Projection of Future Extreme Precipitation and Flood Changes of the Jinsha River Basin in China Based on CMIP5 Climate Models

Projecting future changes in extreme flood is critical for risk management. This paper presented an analysis of the implications of the Fifth Coupled Model Intercomparison Project Phase (CMIP5) climate models on the future flood in the Jinsha River Basin (JRB) in Southwest China, using the Xinanjiang (XAJ) hydrologic model. The bias-corrected and resampled results of the multimodel dataset came from the Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP). Relatively optimal general circulation models (GCMs) were selected with probability density functions (PDFs)-based assessment. These GCMs were coupled with the XAJ model to evaluate the impact of climate change on future extreme flood changes in the JRB. Two scenarios were chosen, namely: a midrange mitigation scenario (Representative Concentration Pathway 4.5, RCP4.5) and a high scenario (RCP8.5). Results show that: (1) The XAJ model performed well in simulating daily discharge and was suitable for the study area, with ENS and R2 higher than 0.8; (2) IPSL-CM5A-LR and MIROC-ESM-CHEM showed considerable skill in representing the observed PDFs of extreme precipitation. The average skill scores across the total area of the JRB were 0.41 to 0.66 and 0.53 to 0.67, respectively. Therefore, these two GCMs can be chosen to analyze the changes in extreme precipitation and flood in the future; (3) The average extreme precipitation under 20- and 50-year return period across the JRB were projected to increase by 1.0–33.7% under RCP4.5 and RCP8.5 during 2020 to 2050. The Upper basin is projected to experience the largest increase in extreme precipitation indices, possibly caused by a warmer climate. The extreme flood under 20- and 50-year return period will change by 0.8 to 23.8% and −6.2 to 28.2%, respectively, over this same future period. Most of scenarios projected an increase during the near future periods, implying the JRB would be likely to undergo more flooding in the future.


Introduction
Flooding is a climate-related disaster. Between 1980 and 2013, economic losses caused by floods have exceeded $1 trillion (2013 values). Over this same period, more than 220,000 people have lost their lives [1]. Global climate change has already begun to affect the earth's hydrologic cycle. Additionally, warmer climate will bring about high atmospheric moisture content, which will further lead to the increase of extreme precipitation events [2][3][4]. As flood is influenced greatly by precipitation, the risk of this kind of extreme hydrologic event is likely to grow. Thus, investigating flood changes in the context of climate change is imperative for risk management.
Most of the research on climate change impact use models for exploring 'what-if' questions. The main train of the research on flood risk management in regional climate change impact assessment and Yunnan-Guizhou Plateau to the mountain area of Southwest Sichuan, with a total length of 3464 km [21]. The Jinsha River's longest tributary is the Yalong River, whose total length is 1187 km (Figure 1). According to the data set provided by Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (RESDC), the total population in the JRB area was about 24 million in 2010, of which about 6.6 million in the Sichuan Province, about 13.8 million in the Yunnan Province and about 0.8 million in the Guizhou Province. The population in Qinghai and Tibet is low. The average annual precipitation of the JRB is approximately 710 mm: the annual precipitation of the lower reaches is approximately 900-1300 mm. The middle and upper reaches are mountainous canyon regions with an average annual precipitation of 600-800 mm. The annual precipitation in the source area upstream is less than 200 mm. Precipitation is mainly concentrated from June to October, which accounts for about 75-85% of annual precipitation (wet season). The spatial distribution of the temperature is similar to the precipitation. The overall trend is increasing from upstream towards downstream and from northwest to southeast. The average annual temperature of the JRB area is below 0 • C (upper reach), approximately 5 • C (middle reach) and above 10 • C (lower reach), respectively. According to statistics of discharge time series of 1950-2011 at the Pingshan station, which can be considered as the controlling hydrologic station, the annual average runoff of the JRB is 143 billion m 3 . The Jinsha River flooding along the mainstream is mainly caused by extreme precipitation over a long time period. Heavy rainfall mainly happens at the middle and lower reaches of the Jinsha River. Floods generally occur in late June to mid-October, and more frequently, from July to September. Jinsha River's longest tributary is the Yalong River, whose total length is 1187 km (Figure 1). According to the data set provided by Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (RESDC), the total population in the JRB area was about 24 million in 2010, of which about 6.6 million in the Sichuan Province, about 13.8 million in the Yunnan Province and about 0.8 million in the Guizhou Province. The population in Qinghai and Tibet is low. The average annual precipitation of the JRB is approximately 710 mm: the annual precipitation of the lower reaches is approximately 900-1300 mm. The middle and upper reaches are mountainous canyon regions with an average annual precipitation of 600-800 mm. The annual precipitation in the source area upstream is less than 200 mm. Precipitation is mainly concentrated from June to October, which accounts for about 75-85% of annual precipitation (wet season). The spatial distribution of the temperature is similar to the precipitation. The overall trend is increasing from upstream towards downstream and from northwest to southeast. The average annual temperature of the JRB area is below 0 °C (upper reach), approximately 5 °C (middle reach) and above 10 °C (lower reach), respectively. According to statistics of discharge time series of 1950-2011 at the Pingshan station, which can be considered as the controlling hydrologic station, the annual average runoff of the JRB is 143 billion m 3 . The Jinsha River flooding along the mainstream is mainly caused by extreme precipitation over a long time period. Heavy rainfall mainly happens at the middle and lower reaches of the Jinsha River. Floods generally occur in late June to mid-October, and more frequently, from July to September.

Observed Hydro-Meteorological Data
Observed daily precipitation, minimum and maximum temperature data was collected from China's Ground Precipitation 0.5° × 0.5° Gridded Dataset (V2.0) and China's Ground temperature 0.5° × 0.5° Gridded Dataset (V2.0) (http://data.cma.cn/), which was provided by the National Meteorological Information Center (NMIC) of the China Meteorological Administration (CMA). This dataset was based on the daily observations from 1961 to the present at 2474 national meteorological stations over the Chinese mainland. Data assessment showed that the gridded value had high

Observed Hydro-Meteorological Data
Observed daily precipitation, minimum and maximum temperature data was collected from China's Ground Precipitation 0.5 • × 0. mainland. Data assessment showed that the gridded value had high correlation and little error with observation; thus, the gridded value can be used to reflect the variations of precipitation and temperature. In this study, 208 grid cells covering the JRB were selected (Figure 2 correlation and little error with observation; thus, the gridded value can be used to reflect the variations of precipitation and temperature. In this study, 208 grid cells covering the JRB were selected ( Figure 2). Daily discharge data was collected from two hydrologic stations in the lower reach ( Figure 1). It was used to calibrate/validate the hydrology model and calculate the extreme flow (see Section 2.3). The basic information of the hydrologic stations is listed in Table 1. In this study, five global climate models in the IPCC AR5 report were selected: GFDL-ESM2M, HADGEM2-ES, IPSL-CM5A-LR, MIROC-ESM-CHEM and NORESM1-M. Information of each GCM is listed in Table 2. Each of these models offers continuous daily-series data, including historical simulations and future projection simulations forced under RCP2.6, RCP4.5, RCP6.0 and RCP8.5. These four RCP scenarios were named according to a possible range of global radiative forcing values in 2100 relative to pre-industrial values. It is assumed that the global radiative forcing will reach the highest point at ~3 W m −2 before 2100. If this value declines to 2.6 W m −2 by 2100, this corresponding scenario is RCP2.6; stabilizations without overshoot pathways to 6.0 and 4.5 W m −2 at stabilization after 2100 means RCP6.0 and RCP4.5; if this value rockets up to 8.5 W m −2 by 2100, the corresponding scenario is RCP8.5 [22]. The data was bias-corrected and resampled by the Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP, http://www.isi-mip.org). It covers the period from 1960 to 2099 on a horizontal grid with 0.5° × 0.5° resolution [23]. The data was bias-corrected to make sure the consistency of the long-term statistical characteristic of the whole data and the observed data from 1960-1999 [24]. This bias-correction method was specifically developed for this project to keep the absolute trends in temperature and relative trends in precipitation and the other variables [25]. There are 208 boxes in and around the JRB and the spatial resolution of data can meet the requirements of hydrological simulation in the JRB. Since this research is focusing on extreme precipitation and discharge, RCP4.5 (medium) and RCP8.5 (high) scenarios were chosen to project the extreme precipitation and discharge variation. Daily discharge data was collected from two hydrologic stations in the lower reach ( Figure 1). It was used to calibrate/validate the hydrology model and calculate the extreme flow (see Section 2.3). The basic information of the hydrologic stations is listed in Table 1. In this study, five global climate models in the IPCC AR5 report were selected: GFDL-ESM2M, HADGEM2-ES, IPSL-CM5A-LR, MIROC-ESM-CHEM and NORESM1-M. Information of each GCM is listed in Table 2. Each of these models offers continuous daily-series data, including historical simulations and future projection simulations forced under RCP2.6, RCP4.5, RCP6.0 and RCP8.5. These four RCP scenarios were named according to a possible range of global radiative forcing values in 2100 relative to pre-industrial values. It is assumed that the global radiative forcing will reach the highest point at~3 W m −2 before 2100. If this value declines to 2.6 W m −2 by 2100, this corresponding scenario is RCP2.6; stabilizations without overshoot pathways to 6.0 and 4.5 W m −2 at stabilization after 2100 means RCP6.0 and RCP4.5; if this value rockets up to 8.5 W m −2 by 2100, the corresponding scenario is RCP8.5 [22]. The data was bias-corrected and resampled by the Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP, http://www.isi-mip.org). It covers the period from 1960 to 2099 on a horizontal grid with 0.5 • × 0.5 • resolution [23]. The data was bias-corrected to make sure the consistency of the long-term statistical characteristic of the whole data and the observed data from 1960-1999 [24]. This bias-correction method was specifically developed for this project to keep the absolute trends in temperature and relative trends in precipitation and the other variables [25]. There are 208 boxes in and around the JRB and the spatial resolution of data can meet the requirements of hydrological simulation in the JRB. Since this research is focusing on extreme precipitation and discharge, RCP4.5 (medium) and RCP8.5 (high) scenarios were chosen to project the extreme precipitation and discharge variation.

XAJ Hydrologic Model
The Xinanjiang (XAJ) model was used in this study for daily discharge simulation. The XAJ model has been widely used in simulating rainfall-runoff, forecasting flood and planning water resources in humid and semi-humid regions [31]. It provides an integral structure to statistically describe the non-uniform distribution of runoff producing areas. It has been proved that, due to its description of vertical spatial distribution of soil moisture storage, the XAJ model performs better than other models [32,33]. Further detailed information relating to the XAJ model can be obtained from relevant references [34,35]. The model is driven by daily precipitation (P) and potential evapotranspiration (PE). In this study, the daily PE in the XAJ model was calculated by using the Hargreaves method based on daily maximum and minimum temperature [36]. The areal daily precipitation and potential evapotranspiration act as input to drive the XAJ model, and can be evaluated through Equation (1). The parameters of the XAJ model were calibrated by fitting the calculated daily discharge against the observed data. Correlation coefficient of linear regression equation (R 2 ) and Nash-Sutcliffe values (E NS ) [37] were chosen to quantify the model's performance of simulating daily discharge.
where P and PE are the areal precipitation and potential evapotranspiration, respectively; P i and PE i are the precipitation and potential evapotranspiration of the grid box i, Si is the area of grid box i; n is the number of grid boxes in the Huatan catchment or Pingshan catchment.

Extreme Precipitation and Flood Indices
The multi-day extreme precipitation and discharge indices, widely used in extreme events analyses [38], were selected to reflect extreme precipitation and flow in this study (Table 3). One type of index measures the annual maximum amount of daily precipitation, such as AMX1p, AMX3p and AMX7p. Another type of index measures the maximum discharge in successive n days, such as AMX1d, AMX3d and AMX7d. Table 3. Definition of extreme precipitation indices in this study.

AMX1p
Annual maximum 1-day precipitation mm AMX3p Annual maximum 3-day precipitation mm AMX7p Annual maximum 7-day precipitation mm AMX1d Annual maximum 1-day discharge m 3 /s AMX3d Annual maximum 3-day discharge m 3 /s AMX7d Annual maximum 7-day discharge m 3 /s A physically more meaningful and more relevant quantity for risk assessment is the probability of the variable (here, multi-day extreme precipitation and discharge) exceeding a certain level. The precipitation and discharge of annual maximum series under given return periods were calculated based on the generalized extreme value (GEV) distribution, which is widely utilized in modeling extreme events in meteorology and hydrology [19]. Equations (2) and (3) are the distribution function: where, ξ, α and k are parameters standing for location, scale and shape, respectively. In this study, the parameters were estimated by using the Maximum Likelihood Estimation method [39].
The flood magnitude (x T ) for T-year return period can be calculated as follows: In this study, return periods of 20 and 50 years were calculated and analyzed.

Selection of GCMs and Projection of Future Extreme Flood Changes
Coupled climate models were evaluated and relatively optimal GCMs were selected to do hydro-meteorological variability projection. The simulation of AMX1p, AMX3p and AMX7p in 208 cells of the JRB were evaluated based on probability density functions (PDFs). An alternative metric-skill score (SS) was defined to illustrate the similarity between two PDFs. The cumulative minimum value of two distributions of each binned value can be calculated with the SS, which means the overlapping area between two PDFs. The skill score (SS) is greater than 0 and smaller than 1. When SS is close to 1, it means that the simulated value fits perfectly with the observed value ( Figure 3a). When SS is close to 0, it means the overlap between simulated value and observed value is negligible ( Figure 3b) [17]. This can be seen in Equation (4): where, n stands for the number of bins; Fs n stands for the frequency of values in a given bin from the model; and Fo n stands for the frequency of values in a given bin from the observed data. Adding up the minimum frequency values over all bins and SS can be obtained. In this research, GEV was utilized to obtain PDFs of the observed and simulated value.
where, ξ, α and k are parameters standing for location, scale and shape, respectively. In this study, the parameters were estimated by using the Maximum Likelihood Estimation method [39]. The flood magnitude (xT) for T-year return period can be calculated as follows: In this study, return periods of 20 and 50 years were calculated and analyzed.

Selection of GCMs and Projection of Future Extreme Flood Changes
Coupled climate models were evaluated and relatively optimal GCMs were selected to do hydro-meteorological variability projection. The simulation of AMX1p, AMX3p and AMX7p in 208 cells of the JRB were evaluated based on probability density functions (PDFs). An alternative metricskill score (SS) was defined to illustrate the similarity between two PDFs. The cumulative minimum value of two distributions of each binned value can be calculated with the SS, which means the overlapping area between two PDFs. The skill score (SS) is greater than 0 and smaller than 1. When SS is close to 1, it means that the simulated value fits perfectly with the observed value ( Figure 3a). When SS is close to 0, it means the overlap between simulated value and observed value is negligible ( Figure 3b) [17]. This can be seen in Equation (4): where, n stands for the number of bins; Fsn stands for the frequency of values in a given bin from the model; and Fon stands for the frequency of values in a given bin from the observed data. Adding up the minimum frequency values over all bins and SS can be obtained. In this research, GEV was utilized to obtain PDFs of the observed and simulated value. The bias was chosen in this research to assess the reconstructing capacity of GCMs to rebuild the precipitation characteristics, so as to test the rationality of using skill score to assess climate models. Specific equations are listed as follows:  The bias was chosen in this research to assess the reconstructing capacity of GCMs to rebuild the precipitation characteristics, so as to test the rationality of using skill score to assess climate models. Specific equations are listed as follows: where, B is the absolute value of relative deviation; P s and P o are the simulated and observed multi-year average extreme precipitation; B is greater than 0 and smaller than 1; the closer B is to 0, the better the performance of the GCM. Driven by the baseline, daily climatology and the projected future daily climate data sets coming from the relatively optimal GCMs, the calibrated XAJ model can output the simulated daily discharge in both baseline (1961 to 1990) and future period (2020 to 2050). With the parameterized GEV distribution in different periods, relative changes in AMXnp and AMXnd under the 20-and 50-year return period can be estimated and represent changes in discharge in the future.

Historical Discharge Simulations with XAJ Model
Based on the 208 cells with daily precipitation, maximum and minimum temperature, input driving data of the XAJ model were generated. The recorded data series was divided into calibration period (before 1990) and validation period (after 1990). Figure 4 shows the simulated and observed daily discharge in the calibration period and validation period. Table 4 summarizes the assessment of the XAJ model in the Huatan and Pinghshan stations. In Table 4, it can be seen that E NS and R 2 are both higher than 0.8 in both calibration and validation period, which means fair performance of the XAJ model. Generally, a daily E NS of 0.65 or higher means the simulation is very good [40]. That is to say, the established XAJ model was acceptable to do daily discharge simulation in the JRB.
where, B is the absolute value of relative deviation; s P and o P are the simulated and observed multi-year average extreme precipitation; B is greater than 0 and smaller than 1; the closer B is to 0, the better the performance of the GCM. Driven by the baseline, daily climatology and the projected future daily climate data sets coming from the relatively optimal GCMs, the calibrated XAJ model can output the simulated daily discharge in both baseline (1961 to 1990) and future period (2020 to 2050). With the parameterized GEV distribution in different periods, relative changes in AMXnp and AMXnd under the 20-and 50-year return period can be estimated and represent changes in discharge in the future.

Historical Discharge Simulations with XAJ Model
Based on the 208 cells with daily precipitation, maximum and minimum temperature, input driving data of the XAJ model were generated. The recorded data series was divided into calibration period (before 1990) and validation period (after 1990). Figure 4 shows the simulated and observed daily discharge in the calibration period and validation period. Table 4 summarizes the assessment of the XAJ model in the Huatan and Pinghshan stations. In Table 4, it can be seen that ENS and R 2 are both higher than 0.8 in both calibration and validation period, which means fair performance of the XAJ model. Generally, a daily ENS of 0.65 or higher means the simulation is very good [40]. That is to say, the established XAJ model was acceptable to do daily discharge simulation in the JRB.

Comparison of GCM Simulations with Observations
With the observed PDF and GCM PDF for each grid square during 1961 to 2000, the SSs estimated for individual cells can be calculated. Average values of SS and B over four sub-catchments (namely, the Upper basin, Yalong River Basin, Middle basin, Lower basin and the Jinsha River Basin) are illustrated in Figure 5. Although several variations exist among different models, the five GCMs in this study performed better in the Lower basin than that in the Middle and Upper basins and better for AMX3p and AMX7p than AMX1p. Compared to the other three GCMs, IPSL-CM5A-LR and MIROC-ESM-CHEM were capable of capturing the characteristics of extreme precipitation statistical distribution and average value, especially for AMX3p and AMX7p. Taking the AMX3p as an example, the average SS (B) across the JRB for IPSL-CM5A-LR and MIROC-ESM-CHEM were 0.60 (0.25) and 0.67 (0.18), respectively, which were superior to the projection of other models. Based on the above analysis, the outputs that came from IPSL-CM5A-LR and MIROC-ESM-CHEM were chosen for the analysis of future extreme flood changes.

Comparison of GCM Simulations with Observations
With the observed PDF and GCM PDF for each grid square during 1961 to 2000, the SSs estimated for individual cells can be calculated. Average values of SS and B over four sub-catchments (namely, the Upper basin, Yalong River Basin, Middle basin, Lower basin and the Jinsha River Basin) are illustrated in Figure 5. Although several variations exist among different models, the five GCMs in this study performed better in the Lower basin than that in the Middle and Upper basins and better for AMX3p and AMX7p than AMX1p. Compared to the other three GCMs, IPSL-CM5A-LR and MIROC-ESM-CHEM were capable of capturing the characteristics of extreme precipitation statistical distribution and average value, especially for AMX3p and AMX7p. Taking the AMX3p as an example, the average SS (B) across the JRB for IPSL-CM5A-LR and MIROC-ESM-CHEM were 0.60 (0.25) and 0.67 (0.18), respectively, which were superior to the projection of other models. Based on the above analysis, the outputs that came from IPSL-CM5A-LR and MIROC-ESM-CHEM were chosen for the analysis of future extreme flood changes.

Changes in Extreme Precipitation for 2020 to 2050
Flood is most likely to be triggered by extreme precipitation [41,42]. Hence, extreme precipitation events variations (AMX1p, AMX3p and AMX7p) from the reference period (1961 to 1990) and the projection period (2020 to 2050) were firstly analyzed. The calculation of relative change between simulated value in projection period and that in the reference period could illustrate the evolution trend, and thus, eliminate the impacts of systematic deviation caused by GCMs. Only the above selected optimal GCMs (IPSL-CM5A-LR and MIROC-ESM-CHEM) were utilized for the analysis of the future extreme precipitation, so as to reduce the uncertainty from the GCMs. Figure 6 shows the relative changes of the average AMXnp under 20-year return period (P = 5%) and 50-year return period (P = 2%) in the four sub-basins and the whole JRB between the baseline and future scenarios. The two relatively optimal GCMs project a consistent increase in AMXnp-Tyr (n = 1, 3, 7; T = 20a, 50a) by 1.0-33.7% across the JRB under RCP4.5 and RCP8.5 from 2020 to 2050. Among the four sub-basins of JRB, the upper basin will face the largest increase in AMX1p (14.4 to 56.3%) and AMX3p (13.9-53.9%) and AMX7p (18.0 to 50.1%). These results indicate that in the near 30 years, the JRB, especially the upstream basin, will suffer from intensified extreme floods.
According to the spatial distribution of extreme precipitation shown by Figure A1, Boxplots of AMXnp change in each grid-box can be obtained (Figure 7). Thus, the spatial difference of extreme precipitation in different combined condition of GCMs and RCPs can be quantitatively described (longer box means bigger difference). It can be seen that in RCP4.5, spatial difference of extreme precipitation projected by IPSL-CM5A-LR was relatively small: extreme precipitation variation in grid-box ranges from 0 to 15%. While the spatial difference of extreme precipitation projected by

Changes in Extreme Precipitation for 2020 to 2050
Flood is most likely to be triggered by extreme precipitation [41,42]. Hence, extreme precipitation events variations (AMX1p, AMX3p and AMX7p) from the reference period (1961 to 1990) and the projection period (2020 to 2050) were firstly analyzed. The calculation of relative change between simulated value in projection period and that in the reference period could illustrate the evolution trend, and thus, eliminate the impacts of systematic deviation caused by GCMs. Only the above selected optimal GCMs (IPSL-CM5A-LR and MIROC-ESM-CHEM) were utilized for the analysis of the future extreme precipitation, so as to reduce the uncertainty from the GCMs. Figure 6 shows the relative changes of the average AMXnp under 20-year return period (P = 5%) and 50-year return period (P = 2%) in the four sub-basins and the whole JRB between the baseline and future scenarios. The two relatively optimal GCMs project a consistent increase in AMXnp-Tyr (n = 1, 3, 7; T = 20a, 50a) by 1.0-33.7% across the JRB under RCP4.5 and RCP8.5 from 2020 to 2050. Among the four sub-basins of JRB, the upper basin will face the largest increase in AMX1p (14.4 to 56.3%) and AMX3p (13.9-53.9%) and AMX7p (18.0 to 50.1%). These results indicate that in the near 30 years, the JRB, especially the upstream basin, will suffer from intensified extreme floods.
According to the spatial distribution of extreme precipitation shown by Figure A1, Boxplots of AMXnp change in each grid-box can be obtained (Figure 7). Thus, the spatial difference of extreme precipitation in different combined condition of GCMs and RCPs can be quantitatively described (longer box means bigger difference). It can be seen that in RCP4.5, spatial difference of extreme precipitation projected by IPSL-CM5A-LR was relatively small: extreme precipitation variation in grid-box ranges from 0 to 15%. While the spatial difference of extreme precipitation projected by MIROC-ESM-CHEM was relatively big: extreme precipitation variation in grid-box ranges from 10% to 35%; in the meantime, the precipitation variation was bigger than that projected by IPSL-CM5A-LR. In RCP8.5, extreme precipitation variation projected by both IPSL-CM5A-LR and MIROC-ESM-CHEM will range from 5% to 25%. MIROC-ESM-CHEM was relatively big: extreme precipitation variation in grid-box ranges from 10% to 35%; in the meantime, the precipitation variation was bigger than that projected by IPSL-CM5A-LR. In RCP8.5, extreme precipitation variation projected by both IPSL-CM5A-LR and MIROC-ESM-CHEM will range from 5% to 25%.     MIROC-ESM-CHEM was relatively big: extreme precipitation variation in grid-box ranges from 10% to 35%; in the meantime, the precipitation variation was bigger than that projected by IPSL-CM5A-LR. In RCP8.5, extreme precipitation variation projected by both IPSL-CM5A-LR and MIROC-ESM-CHEM will range from 5% to 25%.    Figure 7. Boxplot of AMXnp-Tyr (n = 1, 3, 7; T = 20a, 50a) change in each grid-box for different RCPs and GCMs. Y-axis is the ratio of projected value (P f ) to historical value (P h ). P f /P h > 1 means the increase of AMXnp; P f /P h < 1 means the decrease of AMXnp. The box portion represents the 25th, 50th to 75th percentile, and the whiskers represent the 10th percentile and the 90th percentile. Figure 8 illustrated the GEV frequency distributions of AMXnd in Pingshan, which can be used to compare the frequency of extreme floods between baseline and future periods. An increase trend has been detected in most climate change scenarios when frequency was less than 10%. However, in the same frequency range, variation of AMXnd projected by IPSL-CM5A-LR and MIROC-ESM-CHEM showed a significant difference. In RCP4.5, AMXnd projected by IPSL-CM5A-LR varied between 12 to 17%, while AMX1d, AMX3d and AMX7d projected by MIROC-ESM-CHEM were −12 to 6%, −5 to 6% and 8 to 9% respectively. When the frequency was less than 2%, AMX1d and AMX3d showed a decreasing trend; in RCP8.5, AMXnd projected by both IPSL-CM5A-LR and MIROC-ESM-CHEM showed an increasing trend; 18 to 30% and 0 to 15%, respectively.  Figure 8 illustrated the GEV frequency distributions of AMXnd in Pingshan, which can be used to compare the frequency of extreme floods between baseline and future periods. An increase trend has been detected in most climate change scenarios when frequency was less than 10%. However, in the same frequency range, variation of AMXnd projected by IPSL-CM5A-LR and MIROC-ESM-CHEM showed a significant difference. In RCP4.5, AMXnd projected by IPSL-CM5A-LR varied between 12 to 17%, while AMX1d, AMX3d and AMX7d projected by MIROC-ESM-CHEM were −12 to 6%, −5 to 6% and 8 to 9% respectively. When the frequency was less than 2%, AMX1d and AMX3d showed a decreasing trend; in RCP8.5, AMXnd projected by both IPSL-CM5A-LR and MIROC-ESM-CHEM showed an increasing trend; 18 to 30% and 0 to 15%, respectively. To further investigate the relative changes in AMXnd, two different return periods (20-year and 50-year) were chosen. As shown in Figure 9, the largest percentage increases in AMXnd were mainly found for the RCP8.5 scenario of IPSL-CM5A-LR in 2020-2050. Take the 50-year return period as an example: the AMX1d, AMX3d and AMX7d would increase by 12.9%, 15.0% and 16.3%, respectively, in the Pingshan Station. The decrease of AMX1d and AMX3d were found for the RCP4.5 scenario of MIROC-ESM-CHEM in the Pingshan Station (−6.2% in AMX1d and −1.6% in AMX3d for the 50-year return period). AMXnd-Tyr (n = 1, 3, 7; T = 20a, 50a) showed an increasing trend in all the other combination situation of scenarios and GCMs. The range of relative changes in Pingshan are described here, when the results from the RCP4.5 and RCP8.5 scenarios of the two relatively optimal GCMs are considered.. When considering the results in RCP4.5 and RCP8.5 scenarios of the two relatively optimal GCMs, the range of relative changes in Pingshan were described as follows. For AMX1d, the relative changes ranged from 0.8 to 19.9% for the 20-year return period and from −6.2 to 20.1% for the 50-year return period. For AMX3d, the relative changes ranged from 2.9 to 19.9% for the 20-year return period and from −1.6 to 21.4% for the 50-year return period. For AMX7d, the relative changes ranged from 9.0 to 23.8% for the 20-year return period and from 3.6 to 28.2% for the 50-year return period. RCP4.5_P f /P h and RCP8.5_ P f /P h means the ratio between projected value (P f ) and historical value (P h ). P f /P h > 1 means increasing AMXnd while P f /P h < 1 means decreasing AMXnd.

Changes in Extreme Floods for 2020 to 2050
To further investigate the relative changes in AMXnd, two different return periods (20-year and 50-year) were chosen. As shown in Figure 9, the largest percentage increases in AMXnd were mainly found for the RCP8.5 scenario of IPSL-CM5A-LR in 2020-2050. Take the 50-year return period as an example: the AMX1d, AMX3d and AMX7d would increase by 12.9%, 15.0% and 16.3%, respectively, in the Pingshan Station. The decrease of AMX1d and AMX3d were found for the RCP4.5 scenario of MIROC-ESM-CHEM in the Pingshan Station (−6.2% in AMX1d and −1.6% in AMX3d for the 50-year return period). AMXnd-Tyr (n = 1, 3, 7; T = 20a, 50a) showed an increasing trend in all the other combination situation of scenarios and GCMs. The range of relative changes in Pingshan are described here, when the results from the RCP4.5 and RCP8.5 scenarios of the two relatively optimal GCMs are considered.. When considering the results in RCP4.5 and RCP8.5 scenarios of the two relatively optimal GCMs, the range of relative changes in Pingshan were described as follows. For AMX1d, the relative changes ranged from 0.8 to 19.9% for the 20-year return period and from −6.2 to 20.1% for the 50-year return period. For AMX3d, the relative changes ranged from 2.9 to 19.9% for the 20-year return period and from −1.6 to 21.4% for the 50-year return period. For AMX7d, the relative changes ranged from 9.0 to 23.8% for the 20-year return period and from 3.6 to 28.2% for the 50-year return period.

Discussions
The future extreme precipitation and flood variation of the JRB were analyzed in this study. Predominantly, an increase of projected extreme precipitation and flood has been informed by the GCMs. Most of these findings are consistent with previous studies. Both Huang et al. and Zhang et al. found an increasing trend in extreme precipitation indices in the upper Yangtze reaches [18,43]. Su et al. found an increase of 0.3 to 13.1% for peak discharge in the upper Yangtze River Basin during 2036 to 2065 [44]. The study of Gu et al. also showed that flood frequency and magnitude in the upper Yangtze River Basin are likely to increase significantly in the future. The AMX1d and AMX7d with return periods of 50, 20 and 10 years may turn into hydrological extreme events with return periods of approximately 15, 7 and 3 years [45]. There has been evidence showing that in the next decades, the JRB is likely to suffer from more floods.
In this study, we used multiple GCMs and RCP scenarios and one hydrological model to discuss the possible variation in extreme floods. Although we have assessed the performance of multiple climate models and selected the most suitable climate models for this research before analyzing the future variation of flood, uncertainty caused by future climate conditions still exist in future flood projection. Similar research was carried out by Bell et al. and Kay et al. [46,47]. Another uncertainty source is the structure and parameters of the hydrological models. Since the impact of this kind of uncertainty is much less than that in GCMs [8,48], this research did not take the uncertainties from hydrological models into consideration. Still, this does not mean that this kind of uncertainty should be totally ignored. Taking the XAJ model used in this study as an example, the potential evapotranspiration (PET) can be estimated by the Penman-Monteith method or Hargreaves method. More meteorological data, such as solar radiation, is needed when using Penman-Monteith method, while Hargreaves method only needs temperature data. Since the data used by the XAJ model in this research is the gridded dataset provided by National Meteorological Information Center, which only involves precipitation and temperature, Hargreaves method was utilized to calculate PET. As a result, the projection of future flood did not take variation of solar radiation, wind speed or relative humidity into consideration. Furthermore, we assumed that the hydrological model parameters calibrated from observation are still valid when they are used in future climate simulations. It is on this basis that the impact of climate change could be evaluated by a hydrological model [49]. Actually, as pointed by Merz et al., hydrological model parameters may potentially change if calibrated to different periods [50]. Therefore, this study will do further research on the uncertainty produced by model structure and its parameters. Different hydrological models and PET computation methods shall be used to compare differences of response to climate change in the JRB.
The data of climate change scenarios used in this study is provided by Inter-Sectoral Impact Model Intercomparison Project, which has been bias-corrected and resampled with 0.5° × 0.5° resolution, which also means that the original GCMs data has been statistically downscaled. Compared with the dynamical downscaling method based on Regional Climate Model (RCM), statistical downscaling method lacks solid mathematical and physics foundation. Thus, projected flood variation using the data in this research may have difference in the flood projection using RCMs' output data. In the further research, RCMs such as PRECIS could be used to analyze the impact of

Discussions
The future extreme precipitation and flood variation of the JRB were analyzed in this study. Predominantly, an increase of projected extreme precipitation and flood has been informed by the GCMs. Most of these findings are consistent with previous studies. Both Huang et al. and Zhang et al. found an increasing trend in extreme precipitation indices in the upper Yangtze reaches [18,43]. Su et al. found an increase of 0.3 to 13.1% for peak discharge in the upper Yangtze River Basin during 2036 to 2065 [44]. The study of Gu et al. also showed that flood frequency and magnitude in the upper Yangtze River Basin are likely to increase significantly in the future. The AMX1d and AMX7d with return periods of 50, 20 and 10 years may turn into hydrological extreme events with return periods of approximately 15, 7 and 3 years [45]. There has been evidence showing that in the next decades, the JRB is likely to suffer from more floods.
In this study, we used multiple GCMs and RCP scenarios and one hydrological model to discuss the possible variation in extreme floods. Although we have assessed the performance of multiple climate models and selected the most suitable climate models for this research before analyzing the future variation of flood, uncertainty caused by future climate conditions still exist in future flood projection. Similar research was carried out by Bell et al. and Kay et al. [46,47]. Another uncertainty source is the structure and parameters of the hydrological models. Since the impact of this kind of uncertainty is much less than that in GCMs [8,48], this research did not take the uncertainties from hydrological models into consideration. Still, this does not mean that this kind of uncertainty should be totally ignored. Taking the XAJ model used in this study as an example, the potential evapotranspiration (PET) can be estimated by the Penman-Monteith method or Hargreaves method. More meteorological data, such as solar radiation, is needed when using Penman-Monteith method, while Hargreaves method only needs temperature data. Since the data used by the XAJ model in this research is the gridded dataset provided by National Meteorological Information Center, which only involves precipitation and temperature, Hargreaves method was utilized to calculate PET. As a result, the projection of future flood did not take variation of solar radiation, wind speed or relative humidity into consideration. Furthermore, we assumed that the hydrological model parameters calibrated from observation are still valid when they are used in future climate simulations. It is on this basis that the impact of climate change could be evaluated by a hydrological model [49]. Actually, as pointed by Merz et al., hydrological model parameters may potentially change if calibrated to different periods [50]. Therefore, this study will do further research on the uncertainty produced by model structure and its parameters. Different hydrological models and PET computation methods shall be used to compare differences of response to climate change in the JRB.
The data of climate change scenarios used in this study is provided by Inter-Sectoral Impact Model Intercomparison Project, which has been bias-corrected and resampled with 0.5 • × 0.5 • resolution, which also means that the original GCMs data has been statistically downscaled. Compared with the dynamical downscaling method based on Regional Climate Model (RCM), statistical downscaling method lacks solid mathematical and physics foundation. Thus, projected flood variation using the data in this research may have difference in the flood projection using RCMs' output data. In the further research, RCMs such as PRECIS could be used to analyze the impact of climate change on extreme precipitation and flood in JRB and the result of two studies could be compared. Due to the daily precipitation provided by GCM, research on hourly flood peak is hard to carried out using this data set to analyze the impact of climate change on flood. Thus, annual maximum n-day discharge on hydrologic station was chosen to characterize the impact of climate change on flood. In the meantime, the hydrological model (the XAJ model), which can describe the relationship between rainfall and runoff well, was selected. However, the impact of climate change on designed flood hydrograph was the most concerned issue in flood risk management. In the following research, time set method can be utilized to obtain precipitation data with shorter temporal scale (such as hourly scale) based on the GCMs' output. Based on this, the storm event watershed model, e.g. the FLOW-R2D model, can be used to do hydrodynamic simulation, so as to assess the impact of climate change on future typical flood propagation [51][52][53].

Conclusions
This research studied the projected extreme precipitation and flood variation of the JRB in RCP4.5 (medium) and RCP8.5 (high) scenarios based on five CMIP5 GCMs. The XAJ model was employed to simulate daily discharge, using 0.5 • × 0.5 • grid cells across the JRB for the baseline (1961 to 1990) and for the future period (2020 to 2050). Six extreme indices (AMXnp and AMXnd, n = 1, 3, 7) were chosen for the analysis.
Since E NS and R 2 were both higher than 0.8, it was suggested that the XAJ model performed well in simulating the daily discharge. This model can be used to estimate the impact of climate change on future flood in the JRB. Both the PDF-based assessment and model bias were used to evaluate the GCMs performance in the JRB. In general, IPSL-CM5A-LR and MIROC-ESM-CHEM showed considerable skill in representing the observed AMXnp, which can be chosen as the relatively optimal GCMs to analyze the changes in extreme precipitation and flood in the future.
As expected, extreme precipitation indices derived from precipitation projections of the GCMs for JRB exhibit a wide range of possible changes. The average AMXnp (P = 2% and P = 5%) across the JRB were projected to increase by 1.0-33.7% during 2020 to 2050. However, in different combination of climate models and scenarios, extreme precipitation showed a unanimous increasing trend, indicating that extreme precipitation will occur more frequently and more severe in the JRB in the future. According to the extreme precipitation projection revealed in this research, the Upper basin may suffer from the largest increase in extreme precipitation. This phenomenon may be caused by the melting and evaporation of glacier in the mountain area of Yangtze River source region due to global warming.
The projection of extreme floods indicates that AMX1d, AMX3d and AMX7d showed very similar trends during the near future periods (2020-2050). The largest percentage increases in AMXnd were found for the RCP8.5 scenario of IPSL-CM5A-LR in 2020-2050, whereas the AMXnd will increase slightly and may even decrease for the RCP4.5 scenario of MIROC-ESM-CHEM. There is a relatively large variability in the projected ranges of AMXnd under the different future climate change scenarios. However, most of scenarios projected an increase during the near future periods (compared to the baseline 1961 to 1990). Overall, the percentage changes in AMX1d, AMX3d and AMX7d under 50-year return period ranged from −6.2 to 20.1%, −1.6 to 21.4% and 3.6 to 28.2%, respectively. This research is a preliminary assessment of future extreme precipitation and flood in the JRB. Although uncertainties unavoidably exist in the projections, the results should be taken with care because the extreme floods of the JRB would be likely to become more frequent and intense in the next several decades. Furthermore, combining this research with the existing hydraulic engineering conditions, potential submerged area caused by extreme flood could be simulated. Then, the affected inhabitants, buildings and agricultural land could be identified, so as to provide scientific support for the drawing of flood risk map in the context of climate change.
Author Contributions: Z.Y. and J.X. worked together in forming ideas of this paper; Z.Y. and Y.W. worked together in calculating and writing of this manuscript; Z.Y. carried out the calculation and analysis of flood in the study area; J.X. provided supervision during the whole process.

Conflicts of Interest:
The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results.

Conflicts of Interest:
The authors declare no conflict of interest. The founding sponsors had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, and in the decision to publish the results.
Appendix A Figure A1. Changes in AMX1p, AMX3p and AMX7p with 20-/50-year return period from baseline to future.