Increased Occurrence and Intensity of Consecutive Rainfall Events in the China's Three Gorges Reservoir Area Under Global Warming

Consecutive rainfall events (CREs) are important triggers of geological hazards like landslide downhill and mudslide in the Three Gorges Reservoir area (TGR), China. These hazards are not only potential risks for the effective storage capacity of the reservoir but also threats of the safety of the reservoir's Great Dam. The future changes of CREs' occurrence and intensity are analyzed by using the projection experiments from 20 models attending the Coupled Model Intercomparison Project Phase 5 (CMIP5) under three different representative concentration pathways (RCP2.6, RCP4.5, and RCP8.5). Spring and fall are focused on, during which CREs are most frequent. Considering a common overestimate of rainy days number in the state‐of‐the‐art models, a new approach is developed to define CREs based on the percentile of rainfall distribution in observations. The approach yields a similar CREs climatology in models to that in observations and thus is used to identify CREs in models. The results based on multiple model ensemble (MME) and model spread comparison suggest a significant increase in spring and an overall decrease in fall in CREs' occurrence under all three scenarios. As for the intensity, it is projected to intensify in both spring and fall. Particularly, the higher the emission scenario, the greater the spring accumulated rainfall amount during a single CRE. These results imply an increasing risk of geological hazards in the TGR in the future.


Introduction
The Three Gorges Reservoir (TGR) area, spanning 28-32°N latitudinally and 105-112°E longitudinally (see Figure 1), is a mountainous, highly populated region locating in the middle reach of the Yangtze River basin, central China. It often suffers from geological disasters like landslide and mudslide (Chen et al., 2012;Ma et al., 2006). These disasters result in tremendous damages to the lives and properties. One example is the landslide occurring in 1998, which caused a direct economic loss of 610 million RMB (Ma et al., 2005). In addition, they cause rockfall, mud, and debris flows, which block the rivers running to the reservoir; reduce the effective storage of the reservoir (Zhang et al., 2016); and even threaten the safety of the reservoir's Great Dam. Therefore, predicting, warning, and preventing geological hazards are an important national demand in China.
Synoptic processes, especially consecutive rainfall events (CREs), during which it rains for 1 week and even longer with a gentle and moderate intensity, are a substantial trigger to geological hazards, although other factors like short-duration heavy rainfall or earthquake can be also important (Guzzetti et al., 2007;Ye et al., 2009). Corominas and Moya (1999) illustrated that the risk of landslide increases substantially when it rains persistently for several weeks with the moderate accumulated rainfall amount over 200 mm. The size of the landslide may be positively proportional to the duration of CREs (Jibson, 2006). One recent case is the severe landslide occurring in Lishui (28.6°N, 119.9°E), Zhejiang Province on 13 November 2015, which resulted in 38 deaths (Liu, 2015). Prior to the landslide, it rained lightly or moderately for nearly 1 month, with an intensity of only 6-8 mm per day. No precursor was found, and no warning was issued before the disaster. In addition to trigger geological disasters, CREs adversely influence agricultural production, cause local pooling or freezing rain during chilling weather, and affect human health (Ding et al., 2007(Ding et al., , 2008Li et al., 1977;Sun et al., 2016;Wu et al., 2007). Therefore, understanding the future trend of CREs is of substantial importance.
The CREs in the TGR occurred mostly in the spring and fall Zou, 2005). Previous analyses based on instrumental records suggest a decreasing trend in spring CREs' occurrence, duration, and accumulated rainfall amount, but an increase in the mean daily rainfall during the past decades (Zheng et al., 2018;Zou, 2005). In fall the trend is somewhat similar, with a decrease in occurrence despite an increase in intensity (Sun et al., 2016;Wang & Zou, 2015). Whether such a trend persists into the future is unclear.
Under the context of global warming, rainfall features change, including its occurrence frequency, duration, and intensity (Scoccimarro & Gualdi, 2013;Trenberth, 1998;Zhai, 1999). Of particular importance is that rainfall becomes regionalized and intensified, as far as one individual rainy event is concerned (Giorgi et al., 2001;Lau et al., 2013;Sun et al., 2006). This inevitably leads to changes of CREs. Thus, projecting the future trend of CREs in TGR consists of the preliminary aim of the present study.
The reminder of this paper is organized as follows. Section 2 describes data and methods. The projection experiments from the models attending the Coupled Model Intercomparison Project Phase 5 (CMIP5) under three different representative concentration pathways (RCP2.6, RCP 4.5, and RCP8.5) are used (Li et al., 2016;Sillmann et al., 2013). Because of one common bias with too many rainy days in the state-of-the-art models, the canonical method used to identify observational CREs appears inappropriate for modeled precipitation (Boé et al., 2007;Li et al., 2016). Thus, a new approach is developed for the models. Section 3 compares the CREs in the historical experiments with those in observations. Since not all models reproduce the observed CREs well, just those "good" models are selected to project the future trend. Section 4 gives the projection results based on the multiple-model ensemble mean (MME) and an assessment of result diversity in the individual models under different emission scenarios, with the focus on the accumulated rainfall amount and daily rainfall intensity in CREs. Finally, a summary and discussions are given in section 5.

Data Sets
Gridded daily precipitation outputs from 20 models participating in CMIP5 are employed (Table 1). In order to treat all the models equally, only their first run (r1i1p1) is analyzed. The experiments include the historical run with historically evolving forcing for 1961-2005 and the projection runs with prescribed forcing of RCP2.6, RCP4.5, and RCP8.5 for 2006-2099 (Taylor et al., 2012). RCP2.6 is a low, peak-and-decay scenario in Figure 1. The topography distribution of the TGR (the data were obtained from Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences: http://www.resdc.cn). The west and east sections to TGR are the Sichuan Basin and the Jianghan Plain, and the north and south sections to TGR are the Daba Mountains and the Wuling Mountains, respectively. The red dot of upstream is Chongqing City, which is a provincial capital with a population of 30 million, and the purple dot of the downstream indicates the location of the Great Dam of TGR. which radiative forcing reaches the maximum near the middle of the 21st century before decreasing to an eventual nominal level of 2.6 W/m 2 . RCP4.5 is a medium stabilization scenario that follows a rising radiative forcing pathway leading to 4.5 W/m 2 in 2100, while RCP8.5 is a high, business-as-usual emissions scenario with radiative forcing increase to 8.5 W/m 2 by 2100. Details on the CMIP5 models and their configurations are described in the PCMDI website (at http://www-pcmdi.llnl.gov/).
To assess the CMIP5 models' ability in reproducing the observed CREs, the daily gauged grid precipitation data set, referred to as CN05.1, is employed. CN05.1 was produced by data from high-resolution stations across China during the period from 1961 to 2015. It uses thin-plate smoothing splines interpolation for climatology and angular distance weighting interpolation for daily deviation before merging into the full 0.25°× 0.25°grids (Wu & Gao, 2013;Xu et al., 2009). This methodology follows the method by which the Climatic Research Unit data set was created (New et al., 2000). More details about validation information of CN05.1 are given in Wu and Gao (2013). It has been used by a lot of previous studies (e.g., Chen et al., 2014;Li et al., 2020;Pan et al., 2020;Sui et al., 2015). In view of the possible mismatch in horizontal resolutions, both the simulated precipitation and CN05.1 are regridded to a 1.0°× 1.0°grid by using a bilinear interpolation algorithm.

Definition of CREs
In observational studies (e.g., Li et al., 1977;Zou, 2005), one CRE is isolated when there are five or more consecutive rainy days. One rainy day is defined when the accumulated amount is greater than or equal to 0.1 mm within 24 hr from 00 UTC to the next 00 UTC. For CN05.1, because of the rain intensity diffusion and extended rainy days caused by interpolation, one elevated threshold, 1 mm per day, is used to define rainy days. One similar threshold was used in previous studies (Giorgi et al., 2011;Mohan & Rajeevan, 2017;Salinger & Griffiths, 2001). Here one CRE is defined to begin if any one of the following four cases: (1) It has at least 5 consecutive rainy days; (2) it has 6 or 7 rainy days within 7 or 8 consecutive days, despite no 5 consecutive rainy days; (3) it has 7 or 8 rainy days within 9 to 10 consecutive days but has at least Earth and Space Science 1 rainy day within any 2 consecutive days, although it does not meet (1) or (2) above; (4) it has more than 9 rainy days but has at least 1 rainy day within any 2 consecutive days, although it does not meet (1)-(3) above (Sun et al., 2016;Zheng et al., 2018). The CRE termination is defined if there are two consecutive nonrainy days following CRE, and the duration is the day number from the beginning date until the ending date.
For the model outputs, the above definition is inappropriate because models generally overestimate rainy day number but underestimate precipitation intensity (Dai & Trenberth, 2004;Sun et al., 2007). It will cause much-more-than-observed CREs if a same 1 mm threshold is used. Previous studies developed various calibration methods to correct the bias. The first one is the simplest unbiasing method, which simply overlaps the models' climatological mean bias into the simulations (Déqué, 2007). It is straightforward but has an implicit, unrealistic assumption that the modeled mean rainfall follows the observed regardless of the variance. The second is a combination of local intensity scaling with power transformation. It scales the modeled precipitation within the observations and corrects both the climatological mean and variance (Fang et al., 2015;Schmidli et al., 2006;Teutschbein & Seibert, 2012). In details, modeled raw precipitation is calibrated by multiplying the ratio of the observed mean precipitation to the modeled. The method will cause unmatched successive days and weakened rainfall extremes. The third one is probability quantile mapping (Semenov & Stratonovitch, 2010;Themeßl et al., 2012). It adjusts the climatological mean, variance, and probability quantiles distribution of modeled precipitation and has no influences on the extremes of the modeled rainfall. But it fails to correct the temporal autocorrelation properties intrinsic to series and neglects the physical connection between variables (Boé et al., 2007). The fourth is the Artificial Neural Networks technique. It efficiently handles the noisy and unstable data that are typical in weather station observation and maps highly nonlinear relationships between a set of inputs and the corresponding outputs (Luk et al., 2000). The rainfall estimates from Artificial Neural Networks are even more accurate than those based on statistical or dynamic downscaling approaches (Mendes & Marengo, 2010;Skamarock et al., 2008). But this technique is highly sensitive to the quantity and distance of neighboring gauges, and to the local hydrologic system as well (Hung et al., 2009).
In the present study, the long-term trend and future projection of CREs are focused on, so the rainfall intensity in a single day is relatively less important than whether it rains or not on that day. Since rainy day number may be subjective to change in all the above calibration methods, we develop a new approach to defining CREs instead. It is based on the Cumulative probability Distribution Function (CDF) of daily rainy amount. The CDF in models is assumed to follow the observed, which is calculated based on the threshold of 1.0 mm per day. Thus, for the models the threshold to define rainy days can be derived. Subsequently, rainy days number and the CREs can be easily calculated at each grid points.

Variable to Quantify CREs
A total of four variables are used to quantify the CREs, including occurrence frequency (OCF), total rainy days (TRD), accumulated rainfall amount (ACR), and mean daily rainfall intensity (INT). The first, OCF, describes the occurrence climatology of CREs, while the latter three describe the duration and strength of one single CRE. INT is not independent of ACR and TRD but equal to ACR divided by TRD. Previous studies suggest that all the four variables are linked to geological hazards (Corominas & Moya, 1999;Jibson, 2006); thus, they are used for the present analysis. Since CREs occur most frequently in spring and fall in TGR, just these two seasons are focused on. Spring is referred to as 25 February to 4 June, while fall is referred to as 28 August to 4 December considering the possibility that CRE occurrence is not contained strictly in a whole calendar season.

Sen's Slope Estimate
Due to nonnormal characteristics of probability distributions, trends of daily precipitation amount and subsequently CREs cannot be estimated by the least squared fitting. As Santos and Fragoso (2013) and Mohan and Rajeevan (2017), here the trends are estimated by the Kendall's tau-based slop estimator (Sen's slope Q; Sen, 1968) as follows: Specially, for one time series with the length of L, at one time point (say t i , i ¼ 2, …, L), slope can be calculated by using values (X i ) and (X j ) at time points t i and t j , respectively. Here t j precedes t i by at least one unit (j ¼ 1, …, i−1). As such, a total number of (L−1)! slopes can be obtained, and the median among all the slopes is the best estimate of the trend. According to Yue et al. (2002), the Sen's slope is better than the least squared fitting when estimating precipitation trend. After get the trend Q, we use the least squared estimator to estimate the intercept B of the trend-dominated series (y ¼ Qx + B) (Wang & Swail, 2001). Nonparametric Mann-Kendall test is used for significance validation, since it is reliable for both monotonic linear and nonlinear trends in nonnormal distributed series (Gotway, 1992).

Metrics for Model Performance and Selection
To assess model's ability in reproducing spatial pattern of CREs, the Taylor Diagram (Taylor, 2001) is employed, which provides a statistical comparison of simulated and observed CREs, in terms of spatial correlation coefficient, root-mean-square (RMS) difference, and standard deviation. The RMS difference and the standard deviation of various indices in the models are normalized by the observed. Thus, a perfect model has the RMS difference equal to 0, and the spatial correlation and the ratio of spatial standard deviations are both close to 1.
Because of no initialization for the oceanic model component in these CMIP5, one should not expect that their historical runs have the ability to reproduce the observed CREs evolution. Thus, we compromise to assess their normalized temporal standard deviation δ m /δ o (Han et al., 2014;Santer et al., 2009). Here δ m and δ o denote the interannual standard deviation of model simulated and observed seasonal mean CREs variables, respectively. The closer to 1 the value, the better the agreement between simulation and observation.
Considering the substantial importance of CRE occurrences, just OCF and TRD are used to select "good" models. Three criterions are used based on Taylor diagram: (1) The spatial correlation coefficient of models' CRE occurrences with the observed is above 0.31 (significant at the 95% level), (2) the normalized spatial RMS of CREs' occurrences is less than 1.5, and (3) the normalized deviation of modeled spatial CREs' occurrences is smaller than 1.5 but great than 0.5. Besides, another more criterion is considered: The normalized temporal standard deviation of simulated occurrences is smaller than 1.5 but great than 0.5.

Models' Simulation on CREs in the Historical Experiments
The threshold for rainy days is derived before we evaluate the ability of the models in reproducing the observed OCF and TRD. Figure 2 compares the CDFs from the observations and the models. From it, the percentile with the cumulative probability in the observed rainy days below the threshold (1.0 mm per day) is 60.49% in spring (64.17% in fall) in all grid points. Correspondent to the same percentile, the threshold is different from model to model. For example, for MRI-CGCM3 and CSIRO-Mk3.6.0, the threshold is 2.1 mm per day in spring and 1.0 and 0.5 mm per day in fall, respectively. The closer to 1.0 mm the derived threshold, the better the model in capturing the observed rainy days. A higher or lower threshold indicates an overestimate or underestimate in modeled rainy days. Column 3 in Table 2 gives the derived threshold for the individual models.

Earth and Space Science
After the threshold is derived, simulated CREs can be calculated subsequently. Figure 3 is the Taylor diagram, which compares the simulated and observed spatial distribution of CREs. About one half of the models fail to reproduce the spatial pattern of spring CREs, with the correlation less than 0.31. For fall, only a quarter of the models exhibit a significant skill. MME of all 20 models (refer to as MME_A; the character "A" means "all"; refer to Table 2 and Figure 3) shows a pronounced bias. In spring, the spatial correlation coefficient in observed and MME_A OCF (TRD) is 0.74 (0.72), and the standard deviations in MME_A is underestimated relative to the observed. In fall the modeled standardized deviations are close to the observed, but the correlation coefficient is 0.38 (0.33) for the observed and model MME_A OCF (TRD), even lower than that in spring.
By applying the criterions in section 2.2.4, for spring a total of 11 models (FGOALS-g2, IPSL-CM5A-LR, IPSL-CM5A-MR, CNRM-CM5, CanESM2, HadGEM2-AO, HadGEM2-ES, MIROC5, MPI-ESM-LR, MRI-CGCM3, and CSIRO-Mk3.6.0) outstand as "good" models. For fall, a total of five models (IPSL-CM5A-LR, IPSL-CM5A-MR, CNRM-CM5, MRI-CGCM3, and CSIRO-Mk3.6.0) are selected. MME of these "good" models (refer to as MME_G; the character "G" mean "good" in Table 2 and Figure 3) exhibits an evident improvement in reproducing the observed CREs, with the spatial correlation coefficient in the observed and simulated OCF and TRD above 0.84 in spring (0.75 in fall) from these "good" models. Also, the standardized deviations in these "good" models are closer to the observed. Hereafter, just the results from these "good" models are analyzed, and for brevity MME is used to represent MME_G unless it is clarified specially.

10.1029/2020EA001188
Earth and Space Science Figure 4 gives a spatial distribution comparison of CREs occurrences in the observation with those in MME.
In spring, the observation shows a greater OFC and TRD in the southeast than those in the northwest, which is consistent with the MME. In fall, observed CREs exhibit a larger occurrence center concentrated in the southwest. That is also consistent with the MME. Table 2 compares the performances of these "good" models along with their MME and the observations. The seasonal mean accumulated rainfall amount (column 4) in observations in spring and fall is 328.1 and 250.3 mm, respectively, while this value in models varies from 274 to 463 mm in spring (from 163 to 256 mm in fall). The overall consistence in the models' and the observed CREs climatology indicates a qualitative reasonability of this derived threshold. The climatological OCF in individual models and in MME is close to the observation (column 5) in both spring and fall. As for trend, in spring, the OCF in observation (column 6) exhibits a reduction, albeit a lack of significance. About a half of the models yield a same negative trend as the observed, but no significant trend is seen in the models else. The spring trend in MME is nearly neutral (0.01 times per decade), in contrast with a reduction in observations (−0.10 times per decade). In fall the trend in models bears an overall similarity to the observed, which is consistently negative among all the five "good" models albeit being less significant. Also, their MME shows a significant reduction, which is consistent with observations.
For TRD (columns 7 and 8), the observed climatology is 18.1 and 18.6 days in spring and fall, respectively. This value in most of the models is slightly greater both in spring (from 18.4 to 20.5 days, except for MIROC in which it is 17.7 days) and fall (from 21 to 25.2 days). As for trend, 7 models among the 11 "good" for spring and all the 5 "good" models for fall yield a reduction consistent with the observed (column 8). Not surprisingly, MME yields a reduction both in spring and fall, in agreement with a major of the models.
Above the models' occurrence of CREs, climatology and trends have been analyzed. The interannual variability of CREs in these models is also compared with the observations. Figure 5 (left panels to the black dashed vertical line) displays the modeled and observed OCF and TRD evolutions. From it, the uncertainty (model spread) within the models generally conforms to the observed, although the variability is less evident. These analyses suggest that the occurrence of CREs in models is overall comparable to the observed.
Based on OCF and TRD, the intensity of CREs (ACR and INT) is further investigated. The observed ACR (columns 9 and 10) is 143.1 and 148.6 mm in spring and fall, respectively. The modeled value in spring is greater in most of models except for IPSL (122.3 and 125.3 mm, respectively). Six models simulate a  Table 1). The reference ( reduction trend, consistent with the observation (−11.82 mm per 10 year). However, MME suggests an increase trend. It may be not realistic, because it is dominated by CanESM2. In fall it is less unanimous, with two models with a higher value and the three else with a lower value. All five models and their MME show a reduction trend, consistent with observation.
The observed INT (column 11) is 7.8 and 7.9 mm per day in spring and fall, respectively. The value in spring in most of models (8.0 to 12.2 mm per day, except for two IPSL models) is slightly greater, but somewhat smaller in fall (4.8 to 7.0 mm per day). The observed INT trend exhibits negative in both the seasons. Only a small fraction of models in spring (3 out of 11 models) reproduces the observed trend, but so do a major of models in fall (three out of five models) (column 12). Figure 8 (left half to the vertical black dashed line) displays the modeled ACR and INT evolutions in historical runs along with the observations. Although   in the observation with those in the MME of "good" models. a comparison of the evolutions itself does not yield much meaning due to no initialization as mentioned in the above section, it can still provide insights into the interannual variability. From it, the simulated intermodel spread conforms to the observed. These analyses suggest an overall consistence of CREs in these selected models with the observed. This lays a basis for projecting CREs' future change by using these "good" models.

Occurrence
Figure 5 (right to the black vertical dashed line) shows the projected occurrence of CREs (OCF and TRD) averaged over TGR under three emission scenarios. In spring (Figures 5a and 4b), MME shows a significant increase in OCF and TRD under all the three RCPs. The increase is most evident under RCP4.5. Most of the individual models yield a consistent result with MME. For OFC, 7, 8, and 6 models among all the 11 models project the result consistent with MME under RCP2.6, RCP4.5, and RCP8.5, respectively. The numbers are eight, eight, and six for TRD.
In fall (Figures 5c and 5d), a significantly reversed decrease trend is projected in OFC and TRD. The higher the emission, the more obvious the decrease. As for individual models, for OFC, a total of three, three, and five models among all the five "good" models project a decrease under RCP2.6, RCP4.5, and RCP8.5, respectively. The number is two, three, and five for TRD.
In view of regional difference in CREs within TGR from south to north (Zheng et al., 2018), whether the CREs trends vary in different subregions is intriguing. Figure 6a shows the distribution of projected spring OFC and TRD trend in MME. A resemblance is seen between them. First, there is an overall increase in the whole region, particularly its plain western section. Second, the increase is more visible under the  , yellow and black lines represent the observed and MME simulated, respectively. For the projection period (2006-2099) blue, green, and red lines represent three emission scenarios (RCP 2.6, RCP 4.5, and RCP 8.5), respectively, with the correspondent linear fitting indicated with dashed lines. Shading with gray, light blue, light green, and pink denotes the 95% confidence intervals of standard deviation for the "good" models in historical runs and projected runs (RCP 2.6, RCP 4.5, and RCP 8.5, respectively). The anomalies are relative to 1961-2005.

10.1029/2020EA001188
Earth and Space Science lower emissions (RCP2.6 and RCP4.5) than the higher emission (RCP8.5). This has been seen in the area mean above.
Since the result from one single model may dominate MME, this causes uncertainty of projected results. To assess the uncertainty, we analyze the agreement of the models' results. Figure 6b displays the spatial distribution of model number projecting an increase in occurrence of CREs. From it, most of models show a positive trend in OFC and TRD (warm yellow corresponds to an upward trend) in spring. Also, more models are in agreement with MME in the western section. This indicates a larger reliability in the CRE increase there (Figure 6a).
The distribution of projected trend in occurrence MME in fall is displayed in Figure 7a. A decrease in both OFC and TRD is seen across the area, particularly over the southwestern section. The decrease is even obvious under the higher-emission scenario. This is also seen from the distribution of the number of model (Figure 7b). The number of models is represented with deeper blue when they project an overall downward trend in OFC and TRD (Figures 5c and 5d). The southwestern section of TGR projects a consistent decrease under all three scenarios, where CREs occur most frequently in fall (Zou, 2005). Besides, the models' agreement increases along with the enhancement of emissions. Almost all the five "good" models project a reduction trend in occurrence across the region under RCP8.5, and the reduction in about one half of the models is significant in these grid points.

Intensity
Strong precipitation increases the risk of geological hazards (Corominas & Moya, 1999;Guzzetti et al., 2007;Jibson, 2006). Here we analyze the projected intensity of CREs expressed as ACR and INT. From Figure 8 (right to the vertical black dashed line), a significant increase in ACR and INT in spring is seen under all
In fall (Figures 8c and 8d), a negative trend in ACR is projected, being significant under RCP4.5 and RCP8.5. There is (are) one, three, and three model(s) among the five "good" models projecting the result consistent with MME under RCP2.6, RCP4.5, and RCP8.5, respectively. That only one model bears a similar projection to MME implies substantial uncertainty under RCP2.6. For INT, one opposite result is projected, but it may be robust since four, three, and five models among all the five "good" models yield a result similar to MME. Figure 9a shows the spatial distribution of projected spring ACR and INT trend in MME. ACR under all three scenarios shows an increase from north to south, while INT increases in different sections under different emissions. Under the lower scenario, the increase is located in the highly populated southwestern section, but in the northeastern closer to the Great Dam under the higher emissions. The model's spread is checked in Figure 9b. For ACR, almost all models project a positive trend under all the three scenarios, particularly in the southwestern section. A greater spatial homogeneity is seen in INT under RCP4.5 and RCP8.5 than that under RCP2.6.
The spatial distribution of projected fall ACR and INT trend in MME is displayed in Figure 10a. Similar to occurrence (Figure 7a), a decrease in ACR is located in the southwestern area, and it is more pronounced under the higher scenario. INT shows an increase in the western section under RCP2.6 and RCP8.5, but a decrease across the area under RCP4.5. Similar to the previous analyses, the model number projecting a same trend as MME is presented in Figure 10b. From it, the models' agreement is relatively higher in ACR under higher than lower emissions.
Both the accumulated amount and short-duration rainstorm intensity are crucial triggering geological hazards (Corominas & Moya, 1999;Jibson, 2006). Thus, an in-depth analysis on ACR and INT in

Earth and Space Science
individual models is conducted below. Figure 11 shows the frequency distribution of spring ACR bins for different decades under RCP 4.5. In spite of an in-between difference, most of the models yield a visible increase in the future. For example, there is an increased frequency of heavy ACR (exceeding 80 mm) during 2070 to 2099 in CNRM-CM5, HadGEM2-AO, HadGEM2-ES, and MRI-CGCM3. Also, the increase in spring under different scenarios is similar to one another, but with a greater amplitude under RCP8.5 than RCP2.6. In fall, the individual models project a regional nonunanimous result except for CNRM-CM5, which projects increased grids with ACR exceeding 200 mm during the late 21st century. There is no significant change in the two models from IPSL but a slight decrease from the two models else.
The increase in evaporation resulted by warming is greater than the atmospheric capacity in holding moisture; this imbalance implicates a decrease in light to moderate precipitation events and an increase in high precipitation events (Sun, 2006;Trenberth, 1998). The light to moderate precipitation events consisted of a fraction of CREs. To obtain the projection of precipitation intensity in CREs (i.e., INT) in the future, the 90th, 95th, and 99th percentiles obtained by aggregating daily rainfall intensity from all CREs are used to classify four major categories: light rainy, moderate rainy, heavy rainy, and extreme rainy days. Figure 12 compares the projected change in the individual models in 2020-2049 (near future) and 2070-2099 (far Figure 12. The change rate of daily rainfall intensity at different categories during CREs. The increments are (a) 2070-2099 or (b) 2020-2049 relative to the reference period 1970-1999. Response of each "good" CMIP5 models is denoted by different color marks. Different categories of rainfall are shown as LR (light rain), MR (moderate rain), HR (heavy rain), and ER (extreme rain). Part of models' mark in (b) (the ER growth of MRI-CGCM3 are 422%, 526%, and 638% from RCP 2.6 to RCP 8.5, and the ER growth of HadGEM2-ES is the 235% under RCP 8.5) is not shown due to the oversized change rate.
future) relative to 1970-1999. In far future, for spring light rain (Figure 12b) there is about a half of the models projecting a reduction but an increase by the remaining models. For spring moderate rain, more models project an increase with a higher model agreement. Also, almost all models project an increased heavy and extreme rain, and this is particularly evident by HadGEM2-ES, HadGEM2-AO, and MRI-CGCM3. Besides, the increase is most significant under high emissions. In fall, for light, moderate, and heavy rain, most models project weakening in daily rainfall intensity, and the weakening is most prominent under all three scenarios in MRI-CGCM3. In contrast, all the models display an enhanced daily intensity in extreme rain under RCP 2.6 and RCP 8.5.
In near future the projected changes ( Figure 12a) are qualitatively similar to the far future but weaker. This indicates a gradual increase in spring daily rainfall intensity during CREs in the future (Figure 8b). In fall, an overall increase is projected (Figure 8d), although it is not so unanimous in different categories. This increase may be attributed to the growth of extreme rain. This seems reasonable because the precipitable water within the atmosphere increases under a warming context, and it is easier to form bigger particles and rain drops.

Summary and Discussions
The TGR area in China suffers from geological hazards like landslide downhill and mudslide. CRE is a substantial trigger. In this study we used the IPCC CMIP5 outputs to project the future trends of CREs' occurrence and intensity. Just the "good" models are chosen to project based on their historical simulations on the observed CREs.
Considering the common systemic bias with more rainy days in most of the state-of-the-art models, a new approach to defining model's rainy days has been developed based on the CDF of the observed daily rainfall amount. Then, models' rainy days number has been derived to identify CREs. A total of 11/5 models have been selected as "good" models to project the future trends for spring/fall. These models have exhibited a relatively higher skill in reproducing observed CREs' spatial patterns and interannual variability of occurrence.
The results suggest an increase of the occurrence of CREs in spring, being most significant under RCP4.5, but a reduction in fall, being more evident under higher scenarios. The projected change in occurrence is more prone in the southern and western sections of the area. The projected change in accumulated rainfall amount is similar to the occurrence in both seasons. In contrast to difference in occurrence between the two seasons, the projected daily rainfall intensity in CREs increases overall in both spring and fall. The projected increase in occurrence and/or intensifying in daily rainfall intensity imply a higher risk of geological hazards in TGR in future.
It has been well known that CREs occur under a more stable and longitudinally oriented circulation pattern dominated with blocking at middle-high latitudes (Ding et al., 2008;Luo et al., 2013). In the recent decades, the Arctic warms much faster than the middle-lower latitudes, the weakening of the north-south temperature gradient causes a reduction in the atmospheric baroclinicty and subsequently weakens the midlatitudinal westerly and a much broader meridional meanders in middle-high latitudes (Liu et al., 2012;Outten & Esau, 2012). The change might affect the atmospheric pattern related to CREs.
Here just the statistical downscaling scheme based on GCM outputs is used. In addition to the statistical downscaling, the dynamical downscaling with regional climate models is also an effective approach. It bears more physical meaning. Projecting the future trend of CREs in regional climate model like WRF consists of our future work.
There exists some uncertainty in the present study. First, climate simulations have larger uncertainty over mountainous areas like TGR than over plain basins (Palazzi et al., 2013(Palazzi et al., , 2015. Precipitation is much more poorly simulated than other variables such as air temperature, due to its strong localization, relatively sparse instrument samples, and the weaker physical constraints (Allen & Ingram, 2002). Also, the observational gridded data set used here, CN05.1, embraces uncertainty due to the adapted interpolation. Besides, the coarse spatial resolution of the CMIP5 models is also one major source of uncertainties (Birkinshaw et al., 2017). Finally, just several models (IPSL-CM5A-LR, IPSL-CM5A-MR, CNRM-CM5, MRI-CGCM3, and CSIRO-Mk3.6.0) analyzed here incorporate the direct effects and the first indirect effects of aerosols; this affects definitely the identification and projection of CREs since aerosols are essential for precipitation frequency (Jing et al., 2017). This is another source of uncertainty.

Appendix A: Abbreviation in the Context
The table lists all abbreviations mentioned in the context (Appendix A).

Data Availability Statement
Original station gauged data of CN05.1 can be obtained at a Chinese website (http://data.cma.cn/data/ cdcdetail/dataCode/SURF_CLI_CHN_MUL_DAY_V3.0.html) after registration, and relevant English information can be found online (at http://data.cma.cn/en/?r¼data/detail&dataCode¼SURF_CLI_CHN_MUL_ DAY_CES_V3.0). The "r1i1p1" daily precipitation from the 20 CMIP5 models listed in Table 1 is available online (at website https://esgf-node.llnl.gov/search/cmip5/). One can access it after registry; see the historical and RCPs runs from all the available models by selecting "Model" and then "Experiment", obtain daily outputs from the "r1i1p1" member by clicking "Time Frequency" and then "Ensemble", and download precipitation rate through "Variable Long Name on the left panel. Either the separated or packaged files can be obtained through selecting "HTTP" or "WGET".