhouse warming (see also Cross-Chapter Box 4.1). A single-model ensemble (Bethke et al., 2017) demonstrates a future drier Indian monsoon relative to conditions in which volcanic eruptions are not considered, although the effects of GHG warming dominate beyond the mid-term.

The few studies on low-likelihood high-impact scenarios, often in single models, together with findings in SR1.5 (Hoegh-Guldberg et al., 2018), noting the small radiative forcing in 1.5°C or 2°C scenarios, or the absence of large aerosol emissions at the end of the 21st century in RCPs, give us low confidence in abrupt changes to the monsoon on this time scale.

Above, we presented assessments from observational and model attribution studies of the historical period, followed by future climate projections in global and regional models, and storylines approaches including low-likelihood high impact events. Miscellaneous lines of evidence are considered here.

Our assessment could also be informed by attempting to constrain future projections of the Indian summer monsoon using paleoclimate evidence. In modelling work of the mid-Holocene (D’Agostino et al., 2019), the increased obliquity (axial tilt) and altered orbital precession lead to an enhanced monsoon with a stronger dynamic component (strengthening the mean monsoon overturning) controlling the increase in monsoon rainfall. In future climates however, the dynamic contribution decreases (Section 10.6.3.6), yet the increased thermodynamic component (greater moisture availability) overcomes this to cause a wetter monsoon. Monsoon changes under different epochs may not be governed by the same mechanisms (D’Agostino et al., 2019; Hill, 2019), making the mid-Holocene, in particular, unsuitable as a period to compare with.

Finally, the recent national climate-change assessment for India (Krishnan et al., 2020) has distilled multiple lines of evidence to show declining summer monsoon rainfall over the second half of the 20th century, attributable to emissions of anthropogenic aerosols, while future projections informed by CMIP5 modelling and dominated by GHG forcing show increased mean rainfall by the end of the 21st century.

There is very high confidence (robust evidence, high agreement) of a negative trend of summer monsoon rainfall over the second half of the 20th century averaged over all of India. There is medium agreement over trends at the regional level owing to uncertainty among observational products, which hinders model evaluation, downscaling and assessment of changes to extremes. There is high confidence (robust evidence, medium agreement) that anthropogenic aerosol emissions over the Northern Hemisphere and internal variability have contributed to the negative trend, while there is high confidence (robust evidence, medium agreement) that Indian summer monsoon rainfall will increase at the end of the 21st century in response to increased GHG forcing, due to the dominance of thermodynamic mechanisms. No contradictory evidence is found from downscaling methods. The contrast between declining rainfall in the observational record and long-term future increases can be explained using multiple lines of evidence. They are not contradictory since they are attributable to different mechanisms (primarily aerosols and greenhouse gases, respectively). The long-term future changes are generally consistent across global (including at high resolution) and regional climate models, and supported by theoretical arguments. Furthermore, while there are subtle differences found in past periods with a climate similar to the future climate (the mid-Holocene), different physical mechanisms at play suggest that paleoclimate evidence does not reduce confidence in the future projections. In the near term, there is high confidence that internal variability will dominate.

The Mediterranean region is loosely denoted as the region that surrounds the Mediterranean Sea, and it is characterized by complex orography and strong land–sea contrasts. The region contains a dense and growing human population, with large regional differences: whereas the population of the European Mediterranean countries has been relatively stable or even declining during the past decades, the population of countries in Mediterranean areas of the Middle East and North Africa has quadrupled between 1960 and 2015, and the degree of urbanization has risen from 35 to 64% during the same period (Cramer et al., 2018) and during the more recent period 2000–2020 the urban expansion rate has exceeded 5% (Kuang et al., 2021).

The Mediterranean region has experienced significant climate variability over recent decades and has been affected in particular by severe heatwaves and droughts (Sections 8.3, 11.3, 11.6 and 12.4; Hoegh-Guldberg et al., 2018). Increasing summer temperatures will enhance the frequency and intensity of such extreme events and will cause additional environmental and socio-economic pressure on the region.

The Mediterranean has a heterogeneous climate that is partly semi-arid, especially along the southern coast of the Mediterranean Sea (Lionello et al., 2012). It is characterized by mild humid winters and dry warm or hot summers, which are associated with large scale subsidence that is partly related to the downward branch of the Hadley circulation. Other factors affecting the Mediterranean circulation include the monsoon heating over Asia (Rodwell and Hoskins, 1996; Cherchi et al., 2014; Ossó et al., 2019) and circulation anomalies induced by topography (Simpson et al., 2015). Seasonal and interannual variability is strongly linked to natural modes of variability (Section 10.6.4.4). The Mediterranean Sea acts as an evaporation source that dominates the regional hydrological cycle, which is characterized by local cyclogenesis and a separate branch of the mid-latitude storm track (Lionello et al., 2016). It also affects remote locations such as the Sahel (Park et al., 2016; Section 10.4.2.1). Strong storms can develop over the Mediterranean. Among these, Medicanes are particularly destructive and exhibit several similarities with tropical cyclones (Cavicchia et al., 2014; Kouroutzoglou et al., 2015; Gaertner et al., 2018). The Mediterranean region is also characterized by strong land-atmosphere coupling and feedbacks (Seneviratne et al., 2006) generating prolonged droughts and intense heatwaves, which can also affect continental Europe (Zampieri et al., 2009). Other aspects of Mediterranean climate include regional winds, which can be very strong due to the channelling effect (Obermann et al., 2018) and extreme rainfall during autumn (Ducrocq et al., 2014; Ribes et al., 2019).

The Mediterranean region spans a wide variety of countries and economies. This has led to large differences in the existence and availability of observational records, with the southern part of the area being sparsely covered by meteorological stations (Figure 10.20b). Consequently, basin-wide, homogeneous, quality controlled observational datasets are lacking, especially before the advent of substantial satellite observations in the 1970s. Observational uncertainties exist also for those regions that are covered by high quality networks such as European Climate Assessment & Dataset (ECA&D; Flaounas et al., 2012).

Large differences of up to 7°C between the CRU and UDEL datasets have been found especially over mountainous areas, such as the Atlas in Morocco (Zittis and Hadjinicolaou, 2017; Strobach and Bel, 2019). Bucchignani et al. (2016a, b) compared three different datasets (CRU, UDEL, and MERRA) with the available ground observations and found that although the geographical distribution of the bias is qualitatively similar for the three datasets, differences exist, with the absolute bias being generally lower in Modern-Era Retrospective Analysis for Research and Applications (MERRA) especially over North Africa during the summer and winter season. There is high confidence that the sparse monitoring network in parts of the Mediterranean region strongly increases the uncertainty across different gridded datasets (Section 10.2.2.3, Figure 10.20b,c).

The Mediterranean summer climate is affected by large-scale modes of natural variability, the most dominant being the NAO (Annex IV) in winter and the summer NAO in summer (Folland et al., 2009; Bladé et al., 2012), although regional differences exist. The influence of those modes of variability over the eastern Mediterranean is recognized by some studies (Chronis et al., 2011; Kahya, 2011; Black, 2012; Bladé et al., 2012), but disputed by others (Ben-Gai et al., 2001; Ziv et al., 2006; Donat et al., 2014; Turki et al., 2016; Zamrane et al., 2016; Han et al., 2019). During positive summer NAO phase, associated with an upper-level trough over the Balkans, the Mediterranean is anomalously wet (Bladé et al., 2012). Drivers of Mediterranean climate variability include modes of variability such as the AMV (Sutton and Dong, 2012) and the Asian monsoon (Rodwell and Hoskins, 1996; Logothetis et al., 2020). In addition, the increase of GHGs (e.g., Zittis et al., 2019), the decrease of anthropogenic aerosols over Europe and the Mediterranean since the 1980s resulting from air pollution policies (Turnock et al., 2016), and anthropogenic land-use change (Millán, 2014; MedECC 2020) have been shown to be linked to the regional warming. The role of the zonal averaged circulation as a driver for the Mediterranean climate has been stressed by (Garfinkel et al., 2020).

The attribution of observed Mediterranean summer warming to above drivers and implications for future projections will be discussed in Sections 10.6.4.5 and 10.6.4.6.

Observational datasets show large agreement on the historical (1960–2014) temperature evolution at basin-wide scale (Figure 10.20e), with an enhanced warming since the 1990s, and the early decades of the 21st century being on average approximately more than 1°C warmer than late 19th century levels (van der Schrier et al., 2013; Cramer et al., 2018; Lionello and Scarascia, 2018; Figure 10.20e). Over recent decades, the surface air temperature of the Mediterranean including the Mediterranean Sea has warmed by around 0.4°C per decade (Macias et al., 2013). Observed trends over land show large geographical heterogeneity (Figure 10.20d) and notable differences exist amongst different datasets at grid point scale (Figure 10.20c; Qasmi et al., 2021).

Several mechanisms have been proposed for the enhanced Mediterranean warming, although their relative importance and the possible interplay between them are not fully understood. Circulation changes might have contributed to this enhanced warming (Figure 10.20a). Sutton and Dong (2012) argued that the AMV induced a shift around the 1990s towards warmer southern European summers. This mechanism is associated with a linear baroclinic atmospheric response to the AMV-related surface heat flux. Also O’Reilly et al. (2017) related warm summer decades to the AMV, but the connection was shown to be mainly thermodynamic. Qasmi et al. (2021) estimate an increase in Mediterranean summer temperature of 0.2°C–0.8°C during a positive AMV.

Increased warming over land compared to the sea is expected due to the lapse-rate changes associated with tropospheric moisture contrasts (Kröner et al., 2017; Byrne and O’Gorman, 2018; Brogli et al., 2019b; Figure 10.20a). Enhanced land–sea temperature contrast leads to relative humidity and soil moisture feedbacks (Rowell and Jones, 2006), the latter also depending on weather regimes (Quesada et al., 2012). The globally enhanced land–sea contrast in near surface temperature is also a robust result in CMIP5 and CMIP6 models (Section 4.5.1.1).

Due to its semi-arid climate, strong atmosphere–land coupling has contributed to the larger increase of mean summer temperature compared to the increase of annual mean temperature (Seneviratne et al., 2006). In particular, during drought spells, limits to evaporation due to low soil moisture provide a positive feedback and enhances the intensity of heatwaves (Lorenz et al., 2016; Box 11.1). By comparing reanalysis-driven RCM simulations with observations, Knist et al. (2017) found that RCMs are able to reproduce soil moisture interannual variability, spatial patterns, and annual cycles of surface fluxes over the period 1990–2008, revealing a strong land–atmosphere coupling especially in southern Europe in summer. In addition cloud feedbacks can modulate the Mediterranean summer temperature (Mariotti and Dell’Aquila, 2012).

The observed trends over 1901–2010 are outside the range of internal variability shown in CMIP5 pre-industrial control experiments and consistent with, or greater than those simulated by experiments including both anthropogenic and natural forcings (Knutson et al., 2013) and therefore partly attributable to anthropogenic forcing. The decrease of anthropogenic aerosols over Europe including the Mediterranean resulting from European de-industrialisation and air pollution policies (Turnock et al., 2016) has been highlighted as an important contributor to the observed warming (Ruckstuhl et al., 2008; Philipona et al., 2009; de Laat and Crok, 2013; Nabat et al., 2014; Besselaar et al., 2015; Dong et al., 2017; Boé et al., 2020a). Pfeifroth et al. (2018) argue that this brightening is mainly due to cloud changes caused by the indirect aerosol effect with a minor role for the direct aerosol effect, in contrast to Nabat et al. (2014) and Boers et al. (2017) who attribute it to the direct aerosol effect. Using model sensitivity experiments, Nabat et al. (2014) also associated the increase in Mediterranean SST since 1980–2012 with the decrease in aerosol concentrations (Atlas.8.2, Atlas.8.3 and Atlas.8.5).

Over the period 1960–2014, observed trends over land are consistent with those of most of the multi-model or SMILEs ensembles (Figure 10.20f), although large differences exist for individual models and ensemble members. The modelled ensemble-mean trends show large geographical variations. Generally, both global and regional models often underestimate the observed trend especially over parts of North Africa, Italy, the Balkans and Turkey. The cold bias in global models is related to simulated SLP trends that are anti-correlated to the observed trend, which is probably due to systematic model errors (Boé et al., 2020b). Biases in the simulation of soil-moisture and cloud-cover might also have contributed to the underestimation of the warming trend in GCMs (van Oldenborgh et al., 2009). The CORDEX results (at both 0.44° and 0.11° resolution) show consistently smaller values than those in global models and the available datasets (Figure 10.20g; Vautard et al., 2021). This is partly due to the overestimation in the temperature evolution before 1990 (Figure 10.20e), possibly because of differences in the aerosol forcing (Boé et al., 2020a), although the driving global models also have a cold bias (Vautard et al., 2021). Cold biases for recent decades are also found in Med-CORDEX simulations (Dell’Aquila et al., 2018) and by RCM simulations over the southern part of the Mediterranean, Middle East and North Africa region (Almazroui, 2016; Almazroui et al., 2016a, b; Zittis and Hadjinicolaou, 2017; Ozturk et al., 2018), although higher resolution, new bare soil albedo and modified aerosol parametrization significantly improve the results (Bucchignani et al., 2016a, b, 2018). Despite large differences in the multi-model mean trend (Figure 10.20g), in most of the land points the observed trend lies within the model range in all ensembles. For the SST bias exhibited by coupled RCMs the choice of driving global model has the largest impact (Darmaraki et al., 2019; Soto-Navarro et al., 2020).

The Mediterranean is expected to be one of the most prominent and vulnerable climate change hotspots (Diffenbaugh and Giorgi, 2012). CMIP5, CMIP6, HighResMIP and CORDEX (Section 10.6.4.7) simulations all project a future warming for the 21st century that ranges between 3.5°C and 8.75°C for RCP8.5 at the end of this century for those ending at 2100 (Figure 10.21a, b). CMIP6 results project more pronounced warming than CMIP5 for a given emissions scenario and time period (Figure 10.21c; Coppola et al., 2020). However, when analysing the Mediterranean warming in terms of mean global warming levels, the two ensembles largely agree, showing that summer warming is projected to reach values up to 40–50% larger than the global annual warming, largely independent of models and emissions scenarios (Figure 10.21d). Large regional differences exist, with enhanced warming projected over Turkey, the Balkans, the Iberian Peninsula and North African regions (Figures 10.14a, 10.21c; Almazroui et al., 2020a) and reaching, locally, values of up to double the global mean (Lionello and Scarascia, 2018). The enhanced summer warming also increases the amplitude of the seasonal cycle (Yettella and England, 2018).

Figure 10.20 | Aspects of Mediterranean summer warming. (a) Mechanisms and feedbacks involved in enhanced Mediterranean summer warming. (b) Locations of observing stations in E-OBS and Donat et al. (2014). (c) Differences in temperature observational datasets (NOAA Global Temp, Berkeley Earth, CRUTEM4 and GISTEMP) with respect to E-OBS for the land points between the Mediterranean Sea and 46°N and west of 30°E. (d) Observed summer (June to August) surface air temperature linear trends (°C decade^–1) over the 1960–2014 period from Berkeley Earth. (e) Time series of area averaged Mediterranean (25°N–50°N, 10°W–40°E) land point summer temperature anomalies (°C, baseline 1995–2014). Dark blue, brown and turquoise lines show low-pass filtered temperature of Berkeley Earth, CRU TS and HadCRUT5, respectively. Orange, light blue and green lines show low-pass filtered ensemble means of HighResMIP (4 members), CORDEX EUR-44 (20 members) and CORDEX EUR-11 (37 members). Blue and red lines and shadings show low-pass filtered ensemble means and standard deviations of CMIP5 (41 members) and CMIP6 (36 members). The filter is the same as the one used in Figure 10.10. (f) Distribution of 1960–2014 Mediterranean summer temperature linear trends (°C decade^–1) for observations (black crosses), CORDEX EUR-11 (green circles), CORDEX EUR-44 (light blue circles), HighResMIP (orange circles), CMIP6 (red circles), CMIP5 (blue circles) and selected SMILEs (grey box-and-whisker plots, MIROC6, C