GPM is a new generation of global satellite precipitation product after TRMM (Tropical Rainfall Measuring Mission), which has the advantages of wide coverage and sophisticated inversion algorithm. It is widely used in the study of climate change and hydrological processes (Hou et al., 2014; Tang et al., 2016). This research chooses version is GPM-IMERG (Integrated Multi-satellite Retrievals for GPM) V6 Final-Run (https://gpm.nasa.gov/data/directory). It provides global daily precipitation data of 0.1°×0.1° from June 2000 to the present (February 2023).

MSWEP is a multi-source integrated precipitation product derived from the advantages of station data, remote sensing data and reanalysis data, characterized by a long time scale and high spatial resolution (Alijanian et al., 2017; Beck et al., 2019). Compared with MSWEP V1, MSWEP V2.2 used in this study improved the spatial resolution from 0.25° to 0.1°, and the time series also improved from 1979 to 2015 to 1979-2020, providing 3°h gridded precipitation data (http://www.gloh2o.org/).

In this paper, the overlapping period from 2000 to 2020, which can be obtained from multi-source data, is selected for research during downscaling. The projection, rotation, format conversion and unit standardization of GPM and MSWEP data were carried out, and the corresponding monthly, seasonal and annual precipitation data were combined by raster calculator for downscaling.

Tianshan Mountains are characterized by complex terrain with large elevation changes (321–7426 m) and a temperate continental climate with obvious fluctuations (Domrös and Peng, 2012). The extreme temperature in winter and summer leads to uneven precipitation in Tianshan Mountains (Domrös and Peng, 2012; Chen et al., 2016). Therefore, considering the selection of environmental variables from the aspects of terrain factor and climate factor, it provides a necessary basis for the change trend of precipitation downscaling model in the study area. DEM is derived from SRTM data with a spatial resolution of 90 m (http://gscloud.cn/), and slope and aspect are extracted by ArcGIS as topographic factors. Climate factors come from China Meteorological Data Network (http://data.cma.cn/), including temperature (TEM), relative humidity (RHU), wind speed (WIN), sunshine duration (SSD), maximum temperature (MAX) and minimum temperature (MIN). Factors strongly correlated with precipitation were selected as explanatory variables by correlation test, and grid data were formed by Kriging interpolation (Seo et al., 2015). In order to facilitate downscaling, explanatory variables need to maintain the same resolution as precipitation products. The convolution method (Kumar et al., 2017) was used to resampling raster data of altitude, slope, aspect and climate factors to 0.1° and 1 km. Compared with the nearest neighbor method and bilinear method, the convolution method takes into account both the resolution and fidelity of the data, and generally achieves better resampling effect (Awada et al., 2019).

The ground observation data were obtained from the daily scale sites of the National Meteorological Information Network (http://www.nmic.gov.cn/). Quality control was carried out on all the station data (Li et al., 2017), and the missing data was obtained by linear interpolation using the latest available data. Limited by the complex terrain and inconvenient transportation in mountainous areas, the available meteorological stations in Tianshan Mountains are very sparse. The spatial distribution of stations is shown in Figure 1A. A total of 36 stations are used in this study, and the available time series is from 1951 to 2020.

In this study, a combination of geographically weighted regression and Kriging interpolation method is proposed to downscale GPM and MSWEP (spatial resolution is 0.1°×0.1°), and through precision test to generate a high resolution (1 km×1 km) precipitation dataset, which is applied to the spatiotemporal monitoring of drought characteristics in the study area. The realization of downscaling is based on two basic assumptions: first, there is a correlation between precipitation and the selected environmental variable factors, and it can be explained by the established model; Secondly, the model established at a lower spatial resolution consistent with the precipitation product can also use the interpolation results of environmental variables to predict the precipitation at a higher resolution, which is also called the “relational scale invariant” hypothesis (Duan and Li, 2016). The downscaling model construction and data processing process is shown in Figure 2. Specific steps are as follows:

1. At the time scale, the daily precipitation data of GPM and MSWEP and environmental variables were unified into the monthly scale, and the factors with strong correlation with precipitation were selected as explanatory variables through correlation test.

As an extension of the traditional least squares regression method, GWR was first proposed by Brunsdon (Brunsdon et al., 1998), which described the dynamic relationship between the causal variables and explanatory variables in the geographic space, rather than the single regression that took the study area as a whole (Lu et al., 2022). Therefore, it has been widely applied in the study of complex scale-dependent characteristics between meteorological elements and topographic elements in high-altitude mountainous areas (Chu, 2012; Kumari et al., 2017). The mathematical expression of GWR model is as follows: Y j = β 0 ( u j , v j ) + ∑ i = 1 p β i u j , v j X i , j + ε j   i = 1,2,3 , ⋯ , n (1)

In the formula, Y j represents the j th observed value of the dependent variable, and X i , j represents the j th observed value of the i th independent variable. u j , v j , β 0 u j , v j , β k u j , v j and ε j are the geographical coordinates, intercept, slope (regression coefficient) and regression residuals at the j th point, respectively. p is the number of environment variables. This formula is based on the regression coefficient established by the nearby observation point j . The basic idea of the weight matrix is to assume that the closer the distance to the point j , the higher the weight affected. The weight matrix can be expressed as the following formula: β ^ ( u j , v j ) = X T W ( u j , v j ) X − 1 X T W ( u j , v j ) P (2)

Among them, β ^ u j , v j represents the j th point near data partial least-squares regression model is set up, X and P respectively the independent variable and dependent variable; W u j , v j is the weight matrix. In this study, the weight matrix can be expressed as the following formula: w i , j = 1 − d i j 2 θ i k 2 , d i j < θ i k 0 , d i j > θ i k (3)

Where, w i , j is the observation weight of the j th point, used to observe the coefficient at the position of the i th point. d i j represents the Euclidean distance between point j and point i ; θ i k represents the adaptive bandwidth size defined as the k th nearest neighbor distance.

MGWR is an extended form of GWR, which deals with mixed coefficients with spatial variation (local variables) and fixed or constant coefficients (global variables) in calculation (Fotheringham et al., 2017). Generally speaking, MGWR is a generalized GWR model, which introduces additional global variables to better represent the spatial heterogeneity of precipitation in the regression model. In this study, GVT was used to identify spatial variation (local) and fixed (global) environmental variables in the study area. Diff of Criterion (DIF) is adopted as the index to judge the fixed and variable terms in the model (Mei et al., 2006). A positive value of DIF indicates that the factor is specified as a local variable in the GWR model and as a global variable otherwise. In the GWR model, we repeated GVT for all environment variables to determine fixed and varying variables, so as to facilitate the construction of MGWR model. The mathematical expression of MGWR model is as follows: R j = ∑ k = 1 n α k X j , k + ∑ k = n + 1 m β k u j , v j X j , k + ε j , j , k = 1,2,3 , ⋯ , n (4)

Where, R j is the value of the dependent variable; X j , k is the value of the explanatory variable; u j , v j is the coordinate of the j th sampling point; k is the number of independent variables; α k is the regression coefficient of the global variable; β k is the regression coefficient of local variable. ε j is a random error.

The accuracy verification of the datasets includes the performance comparison of the two downscaling models and the accuracy assessment of the spatial-temporal correlation between the downscaling results and the measured precipitation, which is divided into the following two steps:

1. Performance comparison of the two models. Through Akaike information criterion correction (AICc) and residual sum of squares, the downscaling results of two commonly used spatial weight functions (Chao et al., 2018) (adaptive bivariate weight function and fixed Gaussian weight function) under two different models (GWR and MGWR) were respectively compared with the original satellite precipitation data. It is believed that smaller AICc and residual sum of squares are more favorable (Abdollahipour et al., 2021).

Where: n is the length of precipitation series; S_i and G_i are grid precipitation data and site measured precipitation data, respectively. S and G correspond to the average time series of the two precipitation data.

The variation of drought trend is an important index of drought monitoring. This study analyzed the variation of dry and wet trends in different subregions of Tianshan Mountains based on Sen trend and Mann-Kendall (MK) significance test. M-K non-parametric test method can judge the mutation points of the sequence and the trend of each interval, so as to realize the significance test of the trend analysis results (Huang et al., 2015).

The Standardized Precipitation Index (SPI) was proposed by McKee (McKee et al., 1993) in 1993, which only relies on precipitation data to quantitatively describe regional drought conditions. Research (Chang et al., 2018; Adarsh and Janga Reddy, 2019) showed that SPI(SPI3) on a 3-month time scale can better reflect soil water deficit in the underlying surface of arid northwest China. Therefore, SPI3 is selected as the drought research index in this study, and its range is [-3,3], the value represents the degree of dry and wet with positive values representing humidity and negative values representing drought, and the larger the value is, the wetter it is. Supplementary Table S1 is the drought grade table of standardized precipitation index SPI3 obtained according to the national standard of Meteorological Drought Grade GB/T20481-2006.

Based on the run course theory (Yevjevich, 1967), this study identified drought events and quantitatively analyzed the drought characteristics of drought events, as shown in Figure 3. The paper only focuses on drought events with a long duration and a high drought peak: drought (SPI<0) for 3 consecutive months or more and the peak value less than −1.0 (Guo et al., 2022). The time when SPI starts to be negative and recovers to be positive are defined as the beginning and end time of drought event respectively, and the period is the duration of drought (Formula 10). Drought severity is defined as the cumulative value of SPI during drought (Formula 11). Drought intensity is defined as the ratio of drought severity to drought duration (Formula 12). Peak value of drought is the minimum value reached by SPI during drought, and the corresponding month is the peak drought time (Formula 13). M D T = ∑ i = 1 N D T i N (10) M D S = ∑ j = 1 N D S j N , D S = ∑ i = 1 D T S P I i (11) M D I = ∑ j = 1 N D I j N , D I = ∑ i = 1 D T S P I i D T (12) M D P = ∑ j = 1 N D P j N , D P = max 1 ≪ i ≪ D D S P I i (13)

Where, D T 、 D S 、 D I and D P are the duration, severity, intensity and peak value of the drought event respectively. M D T 、 M D S 、 M D I and M D P are the average duration, average severity, average intensity and average peak value of all drought events during the study period. i is the dry month. S P I i is the SPI value of the month i ; N is the total number of drought events identified during the study period; j is a single drought event.

To analyse the accuracy of downscaling results of different products in different time series, the precipitation data of measured stations in Tianshan Mountains and the raster data of corresponding grid points were included into the same series, and the downscaling data were analysed on monthly, seasonal and annual scales, respectively. Figure 6 shows the scatter density of measured monthly precipitation at rainfall stations and GPM, gGPM, mGPM, MSWEP, gMSWEP, and mMSWEP precipitation at corresponding locations during 2000–2020, respectively. It can be seen from Figures 5A, D that both the GPM (BIAS_GPM=−6.21%) and MSWEP (BIAS_MSWEP=−4.83%) underestimated the monthly precipitation in the Tianshan Mountains, and MSWEP (CC=0.63) had a higher correlation with the measured precipitation than GPM (CC=0.51). For GPM, the GWR model can significantly improve the underestimation (BIAS_GPM=−6.21%; BIAS_gGPM=−2.65%) and the degree of aggregation of scatter points (Figure 5B), while mGPM (Figure 5C) is better than gGPM, but the difference is very small. The downscaling comparison results of MSWEP are different from those of GPM. Compared with GWR, the correlation coefficient of MGWR model results (CC_gMSWEP=0.77; CC_mMSWEP=0.93) and Nash coefficient (NSE_gMSWEP =0.70; NSE_mMSWEP =0.88) is further improved, and quantitative error (RMSE_gMSWEP =7.5; RMSE_mMSWEP =1.3) decreases significantly, which has stronger variance interpretation ability for rainfall at stations (R_gMSWEP ²=0.77; R_mMSWEP ² = 0.96). This may be because compared with GWR, which uses a single kernel function to calculate the weight of impact factors, MGWR can capture the difference of impact scales of different variables, so that the impact of terrain and climate factors fitted by the model on precipitation is more in line with reality. For MSWEP, which is itself the result of multi-source data fusion, the data can be revised twice (Sun et al., 2022). However, GPM, based on multi-satellite global observation system, does not consider the influence of local topography on precipitation in the study area (Nwachukwu et al., 2020). The introduction of ground data through GWR model can effectively improve the fitting degree of gGPM and measured precipitation, but whether to consider the different scales of impact factors (MGWR) does not have much positive impact on the improvement of data accuracy. It may introduce more biases.

To further refine the accuracy differences of different downscaling results in each month of the year, the time series of the correlation between GPM (Figure 6A), MSWEP (Figure 6B), their downscaling results and the measured monthly precipitation during the study period are drawn in Figure 6. The correlation of six kinds of data presents regular changes of volatility during the year. The original grid data of GPM showed a weak correlation with the measured data (CC>0.2). The results of GWR and MGWR showed a similar improvement to the original data, especially a very strong correlation with the measured precipitation around June of each year (CC>0.8). MSWEP and the measured data basically showed a moderate correlation (CC>0.4), except for some months, gMSWEP and mMSWEP realized a strong correlation (CC>0.6) and a very strong correlation (CC>0.8) with the measured data, respectively. The six kinds of data show a similar law in each month of the year: they all reach the maximum of annual correlation in June-August when measured precipitation is abundant, and they reach the minimum in December-February of the next year when precipitation is scarce. To further verify the seasonal characteristics of the correlation, the accuracy differences of six kinds of data are compared and verified in different seasons and annual average precipitation scales in Figure 9.

Based on the ground site data and the corresponding grid data extracted from GPM and MSWEP and their two downscaling results respectively, this section analyses the applicability of six kinds of data at the annual scale. The Taylor diagram (Supplementary Figure S3) is used to draw the SDR, CC and RMSE of six kinds of precipitation datasets. The position of each point on the Taylor diagram represents the matching degree between the dataset and the measured data. The closer the distance between the dot of the dataset and the dot of the site, the stronger the correlation and the smaller the error. The results of the six datasets have strong correlation with the observed precipitation data of summer (Supplementary Figure S3C) and autumn (Supplementary Figure S3D) in the Tianshan Mountains. The applicability of the annual scale (Supplementary Figure S3A) ranks the second, and the correlation of spring (Supplementary Figure S3B) and winter (Supplementary Figure S3E) is the worst. This verifies the discussion result of the month series correlation in Figure 6: whether satellite products or their downscaling results, their similarity to the measured data in different seasons is positively correlated with the precipitation abundance, which may be related to the causes of seasonal precipitation in the study area: In winter and early spring, the Tianshan Mountains are affected by the Siberian air mass (Zhang et al., 2012), and the weather is dry and cold, resulting in trace precipitation with very low intensity and perennial snow cover in the high altitude area, which greatly interferes with the precipitation detection of remote sensing products (Zhang et al., 2018). However, in summer and autumn (especially in summer), the precipitation mainly comes from the air mass with water vapor in the Atlantic (Zhang et al., 2016). When they pass through the Tianshan Mountains with complex terrain, they form convective rain with high intensity. Meanwhile, there is less snow and ice cover in summer, which brings higher accuracy for satellite remote sensing monitoring precipitation. As can be seen from Supplementary Figure S3, in most cases, the order of six data dots and observation stations from near to far is: mMSWEP, mGPM, gMSWEP, gGPM, MSWEP, GPM, which is also the precision order of the six to identify the annual precipitation in the Tianshan Mountains. In general, the correlation between the mMSWEP dataset and the measured data at different time scales is more than 0.8, indicating that MGWR has the best scaling effect on MSWEP, and can achieve high-precision precipitation inversion in mountainous areas with high spatial resolution.

Figures 7A–G shows the spatial distribution of satellite products, downscaling results and interpolated monthly mean precipitation at stations in the Tianshan Mountains during 2000–2020. Both the satellite products and the downscaling results can basically capture the spatial precipitation pattern of more precipitation in the West and less in the East, more in the North and less in the South of the Tianshan Mountains. However, both GPM (Figure 7A) and MSWEP (Figure 7B) seem to have missed the high precipitation in the Tianshan Mountains. Especially, it is difficult for GPM to continuously capture the differences in the spatial distribution of precipitation caused by the mountainous terrain. The downscaling models of GWR and MGWR can significantly improve the phenomenon that GPM and MSWEP satellite products underestimate the high value of measured precipitation (Figure 7G). The difference is that although MGWR-GPM (Figure 7C) can better improve the spatial discontinuities of GWR-GPM (Figure 7B) (such as the aggregation of high precipitation values in the West subregion of Tianshan Mountains), the former can significantly weaken the impact of the geographically weighted regression method on precipitation results by integrating background fields of different explanatory variables to achieve downscaling. However, it cannot be completely eliminated and even magnified in local areas (such as the “ox’s eye” phenomenon at the junction peak of the North and South Tianshan Mountains). For MSWEP (Figures 7D–F), compared with the downscaling results of GWR (Figure 7E) and MGWR (Figure 7F), the “ox’s eye” phenomenon was overcome, and the phenomenon of low overestimation and high underestimation of the measured precipitation by satellite products was more significantly improved (Figure 7G).

To further quantify the spatial distribution difference of downscaling results’ error, box charts of the accuracy statistical indicators of GPM and MSWEP and their downscaling results in different subregions are shown in Figure 8. The correlation coefficient and the mean relative deviation between monthly precipitation and measured precipitation of the 6 datasets analysed in Section 3.3.1 (Figure 5) (marked successively on the right axis of Figure 8) were used as the measurement standard for the regional differences. All kinds of products and their downscaling results showed a relatively consistent rule: the correlation with the measured precipitation showed a downward trend from Northwest to Southeast, which was basically consistent with the spatial distribution trend of precipitation in the Tianshan Mountains with more precipitation in the West and less precipitation in the East, more precipitation in the North and less precipitation in the South (Figure 7G). The correlation coefficient is used to reflect the consistency between data, so it is also necessary to discuss the accuracy change of the overall data in different subregions through the spatial distribution of the relative deviation: different data all showed a small underestimation of the precipitation in the West (−6.21%<BIAS<0), and the precipitation in the West subregion near the Ili Valley is stable with little topographic fluctuation and obvious regularity of precipitation (Figure 7G). Satellite remote sensing products can well detect the distribution characteristics of precipitation. The low estimation of precipitation in the South is greater than that in the North, and the GWR and MGWR models effectively narrow the deviation range between the North and South. GPM and MSWEP and their downscaling results show the phenomenon of underestimated and overestimated precipitation in the East subregion respectively, and the relative maximum deviation range in the four subregions. The East Tianshan Mountains are mainly distributed in the Turpan-Ha Basin (Domrös and Peng, 2012), which is far away from the ocean and the terrain is closed. Moreover, under the influence of the hot and dry subtropical continental air mass, it is difficult for the water vapor at the top of the atmosphere to fall to the ground completely (Hu et al., 2016). This results in the monitoring error of precipitation products.

In particular, there are some outliers in the correlation and relative deviation between gGPM and measured precipitation in different sub-regions, and the grid locations of these outliers basically appear at the aggregation of outliers in Figure 7B, which further indicates that GWR has a certain applicability limitation to the downscaling effect of GPM, and there is still a large room for improvement In summary, mMSWEP has the highest fitting degree with the spatial distribution of measured precipitation in Tianshan Mountains, and can reflect the precipitation differentiation characteristics in different subregions with a high resolution (1 km×1 km).

According to the research results in Section 3.3, compared with GPM, the mMSWEP dataset obtained through the downscaling of MGWR model has higher accuracy in different time series and spatial scales, and its correlation of monthly precipitation is basically above 0.7, meeting the basic requirements of drought detection using remote sensing data (Tan et al., 2018). Therefore, in this section, high-resolution (1km×1 km) and long-time series (1979–2020) mMSWEP dataset will be used to invert the spatial and temporal distribution of drought characteristics in the Tianshan Mountains and provide technical support for drought monitoring, early warning and drought relief in the northwest mountainous areas of China.

Based on the 3-month SPI index (SPI3), this section uses the Sen slope coupled MK test method to analyse the spatiotemporal variation of dry and wet conditions in the Tianshan Mountains during 1979–2020, pixel by pixel. The pixels that passed the MK test accounted for 94.5% of the whole Tianshan Mountains. According to the calculation results, the slope estimation results were divided into five grades (Figure 9). Among them, 71% of the areas showed a significant trend of wetting, 15% were slightly wet areas, only 9% showed a significant trend of drying, mainly concentrated in the East Tianshan Mountains, but a small part near the Alatao Mountain in the west of North and the west of the South subregion. The variation trend of dryness and wetness shown in Figure 9 is basically consistent with the distribution pattern of average drought zones reflected in Figure 2 according to the station data over the years -- the North subregion is wetter than the South and the West is wetter than the East, but it shows a richer and clearer spatial texture, and more details of drought are also monitored.

At the subregional scale, SPI3 drought index was calculated based on regional mean precipitation data and time series of Tianshan Mountains and its four subregions were analysed (Figure 10). During 1979-2020, the overall wetting trend of Tianshan Mountains was obvious, with a slope of 0.0016 (Figure 10E), and passed the significance level test of 0.05. In the past 40 years, the basic order of the wetting trend of different regions in Tianshan Mountains is as follows: North (Figure 10B)> West (Figure 10A)> East (Figure 10D) > South (Figure 10C), which is basically consistent with the spatial distribution of Figure 9. Among them, the North and South regions passed the significance level test of 0.05, and the change trend of the West and East only passed the significance level test of 0.1. Since 2000, the overall wetting trend has been intensified, which may be due to the enhanced water vapor transport from high latitudes in the northern Hemisphere with the global warming since the 21st century, and the “three-stage water vapor relay transport” from the tropical Indian Ocean and Arabian Sea to the Central Asian mountains through the western Indian Ocean, the southern Arabian Peninsula and Central Asia (Guo et al., 2022).

To explore the drought events and their characteristics in the Tianshan Mountains, drought events were extracted pixel by pixel based on SPI3 and range-length theory, and the number of drought events and the average of their characteristics (severity, intensity, and peak value) were calculated based on the pixel by pixel. The results are shown in Figure 11.

Figure 11A shows the frequency of drought events identified during 1979–2020. As can be seen from the figure, droughts have occurred more than 30 times in most areas of the East Tianshan Mountains in the past 40 years. In the North and West, the frequency of drought is low, and the frequency is less than 10 times. As can be seen from Figure 11B, the drought in the north slope of the North Tianshan Mountains and the area near Halkatawu Mountain in the West Tianshan Mountains has a longer duration, with an average duration of 10 months Figure 11C shows the spatial distribution of the average severity of all drought events over the past 40 years. It is not difficult to see that most regions with a longer duration of drought also have a higher severity, such as the junction of the West region and the Northern slope of Kunlun Mountains and the vicinity of Yilian Habirga Mountain in the North region, indicating that the drought severity in these areas is mainly affected by the longer duration of drought. The difference is that parts of high severity also occur in the North region of the Tianshan Mountains and the western part of the South region of the Tianshan Mountains, and the drought severity in these areas is mainly affected by the high intensity (Figure 11D).

Based on the SPI index calculated by the average precipitation in the study area, the drought events in the Tianshan Mountains during 1979–2020 were identified in combination with the run theory method, and the duration, intensity, severity, peak value, and other basic characteristics were quantified. Based on the three-month SPI drought index, 16 drought events can be classified by consensus, and the basic characteristics of drought events are shown in Table 3.

Five typical drought events were selected and analysed one by one according to the duration, severity and intensity of drought, the underlined contents in Table 3 were the five typical drought events and their drought characteristics. Figure 12 shows the spatial distribution of peak drought in the five drought events.

Typical drought event one is the drought event with the longest duration and the highest severity. The drought began in January 1983 and lasted for 39 months. It reached its peak in September 1985 with the absolute value of SPI reaching 2.61, which was in an extreme drought state. In March 1986, the drought condition of this drought event returned to the normal level. Figure 12A shows the spatial distribution of drought in the Tianshan Mountains in September 1985. As can be seen from the figure, 96% of the Tianshan Mountain area is in the state of drought, especially the West, North and South of the Tianshan Mountain, and the west of the East region near the Turpan-Hami Basin show SPI value <-2, which is in the state of extreme drought. The area of extreme drought accounts for 30% of the total area of the Tianshan Mountain area. Mahmood (Mahmood et al., 2010) obtained a similar conclusion in related studies that extreme drought events occurred in the Tianshan Mountains during 1983–1986, and the drought was mainly concentrated in the western part of the Tianshan Mountains.

Typical drought event two began in November 1991 and ended in January 1992. As can be seen from Table 3, the drought event lasted only 3 months, with a high drought intensity of 1.22 and a drought severity of 8.04. This drought event is a typical type of short-time high-intensity drought. Figure 12B shows the spatial distribution of SPI drought level in December 1991. It can be seen from the figure that drought covered more than 98% of the Tianshan Mountains in that month, and the areas of severe and extreme drought accounted for 15% and 27%. The North and South Tianshan Mountains were affected by severe and extreme drought levels. The Western and Eastern sections of the East region showed extreme drought and slight drought, respectively. The opposite of this drought event is typical drought event 5 (Figure 12E), which belongs to the typical short-time low-intensity drought type. In addition to the severe and extreme drought in the West of the Tianshan Mountains and the eastern part of the South region, the East region, and the North region the of Tianshan Mountains mainly have mild drought and no drought, respectively. When studying the drought in Xinjiang, Yao et al. (Yao et al., 2018b) proposed that the drought in 2009 was a low-intensity and low-impact drought event, mainly affecting parts of the southern slope of the Tianshan Mountains. According to statistics of Xinjiang Water Resources Department (Communique, 2009) in 2009, drought affected 49.9×10⁴hm² of crops in the whole region, including 1.83×10⁴hm² of no harvest area and 597×10⁴hm² of severe drought grassland. The drinking water of 165 ×10⁴ million people and 37.8×10⁴ million livestock were seriously affected. Among them, Aksu, Baicheng and Kumish areas are suffering from severe drought (which are located near the West of the Tianshan Mountains and the Eastern part of the South).

Typical drought event 3 was the most intense drought event among all the drought events, and its intensity value reached 1.60. It occurred from July 1997 to December 1997 and lasted for 6 months with a severity of 6.60. Such short duration but high intensity drought events often have a greater impact on ecosystems and vegetation than long time, low intensity drought events. Figure 12C shows the SPI classification distribution in August, the most severe month of drought. As can be seen from the figure, about 86% of the area is in a state of drought, and about 25% of the area is in an extreme state of drought. These areas are mainly distributed in the West and middle part of the East Tianshan Mountains.

Typical drought event four belongs to the drought event with long duration but lowest drought intensity. The drought event began in April 2004 and ended in July 2005. The intensity of drought was only 0.46, the duration was 16 months, and the severity was only 4.23. Li et al.(2016) pointed out that 2004 was a typical drought year in Xinjiang, and most regions of Xinjiang were affected by drought to some extent, including the Tianshan Mountains. In July 2004, the drought event was the most severe month. 70% of the Tianshan Mountain area was in the state of drought, but the intensity was not high, and the drought was mainly slight (Figure 12D). Only a small part of the East Tianshan Mountains was in extreme drought. According to statistics of Xinjiang Water Resources Department (Communique, 2009), the drought in 2004 was the longest in 15 years, compared with 1990–2004.

Based on the spatial analysis of drought characteristics (Figure 11) and typical drought events (Figure 12), the West region has drought characteristics of low frequency, long duration, and high severity, while the East region has completely opposite drought characteristics: high frequency, short duration, and low severity. Occasional short - term high - severity drought events occur in North and South regions.