Future Extreme Precipitation in Summer Will Become More Widespread in China Depending on Level of Warming

Abstract

In this study, CMIP6 and single model initial-condition large ensemble (SMILE) simulations were applied to reveal possible changes in the spatial coverage (RSC) of summer extreme precipitation (≥99th percentile) in China. Four different approaches were applied to provide a comprehensive assessment of RSC changes. Results indicate that the trend of RSC in summer for the period 1961–2020 in China is underestimated by raw simulations of CMIP6 multi-model ensembles. Further analyses suggest that the confidence in model simulations reflecting the observed change can be improved based on the two observation constraint approaches and SMILE-based approach. In addition, the reliability of the spatial distribution can also be improved. Projection results indicate that the RSC of summer extreme precipitation increases consistently with the increment of global warming across the different approaches. Among them, the results based on observational constraints and the SMILE-based approach show enhanced reliability and present larger RSC changes than by directly using CMIP6 ensembles. In conclusion, extreme precipitation in summer is expected to become more widespread in China as the level of global warming increases.

Introduction

Frequency and intensity are essential indicators for extreme precipitation and have thus been studied extensively in recent decades. IPCC AR6 (IPCC, 2023) stated that it is a fact that human-induced greenhouse gas emissions have led to an increased frequency and intensity of some weather and climate extremes. In addition, every additional half a degree of global warming will cause discernible increases in the intensity and frequency of heavy precipitation (Schleussner et al., 2016). On the regional scale, extreme precipitation is projected to increase in some areas, including northern Europe, East Asia, North America, and southern Africa (Cardell et al., 2020; Dosio et al., 2019; Torres-Alavez et al., 2021; Yu, Zhai, & Lu, 2018). Besides its frequency and intensity, extreme precipitation has shown a more widespread occurrence in China under the background of global warming (Chen & Sun, 2021), but the potential future changes in the spatial coverage of extreme precipitation are less well studied or understood compared to those in frequency and intensity. It is important to address this knowledge gap to achieve a more comprehensive understanding of the changes in extreme precipitation toward providing additional information for risk assessment.

Usually, projecting extreme events on a global scale is achieved by directly applying a global climate model (GCM) (Mitchell et al., 2016). However, model biases are always present at regional scales because some of the forcings and processes cannot be accurately described. For example, certain variables in a particular region increasing with warming is a well-established fact, but their actual changes may be more severe or less evident (Brunner et al., 2020; Ribes et al., 2021; Zhou & Zhang, 2021). Thus, observational constraints are recommended for GCM-based regional projections. Thus far, multiple approaches have been adopted for observation constraints (IPCC, 2023). One of the most established is the weighted average method, wherein the performances and independence of GCMs are considered, and then those models with high scores are selected to calculate the weighted average (Knutti et al., 2017; Sillmann et al., 2013). By adopting this approach, the uncertainty of extreme events from individual models is efficiently reduced, and the related results outperform any single-model or multi-model average (Brunner et al., 2020; T. Li, Jiang, Zhao, et al., 2021).

Meanwhile, the attribution-based constraint has been developed in recent attribution–projection studies as more results from DAMIP (Detection and Attribution Model Intercomparison Project) experiments have become available from the historical to future-projection time range (Gillett et al., 2016). In this approach, attribution is first conducted using Allen, Stott, and Kettleborough method (Allen et al., 2000; Kettleborough et al., 2007; Stott & Kettleborough, 2002). Then, attribution-based constraint is achieved through scaling factor from the attribution analysis. By applying the attribution-based approach, simulation outputs can be constrained according to the response of the model simulation to anthropogenic and natural forcing (Dong et al., 2020; Ribes et al., 2021). Thus, projected extreme events can be constrained by physical interpretability (Chen & Sun, 2021; C. Li et al., 2020). Moreover, a single model initial-condition large ensemble (SMILE) can be applied to problems of regional projection, which brings the advantages inherent to working with a large ensemble size (Merrifield et al., 2019). It is well-known that the contribution of model uncertainty to projection uncertainty is relatively significant, especially in near-term projection (Lehner et al., 2020). By adopting a large ensemble simulation approach, the projection uncertainty at the regional scale introduced by the multi-model approach can be avoided (Zhou et al., 2020).

As outlined above, different methods have their own special applicability. However, previous studies have tended to only use one approach to conduct projections, with relatively few having used multiple evidence chains simultaneously. Thus, in this study, to comprehensively assess future changes in less well-studied properties of extreme events, the observation-constraint and SMILE-based approaches were applied to project the possible future changes in the spatial coverage of extreme precipitation in China. Following this introduction, Section 2 describes the data and methods employed for assessing the changes in the spatial coverage of extreme precipitation; Section 3 presents the related results, including a model evaluation, attribution analysis, and projection of the spatial coverage of extreme precipitation; and finally, Section 4 summarizes our conclusions and provides some further discussion.

Data and Methods Data

The daily in-situ observed precipitation data of 2,400 meteorological stations from 1961 to 2020 in China provided by the National Meteorological Information Center, China Meteorological Administration (CMA), were employed in this study for evaluation of GCM historical simulations (W. Xu et al., 2013). A quality control procedure that considered the availability of data and station relocations (Yu & Zhai, 2020) was applied to improve the reliability of the result. Ultimately, after this procedure, the precipitation data of 1,250 stations were employed.

The daily precipitation outputs from the sixth phase of the Coupled Model Intercomparison Project (CMIP6) data archive were obtained and 26 of the historical simulations (Historical or ALL forcings) from 1850 to 2014 and Shared Socioeconomic Pathway scenarios with intermediate greenhouse gas (GHG) emissions (SSP2-4.5) from 2015 to 2100 were applied (O’Neill et al., 2016). In addition, the monthly global surface air temperature outputs of Historical and SSP2-4.5 from CMIP6 were obtained to calculate the time ranges for global warming of 1.0, 1.5, and 2°C above pre-industrial level. We used one simulation per model (r1i1p1), and the models employed were ACCESS-CM2, ACCESS-ESM1-5, BCC-CSM2-MR, CESM2-WACCM, CMCC-CM2-SR5, CMCC-ESM2, EC-Earth3, EC-Earth3-Veg, EC-Earth3-Veg-LR, FGOALS-g3, GFDL-ESM4, INM-CM4-8, INM-CM5-0, IITM-ESM, IPSL-CM6A-LR, KACE-1-0-G, KIOST-ESM, MIROC6, MIP-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NESM3, NorESM2-LM, NorESM2-MM, and TaiESM1.

In addition, the daily precipitation outputs from DAMIP (Gillett et al., 2016) were applied. Specifically, 49 historical simulations under GHG (Hist-GHG), aerosol (Hist-AER), and natural (Hist-NAT) forcings for the period 1850–2020 (detailed information given in Table 1) were used. Furthermore, 250 chunks of non-overlapping 60-year (1961–2020) simulations from the pre-industrial control (CTL) experiment were obtained to estimate the internal variability. Moreover, the daily precipitation outputs from 50 simulations of initial-condition large ensembles of CanESM5 with historical simulations from 1850 to 2014 and SSP2-4.5 simulations from 2015 to 2100 were applied (Swart et al., 2019).

Table 1 List of the Model Simulations From DAMIP Used in This Study

Model name	Hist-GHG (run)	Hist-AER (run)	Hist-NAT (run)
CanESM5	10	10	10
MIROC6	3	3	3
NorESM2-LM	3	3	3
ACCESS-CM2	3	3	3
ACCESS-ESM1-5	2	2	2
BCC-CSM2-MR	3	3	3
CNRM-CM6-1	3	3	3
FGOALS-g3	3	3	3
GFDL-ESM4	1	1	1
HadGEM3-GC31-LL	4	4	4
IPSL-CM6A-LR	9	9	9
MRI-ESM2-0	5	5	5
Total	49	49	49

Note. Numbers indicate ensemble sizes for historical simulations under greenhouse gas (Hist-GHG), aerosol (Hist-AER), and natural (Hist-NAT) forcings.

Methods Ratio of Spatial Coverage

In this study, the 99th percentile daily precipitation in summer during 1961–1990 was defined as extreme precipitation with valid precipitation greater than 1 mm per day. As known, observation stations located in China differ in their spatial density, and model simulation grids also have different resolutions. To determine the most suitable interpolation resolution, one consideration is to ensure each interpolated grid contains at least one valid model grid. Thus, the resolution of 3° × 3° was selected. Then, the ratio of spatial coverage (RSC) index was introduced to represent the area of influence for extreme precipitation.

For the observational data, inside a 3° × 3° grid box (i.e., grid i with l_i stations), if k_i,j out of l_i stations is hit by extreme precipitation in year j, the RSC_obs_i,j in the grid box is regarded as k_i,j/l_i. Grid boxes without a station are regarded as missing values. The method applied to build the observed time series of RSC for extreme precipitation in China (RSC_o,j) is as follows: [Image Omitted. See PDF]Here, i refer to the grid, j refers to the year, and n is the total number of valid grids in the observation at 3° × 3° resolution.

Also, the same method was applied to the model simulations. Specifically, if k_m,i,j out of l_m,i grids inside the 3° × 3° grid i had extreme precipitation in summer for year j based on model m, the RSC_model_m,i,j was regarded as k_m,i,j/l_m,I, where l_m,i is the total number of grids in grid i for model m. Then, the simulations were masked according to the availability of observational grids. The time series of RSC for extreme precipitation in China for model m (RSC_m,j) is as follows: [Image Omitted. See PDF]Here, i refer to the grid, j refers to the year, and m refers to the model.

Detection and Attribution

To detect and attribute possible contributions of external forcings, the regularized optimal fingerprinting detection and attribution method was applied in this study (Allen & Stott, 2003; Ribes et al., 2013). This method regresses the observations onto the signals of the external forcings using the total least squares algorithm, as follows: [Image Omitted. See PDF]where y is the observation, x represents the multi-model-simulated signal pattern of the external forcings, v refers to the influence of natural internal variability in the modeled signal patterns, and ε denotes the residual term representing the climate internal variability estimated from the CTL simulations. The regression coefficient β is the scaling factor and if the 90% confidence interval of β does not include 0, the corresponding modeled signals can be detected in the observation. In addition, if the interval contains 1 in their confidence intervals, then the modeled signals can be attributed in the observations. Scaling factors reveal amplitude information about the respective forcing.

In this study, single-signal, two-signal, and three-signal detection and attribution analyses were conducted. For single-signal analysis, the observed changes were regressed onto the external forcing response to ALL forcings; for two-signal analysis, the observed changes were regressed onto the responses to anthropogenic forcing (ANT, equal to ALL minus NAT) and NAT forcing; and for three-signal analysis, the observed changes were regressed onto the responses to GHG, AER, and NAT. A residual consistency test was also applied to test whether the residual term agreed well with the internal variability from the model simulations (Allen & Stott, 2003; Ribes et al., 2013). Moreover, the 250 chunks of non-overlapping 60-year (1961–2020) CTL simulations were divided into two halves—one used to estimate the scaling factors, and the other used for the residual consistency test. If the P-value was larger than 0.1, then the residual consistency test was passed, meaning the regression residuals were consistent with the internal variability.

Definition of Warming Levels

Twenty-year time slices with the respective mean global warming for each model were used to reflect the impacts of different warming levels (Schleussner et al., 2016). This transient approach has been widely used in previous studies (e.g., Yu, Zhai, & Chen, 2018; Yu, Zhai, & Lu, 2018). As shown in IPCC AR6 (IPCC, 2023), the best estimate of the observed warming from the pre-industrial level (1850–1900) to 1995–2014 is approximately 0.85°C. Thus, global mean surface temperature increases of 0.15, 0.65, and 1.15°C relative to 1995–2014 are regarded as global warming levels of 1.0, 1.5, and 2°C above pre-industrial level, respectively.

Different Projection Approaches

In this study, four approaches were used to produce multiple possible future changes in the RSC of extreme precipitation. First, the multi-model ensemble average of CMIP6 outputs was directly applied to represent simulations without bias correction (hereafter referred to as Historical for the historical period and SSP2-4.5 for the projected period). Second, previous studies have indicated that bias-corrected simulations based on the weighted average of selected models with strong simulation performance and good skill in terms of model independence offer a slight improvement in eastern China (T. Li, Jiang, Le Treut, et al., 2021). In this study, only models that could reasonably reveal the trend of observations were selected to conduct equal weighted averaging (referred to as Historical_Weighting for the historical period and SSP245_Weighting for the projected period). It is known that simulation biases exist for estimating precipitation trends in China among CMIP6 models, with some even showing an opposite trend (Figure 1d). Thus, finally, 20 out of 26 models with a positive trend of RSC in China were selected to conduct equal weighted averaging. These models were ACCESS-CM2, BCC-CSM2-MR, CanESM5, CESM2-WACCM, EC-Earth3, EC-Earth3-Veg, FGOALS-g3, KACE-1-0-G, MIROC6, MIP-ESM1-2-HR, MPI-ESM1-2-LR, MRI-ESM2-0, NorESM2-LM, NorESM2-MM, EC-Earth3-Veg-LR, KIOST-ESM. The remaining six models were not considered.

Figure 1. Spatial distribution of changes in the RSC of extreme precipitation from 1961–1990 to 1995–2014 (%) for (a) observation and (b) Historical (blank squares indicate the masked areas).

Third, thanks to different external forcing radiation simulations of DAMIP, another available approach is to constrain model simulations based on attribution with a clear underlying physical mechanism (referred to as Historical_Attribution for the historical period and SSP245_Attribution for the projected period). Finally, possible simulations based on SMILE were also available, benefiting from the initial-condition large ensembles. Here, large ensemble simulations from CanESM5 were applied to show the result from a single model (referred to as Historical_SMILE for the historical period and SSP245_SMILE for the projected period).

Thus, historical and projected changes were combined with multiple lines of evidence to assess the possible future changes in the RSC for extreme precipitation in China. It is worth noting that the period 1961–2020 was regarded as the historical period in the model results to be consistent with the observation period. Meanwhile, the period 2021–2100 in the model results was considered as the projection period.

Results CMIP6 Model Simulations

Figure 1a presents the spatial distribution of RSC from 1961–1990 to 1995–2014 from observation and the multi-model ensemble. Observationally, the difference in RSC in southeastern China is the largest, especially in the middle to lower reaches of the Yangtze and Zhujiang rivers. Meanwhile, it is relatively small in North China and Southwest China, where even negative trends are apparent. However, the spatial distribution of changes in RSC given by the multi-model ensemble is different from the observations (Figure 1b). The multi-model ensemble shows the RSC of extreme precipitation in Northeast China and Northwest China as being larger than that in southeast China. In other words, the spatial distribution of the changes in RSC given by CMIP6 is inconsistent with observation, with the model–observation difference being particularly large for southeast China and North China.

The simulation of RSC trends is a good indicator for model evaluation in this study. The observed changes in the RSC of extreme precipitation exhibit substantial temporal variations (Figure 2). In terms of the temporal evolution of RSC in the model simulation (Figure 2), the uncertainty range of the multi-model ensemble covers the range of variation in the observation series, indicating that the model can capture the observed variability. However, for the trend, the observation is 1.0%/10a, while the multi-model ensemble average is only 0.5%/10a.

Figure 2. Time series of the RSC of extreme precipitation in China relative to 1961–1990 (%), wherein the black and blue lines refer to the time series from observation and the multi-model average of Historical, respectively. The blue shading around the blue line denotes the full uncertainty range of model simulations used in this study. The orange (gray) lines indicate models with a positive (negative) trend with a 5-year running mean.

The trend and variability of the RSC time series reflected by the multi-model average are weaker than those observed, possibly because of China's location in the monsoon region, as well as the complex thermodynamic conditions, which together lead to considerable model uncertainty. Thus, the multi-model average may result in some offsetting among different models and lead to a weak trend. Also, some studies have pointed out that the increased anthropogenic aerosol forcing could have an adverse effect on enhancing extreme precipitation in China, especially in southeast China (H. Xu et al., 2022), which implies that the decline in anthropogenic aerosols and their precursors since 2006 in China may have increased the occurrence of extreme precipitation. However, this decrease in aerosols since 2006 is not accurately reflected in GCMs, which may also contribute to the weak trend in the CMIP6 multi-model ensemble. Thus, the possible future changes of RSC in China may not be accurately reflected using the raw simulation results of CMIP6, which suggests that a constrained estimate is required for the regional climate to provide a bias-corrected projection.

Detection and Attribution of RSC Changes

Figure 3 shows the detection and attribution results for the changes in RSC. For ALL forcing, the 90% confidence interval of the scaling factor β does not contain 0 and contains 1 based on single-signal analysis. This means that ALL signals in the model can be detected and attributed in the observation. Similar analysis is shown for ANT (ALL-NAT) and NAT signals based on two-signal detection analysis, indicating that ANT can be detected but NAT signals cannot in China. Estimates of scaling factors based on three-signal detection analysis are further provided when GHG, AER, and NAT signals are considered simultaneously. As shown, GHG, AER, and NAT signals cannot be detected since the confidence interval of the scaling factor β contains 0. Meanwhile, the best estimates of the ALL scaling factor in China are larger than unity, indicating that the ALL response based on DAMIP outputs underestimates the observed changes. Thus, through the results of multiple detection and attribution analyses, only the ALL signals are detectable and attributable for the RSC changes in China.

Figure 3. Best estimates of the scaling factor β (data points) and their 5%–95% uncertainty range (error bars) for the RSC of extreme precipitation in China based on single-signal detection analysis of ALL (red), two-signal detection analysis of ANT (ALL-NAT, orange) and NAT (green), and three-signal detection analysis of GHG (purple), AER (blue), and NAT (green). The solid and dashed gray horizontal lines denote zero and unity, respectively.

A possible reason for the detection and attribution results may be related to the model's regional simulation ability. GCMs may not accurately capture some processes on the regional scale, which may be affected by multiple factors. Thus, detection and attribution of human influence can be confirmed at the hemispheric scale and in continental regions, but with increasing difficulty in smaller regions (Zhang et al., 2013). In addition, for China, another reason may relate to the inaccurate simulation of aerosol scenarios by the CMIP6 model simulation (Wang et al., 2021), which could lead to the opposite spatial distribution of precipitation (less in South China and more in North China). This might be why the three-signal were not successfully detected and attributed in China from the observations.

In conclusion, the intensified RSC of extreme precipitation can be attributed to human influence dominated by ALL forcing from the perspective of physical interpretability. Thus, we applied the SSP2-4.5 scenarios of CMIP6 to reflect future projections based on attributable forcing with precise physical meaning.

Projected RSC for Different Warming Levels

Besides directly using the multi-model ensemble average, three more approaches were applied to provide full evidence of the changes and related uncertainty ranges of RSC. As shown in Figure 4, the observed trend of RSC (1.0%/10a) is roughly double that of Historical (0.5%/10a) based on the approach of directly using the multi-model ensemble average of CMIP6. This highlights that CMIP6 can reflect but underestimate the changes of RSC in China. Additionally, the RSC trend is improved when the three other approaches are applied. In particular, the RSC trend is 0.8%/10a based on Historical_Weighting, that is, using models that accurately reveal the observed trend; it is 0.8%/10a based on Historical_Attribution, that is, using attributable forcing for the observation; and it is 1.9%/10a based on Historical_SMILE, that is, the large ensemble output based on a single model initial condition.

Figure 4. Time series of the changes in the RSC of extreme precipitation relative to 1961–1990 based on historical simulations from (a) Historical_Weighting, (b) Historical_Attribution, and (c) Historical_SMILE in China for the period 1961–2020 (%), wherein the red line denotes the ensemble average, and the black and blue lines refer to the series of RSC changes from observation and the multi-model ensemble average of Historical, respectively.

Next, the changes of RSC in China at different warming levels were projected based on the four approaches. As shown in Figure 5, the increases in RSC relative to 1961–1990 in China are about 1.5%, 2.0%, 4.6%, and 7.4% for global warming levels of 0.85, 1.0, 1.5, and 2.0°C based on the CMIP6 SSP2-4.5 multi-model ensemble average. The increase in RSC from 1961–1990 to 1995–2014 (approximately 0.85°C relative to the pre-industrial level (1850–1900)) is about 1.5% (Figure 5). The increase is amplified by 1.3, 3.0, and 4.8 times of magnitude when global warming increases to 1.0, 1.5, and 2°C, respectively, which implies that the RSC in China will enhance with an increase in the level of global warming. More areas will experience extreme precipitation in the future.

Figure 5. Projected RSC of extreme precipitation based on SSP2-4.5, SSP245_Weighting, SSP245_Attribution, and SSP245_SMILE in China for the warming levels of 0.85, 1.0, 1.5, and 2.0°C relative to 1961–1990. The width of the boxes represents the 25%–75% uncertainty range of the projected RSC, and the whiskers extend to the full uncertainty range (5%–95% uncertainty range for SSP245_Attribution based on the uncertainty ranges of the scaling factor β).

Moreover, from the results based on SSP245_Weighting, the RSC in China is about 2.5%, 3.1%, 5.9% and 8.8% for global warming levels of 0.85, 1.0, 1.5, and 2.0°C, respectively. The increase in RSC from 1961–1990 to 0.85°C is about 2.5%, and the increase is amplified by 1.3, 2.4, and 3.5 times of magnitude when global warming increases to 1.0, 1.5, and 2°C. Furthermore, the RSC is about 2.5%, 3.3%, 7.5% and 12.1% based on SSP245_Attribution, and about 5.4%, 6.8%, 9.5% and 12.1% based on SSP245_SMILE, for global warming levels of 0.85°C, 1.0°C, 1.5°C and 2.0°C, respectively. The RSC increases 2.5% and 5.4% from 1961–1990 to 0.85°C based on these two approaches, and these increases are amplified by 1.3 and 1.3 times of magnitude when global warming increases to 1.0°C; by 3.0 and 1.8 times of magnitude when it increases to 1.5°C; and 4.8 and 2.3 times of magnitude when it increases to 2°C, based on SSP245_Attribution and SSP245_SMILE. Generally, across different approaches, it is shown that a higher warming level will lead to a larger RSC in China.

Notably, RSC results based on constrained projection and SMILE are more extensive than SSP2-4.5 projection directly applied to the raw CMIP6 multi-model ensemble average. As reflected in Figure 5, the value based on SSP2-4.5 is smaller than the value based on SSP245_Weighting, SSP245_Attribution, and SSP245_SMILE for a specific global warming level. The CMIP6 multi-model ensemble average may somewhat underestimate the future changes in the RSC of extreme precipitation in China. Therefore, based on observation constraints and SMILE-based projection, more areas in China will experience extreme precipitation in the future.

Of interest here is which projected RSC is more accurate and reliable. Therefore, the reliability of projections based on the four different approaches was further analyzed (Figure 6). Here, to quantify the reliability of one possible approach, the changes of RSC in 1995–2014 minus 1961–1990 were first calculated for each grid. In observation (Figure 1a), the RSC in China shows an increase in southeast China and northwestern China, while there is a decrease in southwestern China and North China. This distribution pattern is consistent with precipitation changes in China. Actually, it is currently impossible for GCMs to accurately capture all the changes in every grid. Thus, the percentage of the grid that can capture the observed sign was calculated to judge the reliability of the different constraint approaches. Here, the underlying assumption is that an approach with better model performance in the historical period will show more reliability in terms of future projection.

Figure 6. Spatial distribution of the changes in the RSC of extreme precipitation (1995–2014 minus 1961–1990): (a) Historical_Weighting, (b) Historical_SMILE (blank squares indicate the masked areas).

As shown in Figure 1b, the spatial distribution of RSC changes in Historical is somewhat different from that based on observation. It presents a decrease in southeast China but an increase from southwestern China to northeastern China. In the Historical approach, the corresponding percentage of grids capturing the observed sign is 48.3%, which may be partly due to the simulation bias in southeast China and North China. As discussed above, the ability to reveal the trend of RSC in China still needs some improvement for some CMIP6 model outputs. Subsequently, a similar analysis was conducted for the other three approaches to see if their reliability showed improvement. For Historical_Weighting, a similar spatial distribution of RSC changes is presented based on selected CMIP6 models (Figure 6a). As shown, the percentage of grids capturing the observed sign is 54.0%. Although the value is limited, the improvement mainly manifests in poorly simulated southeast China compared with the Historical approach. This implies that the reliability is improved in Historical_Weighting based on selected models.

The distribution of RSC changes in Historical_Attribution is the same as Historical because they are both based on ALL forcing model output. Meanwhile, RSC changes in Hist-GHG were further analyzed to find if simulations based on Hist-GHG only, not including aerosol, could have a better reliability. In general, its spatial distribution shows a consistent increase in China. The corresponding percentage of grids is 64.4%, which implies that the reliability in Hist-GHG is greatly enhanced, especially in southeast China, where extreme precipitation is significantly enhanced according to observations for the period 1961–2020. In terms of the RSC based on the Historical_SMILE approach (Figure 6b), its spatial distribution of RSC changes shows an increase in northwestern China and southeast China. Moreover, the decreasing sign as observed is revealed for some grids in Southwest China. The percentage of grids capturing the observed sign is 63.2%, meaning that the model performance of the large initial condition ensembles is good in China. As shown in Figure 6b, the improvement is mainly in South China.

In summary, some simulation biases exist for RSC in China based on the CMIP6 multi-model ensemble average. Specifically, the observed changes in RSC show an increase in southeast China but a decrease in North China. However, the changes based on the CMIP6 multi-model ensemble average illustrate a decrease in southeast China and an increase in North China. Then, based on the Historical_Weighting and Historical_SMILE approaches, the confidence increases, especially in southeast China. Combined with the projected RSC as mentioned before, it is concluded that more of China will be influenced by extreme precipitation than directly projected by the CMIP6 simulations alone.

Discussion and Conclusion

CMIP6 simulation data and a CanESM5 large initial condition ensemble were applied to explore the possible future changes in the RSC of summer extreme precipitation (≥99th percentile) at different global warming levels. Four approaches were used to deliver a comprehensive projection result. The key findings are summarized as follows.

Due to the biases of regional simulation and the too high aerosol forcing simulation in GCMs for China, the spatial distribution of the RSC given by the CMIP6 models has a certain deviation, and the trend of the RSC for summer extreme precipitation is underestimated. Therefore, the direct use of CMIP6 multi-model projection results seems to include certain biases for future projection. Furthermore, using the constraint approaches and SMILE-based approach, the confidence to reflect the RSC trend and the reliability of the RSC spatial distribution in China are greatly enhanced. Related projection results show that the RSC for summer extreme precipitation will also increase as the global warming level enhances. In addition, compared with the results from direct projection with the CMIP6 multi-model ensemble, the other three approaches show a larger area of China will face summer extreme precipitation in the future.

Nevertheless, the emergent constraint concept was not applied in this study. An emergent constraint on an Earth system sensitivity requires an emergent relationship between the sensitivity and some measure of variation in the contemporary climate (Eyring et al., 2019). However, robust and physically explainable statistical relations are not observed for RSC changes in China at the present time. Moreover, anthropogenic aerosols and their precursors could have considerable influence on key regional systems (e.g., monsoons, the westerly jet, western Pacific subtropical high, etc.) and variables (e.g., temperature, precipitation.) (Dong et al., 2019, 2022; Wang et al., 2021; H. Xu et al., 2022). Thus, providing accurate simulations of aerosol influence as an input in GCMs is necessary and vital for climate change studies at regional scales.

Acknowledgments

We acknowledge the Program for Climate Model Diagnosis and Intercomparison and the World Climate Research Program's Working Group on Coupled Modeling for making the WCRP multi-model dataset available (https://www.wcrp-climate.org/wgcm-cmip). In addition, we thank the China National Meteorological Information Center for providing the observational data (http://data.cma.cn). This research is supported by the National Natural Science Foundation of China (41905083), the Joint Funds of the National Natural Science Foundation of China (U2142205), the Joint Open Project of KLME & CIC-FEMD, NUIST (KLME202309).

Data Availability Statement

The daily precipitation outputs of CMIP6 are available at https://esgf-node.llnl.gov/search/cmip6/. Here, the data used in this study are obtained by selecting Experiment ID historical and ssp245, Variant Label r1i1p1f1, Frequency day, and Variable pr. The daily precipitation outputs from Detection and Attribution Model Intercomparison Project are available at https://esgf-node.llnl.gov/search/cmip6/. The data are obtained by selecting Experiment ID hist-GHG, hist-aer, hist-nat, ssp245-GHG, ssp245-aer, and ssp245-nat, Frequency day, and Variable pr. The SMILE simulations are available at https://esgf-node.llnl.gov/search/cmip6/. The data are obtained by selecting Source ID CanESM5, Experiment ID historical, ssp245, hist-GHG, hist-aer, hist-nat, Frequency day, and Variable pr. The daily in-situ precipitation dataset of 2,400 meteorological stations from 1961 to 2020 in China is provided by the National Meteorological Information Center, CMA (W. Xu et al., 2013). Figures are generated using NCAR Command Language (The NCAR Command Language (Version 6.6.2), 2019).