Twenty-First Century Drought Analysis Across China using CMIP6 Under Climate Change

14 Under global warming, according to results obtained from offline drought indices driven by 15 projections of general circulation models (GCMs), future droughts in China will worsen but the 16 results are not consistent. We analyzed changes in droughts covering the entire hydrologic cycle 17 using outputs of GCMs of the 6th Coupled Model Intercomparison Project (CMIP6) for SSP2-4.5 18 and SSP5-8.5 climate scenarios, and compared the results with that of popular, offline drought 19 indices (the self-calibrating Palmer Drought Severity Index (scPDSI), Standardized Precipitation 20 Evapotranspiration Index (SPEI) and Standardized Precipitation Actual Evapotranspiration Index 21 (SPAEI)). Among meteorological, agricultural, and hydrological drought indices tested under both 22 SSP scenarios, the results obtained from SPAEI and scPDSI agree better with univariate drought 23 indices than SPEI. scPDSI generally agrees well with agricultural droughts (Standardized Soil 24 Moisture Index with the surface soil moisture content; SSIS). Future droughts estimated using soil 25 moisture analysis are more widespread than that from precipitation and runoff analysis in humid 26 regions of South China by the end of the 21st century. In arid northwestern China and Inner 27 Mongolia, drought areas and severity based on scPDSI and SSIS forced with the SSP scenarios 28 show obvious decreasing trends, in contrast to increasing trends projected in humid regions. Trends 29 projected using SPEI contradict those projected by other drought indices in non-humid regions. 30 Therefore, selecting appropriate drought indices are crucial in project representative future droughts 31 and meaningful information needed to achieve effective regional drought mitigation strategies under 32 climate warming impact.

in the actual evapotranspiration are usually dominated by a change in precipitation rather than in 70 PET (Ayantobo and Wei 2019; Yang et al. 2018a), which could hinder the applicability of SPEI in 71 these regions. Instead of PET, actual evapotranspiration (AET) has been used to compute the 72 Standardized Precipitation Actual Evapotranspiration Index (SPAEI) index, which has been found 73 to show drought conditions more consistent with that of a hydrological drought index than SPEI 74 (Joetzjer et al. 2013). Due to the better physical representation of PDSI, it has been a popular index 75 to assess the severity of historical droughts (Chen et al. 2019) and to project future droughts from 76 outputs of general circulation models (GCMs) (Gizaw and Gan 2016). 77 Outputs of GCMs are widely used for evaluating the effects of climate change on droughts, such 78 as GCMs CMIP5 (the 5 th Coupled Model Intercomparison Project) (Taylor et al. 2012). Based on 79 CMIP5 outputs, many studies indicated that climate warming could intensify droughts (aridity) 80 because the projected increase in water demand (evaporation) generally exceeds the projected 81 increase in water supply (precipitation) ( . Therefore, some studies suggested that directly using actual changes in water cycle variables 91 such as precipitation, runoff, and soil moisture simulated by climate models to project future 92 droughts or aridity trends may be more representative than using offline drought indices. 93 Using multimodel ensemble (MME) simulations of ten GCMs from CMIP6 under SSP2-4.5 and 99 SSP5-8.5 scenarios, we estimated precipitation, runoff, and soil moisture droughts of China, which 100 represent the traditional meteorological, hydrological, and agricultural droughts, respectively. We 101 further compare these three types of droughts with offline drought indices (scPDSI, SPEI and SPAEI) 102 to investigate the applicability of offline drought indices in different climatic regions across China.  Table 1. 114 2.2 CMIP6 model outputs 115 Historical and future climate change scenarios across China simulated by ten CMIP6 GCMs from 116 North America, Europe, and Asia are listed in Table 2. Projected temperatures for future periods for 117 SSP scenarios are SSP1-2.6 (+ 2.6 W m −2 ; low forcing sustainability pathway), SSP2-4.5 (+ 4.5 W 118 m −2 ; medium forcing middle-of-the-road pathway), SSP3-7.0 (+ 7.0 W m −2 ; medium-to high-end 119 forcing pathway), and SSP5-8.5 (+ 8.5 W m −2 ; high-end forcing pathway). For the analysis of future 120 droughts of China, among the SSP scenarios available, SSP2-4.5 and SSP5-8.5 scenarios were 121 selected because they are the most used scenarios that project the global temperature to increase 122 between 1.5 ℃ and 2.0 ℃ by the late 21st century. These projected climate variables of CMIP6 are 123 interpolated into gridded data of 1.5° resolution with bilinear method. 124 3 Research Methodology 125 3.1 Univariate drought indices, SPI, SSI, and SRI 126 To comprehensively assess future droughts across China, three popular univariable drought indices, 127 SPI, SSI, and SRI were estimated using ten GCMs' outputs: precipitation (pr, mm day −1 ), the surface 128 soil moisture content in the top 10 cm (mrsos, kg m −2 ), total soil moisture content (mrso, kg m −2 ), 129 total surface runoff leaving the land portion, excluding base flow drainage from the soil model 130 (mrros, mm day −1 ) and total runoff leaving the land portion, including drainage through the base of 131 the soil moisture (mrro, mm day −1 ). These selected variables essentially represent the three types of 132 traditional droughts -meteorological, hydrological, and agricultural droughts. For clarity, SSI and 133 SRI calculated using mrsos and mrros were denoted as SSIS and SRIS, respectively. The non-134 exceedance probabilities of drought variables were computed using an empirical formula: 135 where n is the series length, i is the sort rank of x, and p(xt) denotes probability of xt. Then the 137 empirical probabilities were transformed into a standardized index (SI) using an inverse Gaussian 138 distribution: So, SI is a Gaussian univariate variable ranging between −3 and +3. 139 (2) 140

141
To assess the performance of SPI, SSI, and SRI, we also estimated popular, comprehensive drought 142 indices, SPEI, SPAEI, and scPDSI. PET is the key predictor in estimating SPEI and scPDSI. Dai 143 (2011) Wells et al. 2004). scPDSI is estimated using the same method as PDSI but with different empirical 161 parameters. It was estimated from the differences between precipitation and climatically appropriate 162 precipitation ( ). represents the demand of the water balance calculated from the water budget 163 based on four variables, evapotranspiration (ET), runoff (RO), loss (L), and recharge (R), and the 164 corresponding potential variables by dividing the soil column into two layers (Zhang et al. 2021a). 165 Weighting factors ( , , ,     ) are used to represent the effects of the local climate on the four 166 potential variables expressed as for a single month, 167 The weighting factors ( , , ,     ) were estimated as follow: where the bar represents the mean value over the calibration period. The moisture departure (d) 171 depicting the difference between precipitation (P) and was calculated as 172 Because the same d could mean different things at different times and different locations, it 174 prevents a straightforward comparison between different values of d (Wells et al. 2004). To 175 overcome the problem, Palmer (1965) proposed correction factors K′ and K as given below, 176 10 2.8 1.5log +0.5, 1, 2, 3 , 12 where p and q are 0.897 and 1/3 derived by Palmer using climatic records from several stations in 183 western Kansas and central Iowa of the United States. But for scPDSI, the climatic characteristic K 184 in Eq. 8 and duration factors in Eq. 9 are replaced with automatically calibrated values based on the 185 historical climatic records in local regions, which makes scPDSI a more representative drought 186 measure (Wells et al. 2004). Therefore, scPDSI is selected in this study. For a given time scale, a 187 drought index value can be discretized into predefined categories that indicate not only the severity 188 of a drought event but also periods of excess moisture availability. 189 Although SPEI, scPDSI, and SPAEI are commonly used to assess meteorological droughts, they can 192 also be used to assess other types of drought because these indices consider the whole 193 atmosphere-land surface water balance processes. Vicente-Serrano et al. (2010a, b), who was the 194 first to propose SPEI, indicated that because of its multi-scalar characteristics, SPEI can identify 195 different drought types, namely, meteorological, hydrological, and agricultural droughts. For 196 example, Potopová et al. (2015) assessed agricultural drought risks using SPEI in the Czech Republic,197 showing that SPEI is applicable in agricultural drought detection. By comparing SPI, SPEI, scPDSI, China (regions 6 and 7) is projected to become drier, resulting in mild to moderate drought 257 conditions in 65% and 70% of the land area under SSP2-4.5 and SSP5-8.5, respectively. By the 258 2080s, more severe and widespread droughts are projected to occur in southern China, and the 259 fraction of land areas suffering from droughts would increase to 84% and 88% under SSP2-4.5 and 260 SSP5-8.5 at regions 6 and 7, respectively. In northern China, especially in the northwest desert 261 region (region 1) and northern Qinghai-Tibetan Plateau (region 3), soil moisture is projected to be 262 wetter with increasing trends similar to the projected precipitation trends. By the end of the 21st 263 century, according to the projected SSIS patterns, 40% (SSP2-4.5) and 62% (SSP5-8.5) of the 264 northwest desert region would be moderately or even very wet. In the same period, most other 265 regions (regions 2, 4, 5) are also projected to be exhibit near-normal or moderately wet patterns, but 266 not as wet as that of the projected SPI. 267

Evaluation of SPEI, scPDSI, and SPAEI in identifying different drought
The projected column soil moisture drought based on SSI exhibits a pattern similar to the near-268 surface soil moisture drought based on SSIS for the same periods, but with more severe and

291
Hydrological drought exhibits drought patterns similar to meteorological drought patterns (Fig. 2), 292 with considerably wetter patterns projected across some parts of China in the future compared to 293 the historical periods (Fig. 4). The most obvious wetting trends are projected in some parts of the 294 northwest desert (region 1), Inner Mongolia (region 2), northeastern China (region 4), and warm-295 temperate northern China (region 5). Most of these regions had modest drought conditions in the 296 1960s, which gradually changed to near-normal in the 2000s and are projected to be wetter in the 297 2050s, and eventually become moderately wet by the end of the 21st century. As expected, these 298 regional changes in hydrological droughts are similar to meteorological droughts, given 299 precipitation is the main driver behind the hydrologic cycle, and therefore the runoff trends. In these 300 northern regions, snow dynamics or changes to snowfall also play an important role, such as a shift 301 to an earlier onset of spring snowmelt, often at the expense of reduced summer runoff due to less snowfall and less surface snowpack in response to a warmer climate (Shi and Wang 2015). Under 303 warming, the total precipitation is also projected to increase in these regions, a higher fraction of 304 rainfall over the total precipitation, and higher future spring snowmelt runoff. Under the SSP2-4.5 305 climate scenario, both SRIS and SRI project a wetter pattern by the end of the 21st century. Under 306 the SSP5-8.5 climate scenario, more than 45% and 47% of the regions are projected to be moderately 307 or very wet based on the SRIS and SRI estimated, respectively. 308 Although some hydrological droughts are projected by the end of the 21st century in some 309 regions of the Qinghai-Tibetan Plateau according to SRIS (Fig. 4d and h), the projected hydrological 310 drought is less extensive compared to the projected agricultural drought. In contrast, some humid 311 regions (regions 6 and 7) are projected to experience opposite trends in hydrological and agricultural 312 droughts, such that the former changes to wet or near-normal conditions but the latter changes to 313 mild or moderate drought conditions. The opposite projections may be attributed to different 314 physical hydrologic processes that affect the basin runoff and soil moisture (

319
The mean scPDSI computed for two historical and two future periods are shown in Fig. 5. scPDSI 320 projects more severe drought conditions than meteorological and hydrological drought indices at 321 the end of the 21st century, but is similar to projected agricultural drought patterns, with drying 322 trends projected in southern China and a minor wetting trend projected in northern China. Compared 323 to SSIS, scPDSI projects future drier conditions in the northwest desert region (region 1), Inner 324 Mongolia (region 2), and some northwestern regions of Qinghai-Tibetan Plateau (region 3) for the 325 historical periods. In 1960s, 61% (71%) and 65% (79%) of region 1 (region 2) were under moderate 326 (mild) drought conditions, respectively under SSP2-4.5 and SSP5-8.5 (Fig. 5a and e). In the same 327 period, most parts of region 1 (region 2) were under mild (near-normal) drought conditions 328 according to SSIS (Fig. 3a and e). In the 2000s, scPDSI still indicates a slightly drier condition than 329 14 of Qinghai-Tibetan Plateau (region 3), but most other regions are under near-normal conditions, 331 which is similar to drought conditions based on SSIS. For both future periods, northern China is 332 projected to be wetter while southern China is to become drier. In 2080s, 60% (58%) and 71% (68%) 333 of region 1 (region 2) are expected to be under mild or wetter conditions, respectively under SSP2-334 4.5 and SSP5-8.5 (Fig. 5d and h). In the same period, 83% and 70% of the southern humid regions 335 (regions 6 and 7) are projected to be under mild or more severe dry conditions, which are marginally 336 less than the areas of drought based on SSIS (Fig. 3d and h) The r values for SPAEI-SPI are generally higher than 0.6 in arid regions (regions 1 and 2), and 371 higher than 0.8 in humid and semi-humid regions, which are generally higher than r values for 372 scPDSI-SPI and SPEI-SPI. The r values for scPDSI-SPI are generally higher than 0.3 at the non-373 humid region (regions 1−5), which are always higher than those for SPEI-SPI at long time scales (> These results suggest that SPAEI and sc-PDSI are likely better than SPEI in monitoring 378 meteorological droughts in arid and semi-arid regions, especially for investigating long-term 379 droughts, which inevitably are the primary interest to most countries coping with the long-term 380 impact of droughts which in recent years have been occurring more frequently and in greater 381 severity. For China, r values for scPDSI-SPI increase with time scales of 12−15 months but then 382 decrease with higher time scales. Therefore, scPDSI is likely more suitable than SPEI in identifying 383 droughts in China lasting more than a year, and the performance tends to be more robust in arid and 384 semi-arid areas. The r values for SPAEI-SPI are similar to those for scPDSI-SPI, but higher than 385 those for SPEI-SPI in northwest arid regions. The r values for scPDSI-SPI are lower in southern 386 humid regions for droughts of short-term time scales. When the time scale of droughts is longer 387 than 12-month, SPEI tends to perform poorly even in humid regions, which shows that SPEI is incapable of identifying long-term meteorological drought. In addition, the effectiveness of SPEI in 389 identifying future meteorological droughts is also dependent on climate scenarios. For example, the 390 r values for SPEI-SPI under SSP5-8.5 are considerably less than those under SSP2-4.5 in all regions. 391 This should be attributed to the larger increase in temperature projected under the SSP5-8.5 climate 392 scenario (Online Resource 1, Fig. S1), which lead to a larger difference between P and PET, and 393 SPEI tends to overestimate droughts with higher PET, especially in non-humid regions, where PET 394 can be considerably higher than AET. Therefore, it seems SPAEI and scPDSI are more suitable than 395 SPEI in identifying meteorological droughts in arid regions. In humid or semi-humid regions, 396 SPAEI performs better than other indices in monitoring meteorological droughts. 397 To better understand the performance of scPDSI, SPEI, and SPAEI in evaluating agricultural 398 droughts, the correlations between scPDSI and SSIS, SPAEI and SSIS, and SPEI and SSIS were 399 investigated.

Correlations between scPDSI and SRIS, between SPEI and SRIS, and between SPAEI and SRIS 419
were also examined to evaluate the performance of these three drought indices in identifying 420 hydrological droughts of China. Fig. 10  scenarios. Similar to its performance in identifying meteorological or agricultural droughts, the 431 performance of scPDSI decreases with time scales beyond 12-month, and it tends to be poorer than 432 the performance of SPEI. Therefore, it seems that SPAEI is more effective in identifying 433 hydrological droughts in both humid and arid regions, SPEI is better than scPDSI in identifying 434 hydrological droughts in humid regions, but scPDSI is better than SPEI in identifying hydrological 435 droughts in drier regions. 436

437
From the above analysis, scPDSI shows more robust performance in identifying agricultural 438 droughts, especially in non-humid or arid regions. Therefore, it is chosen to further analyze the 439 evolution of drought characteristics in China. SSIS is also selected to analyze the drought 440 characteristics of these regions together with scPDSI. Fig. 11  Due to differences in the range of drought severities estimated from scPDSI and SSIS, it is 463 difficult to directly compare the drought severities estimated from these two indices. Therefore, the 464 time series of drought severities in seven regions obtained from scPDSI and SSIS are separately 465 shown in Fig. 12 and Fig. 13. The drought severity based on scPDSI shows a decreasing trend before 466 the 2050s but an increasing trend after 2050s in northwestern China, with an average scPDSI of 467 of the 21st century. In humid regions, drought severity and drought areas based on scPDSI and SSIS 476 both show obvious increasing trends, which likely means that these regions will experience more 477 severe dry spells compared to historical periods. Similar to drought areas, drought severity has no 478 obvious trend in semi-humid regions of China (regions 4 and 5). Drought severity obtained from 479 scPDSI agrees better with that from SSIS in humid regions, which is similar to drought areas. 480

481
It is beneficial to understand how the occurrence of different types of droughts will change under 482 the impact of climate change. We investigated drought responses to warming under SSP2-4.5 and 483 SSP5-8.5 climate scenarios using outputs from GCMs of CMIP6 in seven climatic regions across 484 China. Three types of univariate drought indices based on projected precipitation, soil moisture, and 485 runoff were used to investigate future drought conditions, including all three traditional drought 486 types -meteorological, hydrological, and agricultural droughts. In addition, three popular drought 487 indices, scPDSI, SPEI, and SPAEI, were also estimated to compare the projected drought patterns 488 of China in the 2050s and 2080s with these three types of univariate drought indices. Finally, 489 drought areas and severity were also analyzed using scPDSI and SSIS for 7 climatic regions of 490 China. In this study, we only estimated SPI, SPEI, and scPDSI patterns using original 491 meteorological variables (precipitation, temperature, etc.) simulated by GCMs so that we can 492 compare these indices with SRI and SSI estimated from soil moisture and runoff data of GCMs,493 which are difficult to be downscaled due to the lack of observed data to develop empirical 494 relationships between simulated and observed data. To assess the validity of using original GCM 495 data in the drought analysis, we have compared meteorological data (precipitation and temperature) 496 anomaly at 7 climatic regions and drought patterns (SPI and SPEI) derived from original GCM 497 simulations with those derived from statistically downscaled GCM simulations by the BCSD 498 method from 1979 to 2100 (Online Resource 1, Fig. S2−Fig. S5). The downscaled precipitation and 499 temperature show a similar increasing trend to the original GCM precipitation (Online Resource 1, 500 Fig. S2 and Fig. S3). Drought indices (SPI and SPEI) estimated using original GCM simulations 501 also have similar patterns with those estimated using statistically downscaled GCM simulations at three periods (the 2000s, 2050s, and 2080s) (Online Resource 1, Fig. S4 and Fig. S5). Therefore, 503 we expect the projected changes to meteorological, hydrological, and agricultural droughts based 504 on the original GCM data to be generally representative. 505 Based on results obtained from different drought indices except for SPEI, it seems that northern 506 arid China is expected to become wetter. Projected meteorological and hydrological droughts 507 consistently show that most parts of China will become wetter in the future. This is mainly attributed 508 to the increase of precipitation. Under a warmer climate, the atmospheric moisture is expected to be 509 higher as the near-surface atmospheric water holding capacity will increase at about 7% per °C of 510 warming (Clausius-Clapeyron scaling) (Zhang et al. 2021a). Further, the possible convergence of 511 storm systems could also result in higher rainfall totals during wet events. According to CMIP6 512 models, the global annual mean precipitation over land, especially at high latitudes of the Northern 513 Hemisphere will increase. Under both SSP scenarios, precipitation increases faster in winter than 514 that in summer, especially in northern China (Online Resource 1, Fig. S6). On the other hand, under 515 a warmer climate in some cold and alpine regions of China, the snowpack will decrease partly due 516 to less snowfall and higher spring snowmelt runoff, which together with higher evaporative losses 517 will result in less summer runoff in these regions. Compared to meteorological and hydrological 518 droughts, agricultural drought is widespread by 2080s in southern China, which is consistent with 519 results obtained using the scPDSI index. This may be partly attributed to the increase in water vapor 520 deficit (Dai et al. 2018)  In arid and semi-arid regions, PET is mostly much higher than AET which represents the atmospheric water demand. However, in these regions, it is precipitation rather than PET that 532 determines AET (Online Resource 1, Fig. S7). On the other hand, in humid regions, the change in 533 AET is more dependent on PET than on precipitation. Because of differences in the relationships 534 between precipitation, AET, and PET at different climate zones, SPEI which uses the difference 535 between precipitation and PET in estimating surface water deficit/surplus could overestimate 536 droughts (Zhang et al. 2019a). In general, using as the water demand metric is physically more 537 reasonable than PET in estimating the surface water-energy balance, especially in arid regions.   Mean SSIS (top panel) and SSI (bottom panel) drought indices of China estimated from ten GCMs' simulations for two historical  and two future (2041-2070, 2071-2100) periods under SSP2-4.5 (a-d for SSIS and i-l for SSI) and SSP5-8.5 climate scenarios (e-h for SSIS and m-p for SSI), respectively Mean SRIS (top panel) and SRI (bottom panel) drought indices estimated from ten GCMs' simulations for two historical  and two future (2041-2070, 2071-2100) periods under SSP2-4.5 (a-d for SRIS and i-l for SRI) and SSP5-8.5 climate scenarios (e-h for SRIS and m-p for SRI), respectively Mean scPDSI estimated from ten GCMs' simulations for two historical  and two future (2041-2070, 2071-2100) periods under SSP2-4.5 (a-d) and SSP5-8.5 climate scenarios (e-h), respectively Figure 6 Mean SPEI estimated from ten GCMs' simulations for two historical  and two future (2041-2070, 2071-2100) periods for SSP2-4.5 (a-d) and SSP5-8.5 scenarios (e-h), respectively Mean SPAEI estimated from ten GCMs' simulations for two historical  and two future (2041-2070, 2071-2100) periods for SSP2-4.5 (a-d) and SSP5-8.5 scenarios (e-h), respectively Pearson correlation coe cients between SPEI and SSIS, SPAEI and SSIS, and scPDSI and SSIS at 1-to 48-month time scales in the seven climatic regions of China under SSP2-4.5 and SSP5-8.5 scenarios, respectively Figure 10 Pearson correlation coe cients between SPEI and SRIS, SPAEI and SRIS, and scPDSI and SRIS at 1-to 48-month time scales in the seven climatic regions of China under SSP2-4.5 and SSP5-8.5 scenarios Figure 11 The percentage of drought areas estimated from scPDSI and SSIS of the seven climatic regions of China over the historical and future periods (1950-2100) projected under SSP2-4.5 and SSP5-8.5 climate scenarios of ten GCMs of CMIP6 Figure 12 The severity of droughts estimated from scPDSI for the seven climatic regions of China over the historical and future periods (1950-2100) projected under SSP2-4.5 and SSP5-8.5 climate scenarios of ten GCMs of CMIP6 Figure 13 The severity of droughts estimated from SSIS for the seven climatic regions of China over the historical and future periods (1950-2100) projected under SSP2-4.5 and SSP5-8.5 climate scenarios of ten GCMs of CMIP6 Supplementary Files