Spatial gathering characteristics of drought in the Qinghai-Tibet Plateau

Due to climate change, drought has caused serious impacts on the eco-environment, hydrology and agriculture, and drought events in the Qinghai-Tibet Plateau (QTP) have become more severer and frequent; therefore, understanding the characteristics and variations of drought is crucial to reduce its eco-environmental and socio-economic impacts. This study used the Standardized Precipitation Evapotranspiration Index (SPEI) to identify drought and assessed its interannual changes from 1980 to 2020. Then, the Hurst exponent and intensity analysis were used to identify future drought trend and the characteristics of drought intensity. Moreover, Empirical Orthogonal Function (EOF) analysis was performed to examine the main spatial gathering characteristics. The results indicate that: 1) the QTP was becoming wetter in general, and drying places were mainly distributed in the southeast and northeast of the QTP, as well as the Qaidam Basin accounting for 27% areas of QTP. 2) The trend of wet and dry in the future in most regions would be the same as the present, only 10% of the regions would have the reverse trend. 3) The rate of wet/drought changes was faster in the 1980s and 2000s. 4) The EOF mode1 revealed a gathering distribution structure with negative values in the southeast and east of the QTP and positive values in the center and west. The west was more sensitive to the variation of dry and wet, and most areas will continue to be wet in the future. EOF mode2 and mode3 mainly indicated opposite gathering patterns of southwest-northeast and south-north. The results provide favorable evidence for policymakers to better understand and prevent drought.


Introduction
Few extreme events can lead to more tremendous damage than drought (Dai, 2011). Drought can lead to serious negative impacts on water resources, agricultural production, environments, and ecosystems (Potop et al., 2014). Though the number of drought disasters amounts to only 5% of the total disasters, the loss can reach 30% of the loss resulting from all natural disasters (Wang et al., 2014). Drought brings the world about 221 billion dollars loss annually for the past 60 years (Tong et al., 2018). The number and duration of drought events would increase in the future (Blunden et al., 2011), and dryland will occupy 50% of the world's land area before the end of the 21st century (Huang et al., 2016). So it is essential to accurately understand and identify the characteristics and variations of drought to provide an early warning for decision-making groups (Hayes et al., 2011).
Generally, there are four categories of drought: meteorological drought, hydrological drought, agricultural drought, and socio-economic drought (Mishra and Singh, 2010). Many indices have been developed to characterize different categories of drought, such as the Palmer Drought Severity Index (PDSI) (Palmer, 1965), Standardized Precipitation Evapotranspiration Index (SPEI) (Vicente-Serrano et al., 2010), Standardized Precipitation Index (SPI) (McKee et al., 1993), Surface Water Supply Index (SWSI) (Mishra and Singh, 2010), and Standardized Runoff Index (SRI) (Shukla and Wood, 2008). Meteorological drought is usually the first step in drought occurrence and also contributes to the occurrence of the other three drought categories (Eslamian et al., 2017). After considering the applicability of the meteorological drought index, SPEI is chosen to identify drought in this paper. It combines the advantages of PDSI sensitivity to potential evapotranspiration (PET) as well as takes into account the advantages of the SPI multi-time scale (Yu et al., 2014).
The Qinghai-Tibet Plateau (QTP), known as the primary water source for Asia, provides more than two billion people with water directly and indirectly (Rangwala and Miller, 2012). The QTP is sensitive and vulnerable to climate change, which will have a significant influence on the surrounding regions and the worldwide climate (Kang et al., 2010). Due to climate warming (You et al., 2021), drought is particularly prevalent in the QTP, affecting vegetation carbon uptake (Ye et al., 2020), growth , and further aggravating water stress . At the same time, the drought had a serious impact on some basins, approximately 29 drought events occurred in the Yangtze River Basin from 1979 to 2012 (Zhang et al., 2016), and the long-term average drought in the Yellow River Basin lasted nearly 5.8 months from 1953 to 2012 (Huang et al., 2015). Moreover, the discharge of rivers that originate in the QTP shrunk drastically due to drought (Wang et al., 2011). Since 1960, the Yellow River discharge has dropped to zero 3 times, and the longest duration recorded is 226 days (Liang et al., 2010). The Yangtze River Basin experienced the severest drought in the spring of 2011, and the rainfall is at the lowest level since 1961 (Lu et al., 2014). Therefore, studying the long-term drought development in the QTP is important to understand the causes and characteristics of drought better, guide the decision-making departments to monitor and analyze drought, and prevent local agriculture and animal husbandry from drought disasters.
In recent decades, drought indices are beginning to be used to identify and analyze the drought characteristics in the QTP. A study based on the SPEI found that drought intensity occurred more seriously before the 1990s in the QTP, then the

FIGURE 1
Location and elevation of Qinghai-Tibet Plateau.

Frontiers in Environmental Science
frontiersin.org degree of drought trended to ease, and the evolution of drought had noticeable interdecadal differences after 1997 (Liang et al., 2018). During the growth season, though the northeastern and southern parts experienced a significant wetting trend, the eastern fringe of the QTP tended to be dry, and this trend will continue (Zhang et al., 2019). The altitude dependence of drought characteristics was investigated based on SPEI and 4,000 m a. s. l was determined as the dividing line, the higher the altitude, the greater the changing trend (Feng et al., 2020). However, the spatial gathering characteristics of drought resulting from climate change are still unclear in the QTP, particularly in the conditions of the complex topography, environments, and lack of observational meteorological data. In this study, the 12-month-scale SPEI over 41 years from 1980 to 2020 was used to identify the specific spatial gathering characteristics of the QTP. The objectives of this paper are: 1) to depict the changing trend of drought; 2) to predict the future change trend of drought; 3) to analyze the rate and intensity of drought transformation in each time interval and 4) to detect the spatial gathering characteristics of drought.

Study area and data
The QTP, referred to as the "Asian water tower" and "The Third Pole," is located in western China and covers an area of 2.5 million square kilometers (Immerzeel et al., 2010). The QTP is from 300 km to 1,500 km wide from 24°37′27″N to 41°09′13″N, and 2,800 km long from 73°18′52″E to 106°19′30″E, with an average altitude of above 4,000 m a. s. l. The plateau suffers from strong solar radiation, low temperature, and air pressure, as well as distinct seasonal and spatial precipitation heterogeneity due to its unique geographic location and large-scale terrain (Yao et al., 2012). The annual mean temperature varies between −2.2 and 0°C, and the annual mean precipitation ranges from over 1,000 mm in the southeast to less than 50 mm in the northwest. Precipitation between June and September contributes to more than 60%-90% of total yearly precipitation (Kuang and Jiao, 2016), and the mean annual potential evapotranspiration and evapotranspiration are approximately 940 mm and 380 mm, respectively Feng et al., 2021). The QTP is the headwaters of many prominent rivers in Asia (Sun et al., 2021) ( Figure 1). Glaciers and permafrost are widely distributed on the plateau, as well as alpine meadow and alpine grassland (Feng et al., 2019).
This study focuses on the meteorological drought in the QTP. In order to fully understand the spatial and temporal characteristics of drought, long-term datasets related to precipitation and potential evapotranspiration are fundamental. However, there are insufficient meteorological stations covering QTP characterized by complex topography and harsh climate, so the grid datasets of potential evapotranspiration (PET), precipitation, and temperature derived from the European Centre for Medium-Range Weather Forecasts (ECMWF) were served as substitutes of station observations (Zhou et al., 2021). ERA5 is the fifth generation of the ECMWF reanalysis for the global climate and weather (https://cds.climate.copernicus.eu/cdsapp#!/ home). The datasets cover the period from 1980 to 2020, and the temporal coverage is 1 month, while the spatial resolution is 0.25°× 0.25°for each pixel. We used the boundary of the QTP to extract the raster of potential evapotranspiration, temperature, and precipitation and then conduct the following postprocessing.

SPEI calculation
SPEI is an ideal indicator to monitor drought, its principle is that the drought in a region can be aggregated by the deviation of the difference between precipitation and PET from the average state (Lu et al., 2019;Hu et al., 2021). The calculation process please refer to Vicente-Serrano et al. (2010) for details. In this study, 12 months scale of SPEI was computed as the SPEI-annual, and the drought classification based on the SPEI value is shown in Supplementary Table S1.

Trend analysis
Sen's slope and Mann-Kendall trend test are employed to examine the trend of drought. Sen. slope estimation is a non-

FIGURE 2
Spatial distribution of trend in annual SPEI. (·means that 95% of the significance test has passed).

Frontiers in Environmental Science
frontiersin.org parametric trend slope calculation method, which is not affected by the singular values of a series and can well reflect the degree of change trend of the time series. Mann-Kendall trend test is a nonparametric test to assess the trend of climate and hydrological factors of a time series, using the statistic Z to check the significance of the changing trend (Gocic and Trajkovic, 2013).

Hurst exponent
The Hurst exponent serves as a practical statistical tool to predict the time series persistence. The rescaled range (R/S) analysis has been used by many studies to calculate the Hurst exponent, and the calculation steps are as follows (Tong et al., 2018): Firstly, the time series of SPEI {X i } (i = 1, 2, 3, . . . , n) is grouped into a number of sub-series and then calculate the average value.
The cumulative deviation is calculated by The range of each sub-series is computed by The standard deviation series is given as Finally, we calculate the rescaled range R/Ŝ The value of H is in the range of 0-1. The value in the range of 0-0.5 means that there is anti-persistence in the time series, 0.5-1 indicates long-term memory, and H = 0.5 is an independent process (García and Requena, 2019).

Intensity analysis
The intensity analysis was initially applied to the analysis of land use and land cover change (Huang et al., 2012), then it was used as an effective method to analyze drought characteristics (Wang et al., 2019a). Intensity analysis can account for the change in interval, category, and conversion level. Interval intensity is used to analyze the change rate and identify whether the change rate is fast or not at a certain time interval, and category intensity can determine the state of category change, namely relatively active or dormant, while transition intensity is to determine which one is in the dominant position in the transition during the time interval (Abdullah and Nakagoshi, 2006;FIGURE 4 The interval intensity analysis for the 1980s, 1990s, 2000s, and 2010s.

FIGURE 6
Intensity analysis of transition category in each interval.
Frontiers in Environmental Science frontiersin.org 05 Aldwaik and Pontius, 2012). The equations related to intensity analysis are shown in Supplementary Table S2.

Empirical Orthogonal Function
EOF is one of the most widely used tools in meteorological research (Dai et al., 2004), and it aims to convey as much information provided by raw data as possible with fewer independent variables. EOF reduces the dimensionality of a large amount of raw data to obtain the dominant spatial pattern and temporal coefficients . The significant test was conducted to determine whether the decomposed eigenvectors are physically meaningful (North et al., 1982). Figure 2 shows the annual spatial change trend of drought distribution characterized by SPEI from 1980 to 2020. About 73% of the areas showed an increasing trend, suggesting that a majority of the areas were becoming wetter, especially in the western and southwest regions. Only 27% of the areas were prone to drying, and it was mainly distributed in the southeast and northeast of the QTP, as well as the Qaidam Basin. The areas with the significant trend all passed the 95% significance test. Figure 3 shows the Hurst exponent from 1980 to 2020. In general, 90% areas of the QTP had the value of Hurst exponent between 0.5 and 1.0, indicating that the current state would have a persistent trend in the future, and was primarily distributed in the west and the south of the QTP ( Figure 3A). In the remaining 10% of the area with the Hurst exponent less than 0.5, the current state would be reversed. Combining the trend analysis of SPEI ( Figure 2) and Hurst exponent, 65% areas of the QTP would continue to be wet in the future. The wetting trend would be distributed in the western regions and some of the eastern QTP. Areas with persistent drying trend accounted for around 24% of the QTP and were primarily found in the southeast and northeast of the QTP, as well as the Qaidam Basin. The remaining 10% of the QTP would be transformed from dry to wet or from wet to dry ( Figure 3B).

Changes in drought intensity characteristics
We divided the drought into nine categories for the 1980s, 1990s, 2000s, and 2010s. The number of grids for each drought category was calculated from one-time interval to the next, as shown in Supplementary Table S3. Figure 4 shows the overall rate of change in the drought during each time interval. The horizontal axis shows the time interval, and the vertical axis displays the percentage variation in the number of pixels per year, with the horizontal solid line indicating the uniform annual change. If the bar does not exceed the uniform annual change line, it means that the interval changes slowly. On the contrary, the interval changed quickly if the bar is above the uniform annual change line. The rate of change was relatively quick for the 1980s and 2000s, and slow for the 1990s and 2010s. The 1980s was the fastest rate of change. Figure 5 shows whether the change rate of the nine drought categories in each interval was an active or dormant state. Each category has two horizontal bars, the lower bar shows the number of loss grids and the annual intensity of change, while the upper bar represents the gain. The solid line is the uniform line, if the bar finishes on the right side of the uniform line, the variation in the interval is relatively active, and if the bar ends on the left side of the uniform line, the variation in the interval is dormant.
In the 1980s interval, all wet categories gained more grids than lose grids. The intensity of both extreme wet and drought exceeded the uniform line in the interval, close to 10%, indicating that they were relatively active. The near normal was negative incomes and relatively dormant, with the largest grid changes of more than 1 k ( Figure 5A). In the 1990s interval, except for near-normal and moderate drought (loss) in a dormant state, the rest were relatively active. Moderate wet and below moderate drought were positive incomes, but the light drought was negative, and the largest number of grid changes was near normal ( Figure 5B). In the 2000s interval, the grids changes of each wet category were the same basically, but light wet and near-normal were negative incomes, and all drought types were positive gains ( Figure 5C). In the last interval, the average intensity was relatively small, all types were above the uniform line and were in an active state, the gross gain of wet was positive, and the gross gain of the moderate drought was large, which could indicate that the number of grids transferred to moderate drought was large, and the rest were the same in general ( Figure 5D). Figure 6 demonstrates the intensity analysis of the transition level. The upper vertical bar is the transition intensity from category i to category m (gain), and the lower vertical bar is the transition intensity from category m to category i (loss, m≠i). The solid red line indicates the uniform transition intensity. If the vertical bar exceeds the uniform line, it shows that the transition intensity of this drought category is dominant.
Combining Supplementary Table S3 with Figure 6, the results indicated that the 1980s were primarily dominated by moderate drought and slight wet transition. The gross grids transformed between drying and wetting were the same, indicating there was no obvious transformation of the drying and wetting area during this interval. But in the 1990s, the drying areas were more than the wetting areas in general, light drought, near normal, and light wet Frontiers in Environmental Science frontiersin.org were dominant, and the number of interconverted grids among near normal, severe and light drought, moderate and light wet were high. In the 2000s, it was obvious that all of them were transformed to the lower level of drought, near-normal and light wet were the dominant transition, and the drying area was increasing. It was dominated by light wet and moderate drought in the 2010s, the number of wetting grids was significantly greater than the drying grids, and the area of wetting was expanding.

Main spatial gathering patterns of drought
The 12-month-scale SPEI was used to perform EOF analysis to obtain the main spatial gathering characteristics of drought in the QTP, and the significance test was performed using the North significance test. The spatial distribution characteristics of the first few eigenvectors having passed the significance test could maximize the distribution gathering characteristics of drought. If each component of the eigenvector has the same sign, the variation of this region shows the same, while if the components of an eigenvector are distributed in a positive and negative pattern, they represent two different types of distributions. The larger absolute value is the center, and the value of the eigenvector multiplies by the value of the time coefficient represents the actual value. As shown in Supplementary Table S4, the cumulative variance contribution rate of the first five eigenvectors accounts for 70.8%. The eigenvectors with a variance contribution rate of less than 10% are not strong enough to explain, while the cumulative

Frontiers in Environmental Science
frontiersin.org variance contribution rate of the first three eigenvectors close to 60%. Therefore, the first three eigenvectors are used to characterize the main spatial gathering characteristics of SPEI in the QTP.
The variance contribution rate of the EOF mode1 was the largest with 23%, which was the principal spatial gathering distribution of drought, and exhibited a gathering distribution structure with negative values in the southeast and east of the QTP and positive values in the center and west ( Figure 7A). The low-value center was in the west, indicating that the variation of dry and wet was more obvious there, and the regional changes were consistent. Combined with the time coefficients, the turning point of dry and wet occurred about 1998, indicating that the west was dry before 1998 and wet afterward. Extreme drought years occurred in 1985, 1995, and 2015, respectively, and extreme wet years occur in 1998, 2000, 2002, 2017, and 2018, respectively. From the trend line of the time coefficients, we could also determine that wetting trend would continue in most areas in the future.
The spatial gathering distribution of the EOF mode2 was shown in Figure 7B. The variance contribution rate reached 20%, close to the first mode, reflecting the eigenvector field of the southwest and northeast inversions with the low-value center in the southeast and the high-value center in the northeast. The time coefficients before 1987 and after 2012 were positive, indicating that the southwest was dry while the northeast was wet. The extreme drought years in the southwest were 1989, 1993, 2009, and 2012, and the extreme wet years were 1980, 1994, 1997, and 2013 respectively.
The EOF mode3 showed an opposite trend between the north and south with a 13% variance contribution rate ( Figure 7C). The high-value centers were mainly located in the east and south, while the low-value centers were mainly found in the west and parts of the north. The 2004 year was a turning point. The south was in a wet situation before that, and it had been in a dry state after that. The years of extreme drought in the south occurred in 1981, 1983, 1986, 1994, 2002, 2006, and 2015, respectively, and the extremely wet ones were in 1985, 1990, 1991, 1998, and 2020, respectively. In order to analyze the causes of the three distinct spatial gathering characteristics, we also investigated the gathering characteristics of precipitation, temperature, and potential evapotranspiration, taking into account all spatial characteristics with contribution rates of more than 5%, as shown in Supplementary Table S5. It indicates that the principal spatial gathering distribution of the EOF mode one might correspond to the altitude and temperature ( Figure 8A); The southwest-northeast mode might be closely connected to potential evapotranspiration, and potential evapotranspiration changes have a greater impact on the trend of the dry and wet change ( Figure 8B); Precipitation and potential evapotranspiration might be related to the south-north mode ( Figure 8C), the superposition of the precipitation decreasing and potential evapotranspiration increasing might be the main reason for the north-south difference.

FIGURE 8
Spatial distribution of altitude(A) and the spatial gathering distribution of the second mode of PET(B), the first mode of P(C).

Frontiers in Environmental Science frontiersin.org
We also analyzed the correlation of precipitation, temperature, and potential evaporation with SPEI, as well as its trend analysis. The result showed that SPEI was strongly positively connected to precipitation but negatively associated with potential evapotranspiration. The positive correlation with the temperature only occurred in a few areas (Figure 9). Wetting regions had a significant increase in precipitation and a significant decrease in potential evapotranspiration, whereas precipitation in dry areas decreased and potential evaporation increased. The reasons of drought in these regions might be associated with the static wind and central plateau convection, which caused the unequal distribution of rainfall, while warming led land evapotranspiration to increase. The superposition caused runoff decrease and the occurrence of drought (Yang et al., 2014;Yang et al., 2011).

Discussion
The QTP was wetter in general, and only 27% of the areas tended to be dry from 1980 to 2020. Several recent studies also concluded that the QTP became wetter in recent years (Li et al., 2010;Chen and Sun, 2015;Gao et al., 2015;Liu et al., 2015;Wang et al., 2015;Liu et al., 2019b). However, some studies pointed out that the southern regions were becoming wetter (Yu et al., 2014;Wang et al., 2017). To understand the differences with other studies, we also compared some studies with different drought indices and data in the QTP, and their findings were listed in Table 1. Furthermore, The PET of SPEI calculated by the Thornthwaite method would exhibit significantly greater differences than SPEI_PM (Chen and Sun, 2015). The unequal period covering the study and

Frontiers in Environmental Science
frontiersin.org usage of the different datasets as well as different methods might affect the results. Apparently, the causes of drought are extremely complex. Human activities, such as land use/cover change (LUCC) and urbanization, change the exchange of water and energy between the atmosphere and the soil by modifying the Earth's surface, further aggravating drought (Meshesha et al., 2012;Yang et al., 2017). The interaction of dust-cloud, atmosphere-land, atmosphere-ocean has a significant impact on regional climate change and hydrological cycle, leading to the variations of precipitation, temperature, PET and soil moisture, which can affect drought (Huang et al., 2017). The drought might also be related to geographical environmental characteristics and elevation (Feng et al., 2020). Therefore, in order to better understand the drought, more detailed investigations in drought mechanisms are needed in future studies.
Though in this study we analyze the drought characteristics in the QTP from 1980 to 2020 using grid datasets and have made some findings, there are still some limitations. Firstly, the dataset is not as accurate as traditional meteorological data and contains some anomalies. Besides, the causes of drought are complicated, we did not accurately distinguish whether natural factors or human activities influenced the drought, e.g., the government's afforestation project, returning farmland to forests, and protection of pastures, which may be responsible for the regional wetness (Li et al., 2013). Furthermore, the causes of the inversion of dry to wet and wet to dry are not addressed in this study. In the future, high-resolution grid data could be used to analyze different drought indices, and to compare their similarities and differences. The influence mechanism and propagation time of different climate variables on drought are also very interesting topics and need to be further explored.

Conclusion
Based on the trend analysis, Hurst exponent analysis, intensity analysis, and EOF analysis methods, we analyzed the dry and wet changes, future persistent changes of drought, drought intensity, and transfer, as well as spatial structure characteristics of drought from 1980 to 2020 in the QTP. The following are the main conclusions: 1) In the past 41 years, there was an overall wetting trend in the QTP, with about 27% of the southeast and some of the northeast of the QTP experiencing regional drying. In the 1980s and 2000s, the drought intensity was relatively fast, while it was slow in the 1990s and 2010s. 2) The future trend of wet and dry in most areas would be the same as the current trend, only 10% of the areas had the reverse transformation trend. 3) EOF mode1 revealed a gathering distribution structure with negative values in the southeast and east of the QTP and positive values in the center and west. The west was more sensitive to climate change, and the wet/dry changeover occurred about in 1998. Most areas will continue to be wet in most areas in the future. A contrast gathering pattern between southwest and northeast was found by analysis of EOF mode2. The southwest was dry while the northeast was wet before 1987 and after 2012, the reverse occurred between 1987 and 2012; EOF mode3 showed the opposite phases between the south and north regions, and the southern regions were prone to be dry with a high amplitude, and the year of transition was in 2004.

Data availability statement
Publicly available datasets were analyzed in this study. This data can be found here: https://cds.climate.copernicus.eu/ cdsapp#!/home.