3.1. Seasonality variability in CMIP6 historical data
The four seasons are distinct in the IRB of northwest China, the climate is dry, and the distribution of precipitation is extremely uneven during the year. The maximum annual average precipitation in winter and autumn in various regions is around 200mm, while the value in spring and summer is around 450mm. The average summer rainfall in the Gangdise Mountains in the southeast exceeds 450mm, while the average winter rainfall in the desert area does not exceed 5mm. Fig. 2 shows the comparison of the seasonal variation between the actual precipitation in CN05.1 and the selected four CMIP6 climate model datasets.in the IRB. These data were divided into four seasons, namely winter (December-February), spring (March-May), summer (June-August), and autumn (September- November).
Figure 2 displays that the maximum annual average precipitation is 240mm in winter (Fig. 2a-2e), 500mm in spring (Fig. 2f-2j), 550mm in summer (Fig. 2k-2o), and 220mm in autumn (Fig. 2p-2t). The IRB has complex landform types, including snow-capped mountains, oases, and deserts, which makes precipitation in seasons show huge spatial differences. Comparing the basic data of the selected CMIP6 climate models, an almost similar spatial distribution of precipitation was observed in all climate models. In spring, the simulated precipitation of each model shows the decreasing trend from northwest and southwest to northeast and southeast, while the precipitation shows a decreasing trend from southwest to northeast in summer.
Comparing the precipitation distribution maps of each season horizontally, the precipitation data of the IL model is the closest to the historical observations. The IL model can well represent the extremely high value of precipitation in the precipitation data of spring and summer, while the GE model can capture that in winter and autumn. The MI model overestimates the precipitation in the IRB, especially in the southwestern region of each season. The ME model's estimation of precipitation in this area is not stable enough. The winter is generally low, and the other three seasons are low in some places and high in some places.
3.2. Model evaluation and bias correction
Figure 3 shows the percentage biases of the selected CMIP6 GCMs and the change results of the RMSE concerning the observed precipitation of each grid during the period 1985-2014. In Figs. (3a-3d), the negative and positive values respectively represent the underestimation and overestimation of the precipitation data based on the four models of CMIP6 GCMs compared with the observed data. In the GE, ME, and MI plots, the deviation values of most regions are between -50% and 100%, while the deviation values of most regions in the IL plot are between -100% and 75%. Figs. 3a-3d shows that the deviations in the northern and southern regions are generally between -50% and 50%, and the areas with large deviations are mainly concentrated in the southwest and central desert regions. The comparison of the deviations of these four plots shows that the deviation of the GE plot is relatively small, while the deviation of the MI graph is relatively large. In Figs. (3e-3h), the RMSE of almost all regions remains at 0-5, and only the RMSE of individual regions in the northeast and southwest exceeds 5. Comparing the RMSEs of the four plots, the RMSE of the IL plot is overall smaller, while the MI graph is overall larger. It reflects that the IL model-based precipitation data and the observed data have a small degree of dispersion, which tends to be stable.
Figure 4 reflects the deviation percentages of the four models after correction with the EDCDF compared to the observed data. Figs. 4a-4c shows that the deviation percentages in most regions are between -1% and 1%, and the deviation percentages in individual central regions such as the Taklimakan Desert are between -4% and 4%. The overall deviation correction effect of the IL plot is better, except that the deviation value of the individual areas in the middle is kept at about 5%.
3.3. Variability in precipitation extremes during historical events
Figure 5 shows the average extreme precipitation indices calculated using the observed precipitation datasets during the period 1961-2014. Figs. 5a-5i shows the average values of the observed precipitation, PRCPTOT, SDII, CDD, CWD, P1025, PG25, Rx1day, and Rx5day calculated from 1961 to 2014, respectively. It can be seen that the southeast region and the vicinity of the Tianshan Mountains have the highest precipitation from Fig. 5a, while the central desert region has the least precipitation. The PRCPTOT in Fig. 5b and annual average precipitation in Fig. 5a shows similar distributions. From Fig. 5b and Fig. 5c, it can be seen that the PRCPTOT and the SDII have a strong correlation, especially in the southeast and northwest regions, while the correlation is poor in the central desert area and the Kunlun Mountain area.
The maximum observed CDD appears in the Taklimakan Desert area, and the minimum appears in the northwestern region (Fig. 5d). The CWD value of the IRB in the northwest is generally small, and the CWD value is large only in the eastern part of Qilian, near the Tianshan Mountains and the Qinghai-Tibet Plateau in the south (Fig. 5e). The number of moderate rain days with daily precipitation of 10-25mm only occurs in the Tianshan Mountains, the southern part of Gangdise Mountains, the eastern part of Qilian Mountains, and the Daxingan Mountains in the northeast (Fig. 5f). Heavy rain with a daily rainfall of more than 25mm only occurs in the Daxinganling Mountains in the northeast, and the annual average value is 1.5 days (Fig. 5g). The distribution of Rx1day and Rx5day in the IRB is highly consistent, showing a trend of decreasing from the southeast and northwest to the middle (Fig. 5h-5i). The maximum value of Rx1day is about 35mm, and the maximum value of Rx5day is about 60mm.
3.4. Analysis of extreme precipitation in the future
Figures 6a-6d and Figs. 6e-6h respectively shows the change rate of the PRCPTOT of the four selected GCMs in the NF (2021-2050) and the FF (2071-2100) compared to the observed data, and Figs. 6i-6l and Figs. 6m-6p respectively shows the percentage change of SDII in the two periods. The PRCPTOT change rate of different GCMs in different periods showed a trend of decreasing from the center to the southeast and northwest. The spatial distribution of the PRCPTOT change rate in the NF and the FF of the same GCM is relatively similar. The rate of change of the PRCPTOT in the NF and the FF has not changed much in the Altai Mountains, Tianshan Mountains, Gangdise Mountains, Qilian Mountains, Helan Mountains, and Daxingan Mountains. However, the change rate of the PRCPTOT has changed greatly in the Gurbantunggut Desert, the Taklimakan Desert, and the Hami-Yinshan region. The PRCPTOT change rates of GE and ME in the FF are greater than those in the NF in the central Taklimakan Desert, the Gurbantungut Desert, and the Hami-Yinshan region (Figs. 6a-6b and 6e-6f). The PRCPTOT index of the Taklimakan Desert in the FF is 6 times that of the observed data, while that of the NF is only 2 times. In the MI and IL models (Figs. 6c-6d and 6g-6h), both times (i.e., NF and FF) have shown a similar spatial distribution of the PRCPTOT, that is, the annual precipitation has increased significantly in the desert area.
In the GE and ME models (Fig. 6i and 6j), the percentage change of SDII in the NF shows that the surrounding areas such as Tianshan, Kunlun, Qilian, and Daxinganling Mountains are around -10%, while the central desert area is around +10%, and the Gangdise Mountain is around +30%. The percentage change of the SDII in the FF of the Tianshan, Kunlun, Qilian, and Taklimakan Desert areas will decrease compared to the NF, while the Gangdise Mountains in the southeast will increase (Fig. 6m and 6n). In the MI and IL models (Fig. 6k and 6l), the percentage change of the SDII in the NF is similar to the other two models, except that the values of some grids in desert areas are more negative. However, some grid values in the central desert area will decrease in the FF and will decrease in the south and east, especially in the Gangdise Mountains and Qilian Mountains (Fig. 6o and 6p).
Figures 7a-7d and Figs. 7i-7l respectively displays the percent of change CDD and CWD indices of the four selected GCMs compared to the observed value during the NF time (2021-2050), while Figs. 7e-7h and Figs. 7m-7p displays that of the FF (2071-2100). In the GE and ME models (Figs. 7a-7b and 7e-7j), it can be seen that some grids in the southeast and northeast regions show positive changes in different periods, such as Qilian Mountains, Gangdisi Mountains, and Daxinganling Mountains, while most of the grids in the desert, northwest and southwest regions, such as the Kunlun Mountains and the Altai Mountains show negative changes. In the MI and IL models (Figs. 7c-7d and 7g-7h), the distribution of the CDD change percentage in the NF is similar to the other two models, but there will be subtle changes in the FF, which are mainly reflected in the Tianshan Mountains and the Altai Mountains. The change rate of some grids in the Altai Mountains and Tianshan Mountains has increased by more than 80%. In general, the CDD change percentage of most grids in different periods and different models corresponds to -60% to +20%.
In the GE and ME models (Fig. 7i and 7j), the CWD change percentage of most grids will exceed 40%, and only some grids will show negative values in the NF, such as Tianshan, Kunlun Mountains, and Qilian Mountains. However, most of the grids in the FF are above 80%, except for the Qinghai-Tibet Plateau from the Kunlun Mountains to Gangdis Mountains, Qilian Mountains, and Tianshan Mountains (Fig. 7m and 7n). In the MI and IL models (Fig. 7k and 7l), the distribution is quite different from the other two models in the NF, and most of the grids exceed 80%. Only a small part of the grids are negative, mainly distributed in Tianshan, Kunlun, and Qilian Mountains. The distribution of the percentage change of the CWD is nearly half of the negative value in the FF, distributed in mountainous areas, such as the Altai Mountains, Tianshan Mountains, and the Kunlun Mountains, as well as desert areas (Fig. 7o and 7p). In general, the CWD change percentage of most grids in different periods and different models corresponds to -10% to +100%.
Figures 8a-8d and Figs. 8i-8l respectively shows the percentage of change in P1025 and PG25 of the four selected GCMs in the NF, while Figs. 8e-8h and Figs. 8m-8p shows that of the FF. In the GE model (Fig. 8a), most of the grids in the entire watershed show positive values in P1025, especially in desert areas, while a few grids in Tianshan, the Kunlun Mountains, and Qilian Mountains show negative values. Compared with the NF (Fig. 8e), the spatial distribution of the percentage change has a strong consistency in the FF, but the positive value is not obvious in the desert area. In the ME model (Fig. 8b and 8f), both NF and FF show a spatial distribution similar to the GE model in P1025. However, the percentage changes of the grids in the central desert area and the Gangdise Mountains are slightly different from those of the GE model in the NF. In the MI and IL models (Figs. 8c-8d and 8g-8h), the percentage change of the P1025 of most grids in the NF and the FF corresponds to 0 to 10%, some grids in desert areas exceed 20%, and the grid values of Tianshan, Kunlun, and Gangdese mountains correspond to -5%.
In the GE and ME models (Figs. 8i-8j and 8m-8n), the percentage change of PG25 for most grids in the Tianshan, Kunlun, Qilian, and Daxinganling Mountains regions is negative in different periods, while most grids in the Taklimakan Desert, Hami-Yinshan, and Gangdise Mountains correspond to +40%. Comparing the grid changes of the two models in different periods, it is found that the positive grid in the FF shows a significant decreasing trend. In the MI and IL models (Figs. 8k-8l and 8o-8p), compared with the observed value, the percentage change of PG25 is concentrated in the Taklimakan Desert, the Gangdise Mountains of the Qinghai-Tibet Plateau, the Hami-Yinshan Mountains, and the Qilian Mountains.
Figures 9a-9d and Figs. 9i-9l respectively shows the percentage of change in Rx1day and Rx5day of the four selected GCMs in the NF, while Figs. 9e-9h and Figs. 9m-9p shows that of the FF. For most of the IRB, the range of variation is from -30% to +100%. In the GE and ME models (Figs. 9a-9b, 9e-9f, 9i-9j, and 9m-9n), they display almost similar characteristics of Rx1day and Rx5day in different periods. The grids of each graph are positive, except for some grids in the Altai Mountains, Tianshan Mountains, Kunlun Mountains, Qilian Mountains, and Hami regions. However, the spatial distributions shown by MI and IL models are different from GE and ME (Figs. 9c-9d, 9g-9h, 9k-9l, and 9o-9p). And the percentage of changes in the grid reflected by them is generally low.