An Evaluation of the Capability of Global Meteorological Datasets to Capture Drought Events in Xinjiang

Abstract

With the accelerating pace of global warming, the imperative of selecting robust, long-term drought monitoring tools is becoming increasingly pronounced. In this study, we computed the Standardized Precipitation Evapotranspiration Index (SPEI) at both 3-month and 12-month temporal scales, utilizing observational data from 102 stations across Xinjiang and gridded observations spanning China. Our objective encompassed an assessment of the efficacy of three widely employed global meteorological estimation datasets (GMEs) in the context of drought monitoring across Xinjiang over the period of 1960–2020. Moreover, we conducted an in-depth examination into the origins of discrepancies observed within these GMEs. The findings of our analysis revealed a notable discrepancy in performance among the three GMEs, with CRU and ERA5 exhibiting significantly superior performance compared to NCEP-NCAR. Specifically, CRU (CC = 0.78, RMSE = 0.39 in northern Xinjiang) performed excellently in capturing regional wet–dry fluctuations and effectively monitoring the occurrence of droughts in northern Xinjiang. ERA5 (CC = 0.46, RMSE = 0.67 in southern Xinjiang) demonstrates a stronger capability to reflect the drought dynamics in the southern Xinjiang. Furthermore, the adequacy of these datasets in delineating the spatial distribution and severity of major drought events varied across different years of drought occurrence. While CRU and ERA5 displayed relatively accurate simulations of significant drought events in northern Xinjiang, all three GMEs exhibited substantial uncertainty when characterizing drought occurrences in southern Xinjiang. All three GMEs exhibited significant overestimation of the SPEI before 1990, and notable underestimation of this value thereafter, in Xinjiang. Discrepancies in potential evapotranspiration (PET) predominantly drove the disparities observed in CRU and ERA5, whereas both precipitation and PET influenced the discrepancies in NCEP-NCAR. The primary cause of PET differences stemmed from the reanalysis data’s inability to accurately simulate surface wind speed trends. Moreover, while reanalysis data effectively captured temperature, precipitation, and PET trends, numerical errors remained non-negligible. These findings offer crucial insights for dataset selection in drought research and drought risk management and provide foundational support for the refinement and enhancement of global meteorological estimation datasets.

1. Introduction

Drought is among the most severe natural disasters and is distinguished by its higher frequency, longer duration, broader impact range, and greater damage compared to other natural calamities. It exerts significant adverse effects on socio-economic development and ecosystem stability [1,2,3]. In recent years, the frequency and intensity of droughts in many regions worldwide have been projected to increase due to global warming and human activities [4,5]. To prevent and mitigate the losses caused by droughts, it is essential to accurately monitor their characteristics in a timely manner, focusing particularly on the evolution of spatial patterns. Consequently, there is an urgent need to develop more effective drought monitoring tools and more accurate datasets.

Drought is typically caused by a deficiency in moisture and can be categorized into meteorological drought, hydrological drought, and agricultural drought based on its duration, intensity, impact, and recovery rate. Currently, drought indices are among the key tools for quantitatively monitoring and assessing different types of drought [6,7,8,9,10]. The commonly used indices for evaluating dryness and wetness variations in Xinjiang include the Palmer Drought Severity Index (PDSI) [6], the Standardized Precipitation Index (SPI) [9], and the Standardized Precipitation Evapotranspiration Index (SPEI) [11]. The SPEI considers the effects of temperature and precipitation on drought, is highly sensitive to changes in potential evapotranspiration (PET), and is adaptable to various time scales [12], making it an effective indicator for studying and monitoring changes in wetness and dryness under global warming conditions. In recent years, it has been widely applied in the assessment and prediction of drought in Xinjiang [11,12,13,14,15,16,17,18,19,20]. Generally, the variables used to calculate SPEI, such as temperature and precipitation, originate from ground-based station observations. Due to constraints such as terrain, environmental conditions, and maintenance difficulties, the installation and upkeep of meteorological stations pose significant challenges [21]. This results in there being an uneven distribution of observation stations in remote and underdeveloped regions. Additionally, the accuracy of ground-based observations is affected by factors such as wind, evaporation, human operations, and instrument degradation [22]. Therefore, monitoring dynamic drought information based on limited station observations involves considerable uncertainty. In recent years, to address the data deficiencies in regions with sparse observations, various Global Meteorological Estimate Datasets (GMEs) have been produced. These datasets were created through the interpolation of existing data and/or the integration and assimilation of different multi-source data systems, resulting in a high spatial resolution and temporal continuity. Examples of these include the Climatic Research Unit gridded Time Series (CRU TS) from the Climate Research Unit at the University of East Anglia [23], the fifth-generation global climate reanalysis data ERA5 [24] released by the European Centre for Medium-Range Weather Forecasts (ECMWF), and the NCEP/NCAR Reanalysis I [25] jointly released by the National Centers for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR).

Recent studies extensively analyzed the accuracy of the three aforementioned GMEs in assessing near-surface meteorological elements [21,26,27,28,29,30,31]. For instance, CRU was demonstrated to be an effective substitute for hydrological observations in semi-arid regions [32,33], despite CRU often underestimating precipitation and PET values [34,35]. Wang et al. [36] found that CRU underestimated the decadal variability in precipitation in China and exhibited significant discrepancies from observed precipitation in areas near large mountain ranges in western China. Moreover, CRU performed well in estimating the temporal and spatial variability in drought in Africa [37]. For reanalysis products, variations in input data, numerical assimilation models, parameterization schemes, and the spatiotemporal resolution of the final products led to significant differences in their performance across different regions. NCEP-NCAR successfully captured the spatial pattern of the East Asian Summer Monsoon (EASM), though it generally overestimated precipitation and temperature in China [38,39,40]. While NCEP-NCAR accurately captured the significant warming trend in eastern China, it failed to account for the summer temperature rise in high-altitude regions, displaying considerable deviations in warming magnitude in western China [41]. Additionally, NCEP-NCAR performed well in simulating the decadal variation in surface wind speed in the northwest region, but overestimated relative humidity in all seasons, particularly in spring and winter [42]. Recent studies validated that ERA5 is suitable for hydrological and climate research in China [43,44,45,46,47], but there is still significant uncertainty regarding precipitation in western China. For instance, while ERA5 precipitation captured the spatiotemporal characteristics of observed regional precipitation events, it tended to overestimate precipitation frequency and underestimate precipitation intensity, failing to accurately simulate short-term (hourly and daily) precipitation [48,49]. Jiao et al. [44] validated that ERA5 overestimated precipitation on both seasonal and annual scales and found that it could not accurately simulate the spatial distribution of trend changes. Furthermore, ERA5 faced challenges in simulating extreme temperature and precipitation events in western China [50,51].

Numerous studies compared the accuracy of various GMEs, including NCEP-NCAR, CRU, and ERA, for different meteorological variables [26,32,38,39,41,52,53,54,55,56,57,58]. However, due to updates in GMEs (such as ERA-Interim to ERA5), more timely evaluations of these products are necessary. Recently, numerous related studies conducted worldwide explored the application of high-resolution products in drought detection, though these studies generally focused on satellite-based datasets. Few studies [59,60,61] evaluated interpolated data and/or reanalysis datasets, and in assessing the application of high-resolution products in drought monitoring, these studies typically considered precipitation as the sole factor, neglecting the influence of evapotranspiration. To our knowledge, very few evaluations have addressed the suitability of different GMEs for drought monitoring based on SPEI. Currently, only a limited number of studies have been conducted in China and Pakistan [19,62,63].

Xinjiang, as an important part of the arid region in Central Asia, faces extreme water scarcity and is highly sensitive to global climate change [11]. The region’s water resources are diverse, with widespread mountain glaciers and significant variations in the proportion of total runoff contributed by snow and glacier meltwater across different basins [64]. With the intensification of global warming, glaciers in the mountainous areas are rapidly retreating, which is profoundly impacting the region’s water resources, ecological environment, and socio-economic development [14]. Accurately assessing the wet–dry changes in Xinjiang and effectively monitoring drought in the region remain urgent scientific challenges that need to be addressed. CRU TS4, ERA5, and NCEP-NCAR Reanalysis 1 (hereafter referred to as CRU, ERA5, and NCEP-NCAR) are among the limited products with long-term records available for Xinjiang, holding significant potential for the long-term monitoring of drought in the region. The drought simulation capability of the SPEI, determined based on CRU and ERA5 data, has been effectively validated in Pakistan, India, and the Arabian Peninsula [19,65,66]. However, there remains a research gap in Xinjiang. Given the uncertainty inherent in gridded meteorological datasets, GMEs exhibit different qualities in representing various aspects of drought occurrence, such as size, frequency, intensity, seasonality, and geographical distribution. Therefore, evaluating their accuracy and applicability is a prerequisite for their effective use.

Based on the aforementioned motivations, this study conducted a comparative analysis of the potential of GMEs (CRU, ERA5, and NCEP-NCAR) to estimate the SPEI for the Xinjiang region. The aim of this research was to comprehensively evaluate the drought monitoring capabilities of these GMEs to aid in selecting appropriate datasets for future studies on drought events in Xinjiang. Our evaluation results will identify the issues associated with different GMEs in a region as complex in topography and climate systems as Xinjiang, providing valuable feedback for the updating and improvement of GMEs.

2. Data and Methodology

2.1. Study Region

As a typical “arid and semi-arid region”, Xinjiang is characterized by a fragile ecological environment and sensitivity to climate change, making it a key area for global climate change research. Xinjiang is one of China’s major agricultural production bases, with a long agricultural production cycle and abundant natural resources. The region is also rich in mineral resources, including petroleum, natural gas, and coal. Climate change, including extreme weather events, rising temperatures, changing precipitation patterns, and drought and flood disasters, is having significant impacts on agricultural production and the energy industry. Therefore, the accurate assessment of climate change in Xinjiang is of paramount importance. Xinjiang is surrounded by high mountains, with the Tianshan Mountains running through its center, creating a unique topography of “three mountains and two basins.” Based on the climate zoning map of China provided by Zheng et al. [67], we divided Xinjiang into two parts, using the Tianshan Mountains as the boundary [68]: northern Xinjiang (I) and southern Xinjiang (II) (Figure 1). Northern Xinjiang belongs to the mid-temperate zone, with an average annual temperature of 5.9 °C and an average annual precipitation of 253.0 mm. This region primarily includes the northern slopes of the Tianshan Mountains, the Ili River Valley, the Junggar Basin, and areas to the north. Southern Xinjiang belongs to the warm temperate zone, with an average annual temperature of 10.5 °C and an average annual precipitation of 68.5 mm. This region mainly includes the southern slopes of the Tianshan Mountains, the Turpan Basin, the Tarim Basin, and areas to the south.

2.2. Data Description

This study evaluates the performance of three GMEs in capturing the spatial and temporal patterns of drought in Xinjiang. These are among the few products with long-term records and a rich set of variables available for the region. The data used in this study mainly included the following: (1) reference data—a quality-controlled station recorded from 102 sites and we used the 0.25° China CN05.1 gridded observational data; (2) station-interpolated data: this included the high-resolution gridded dataset from CRU; (3) reanalysis data: this included the ERA5 reanalysis dataset and the NCEP-NCAR Reanalysis 1 dataset. These products varied in terms of spatial resolution, the quantity and type of data inputs used for data construction (based on observations and reanalysis), and the methods or algorithms used for data construction. The reasons for selecting these datasets were their long time spans, the variety of variables available (used to generate the monthly scale SPEI), and their wide application and accessibility in both research and operational activities. Based on the sequence length of station data, the station-interpolated and reanalysis data used in this study covered the period from 1960 to 2020. For ease of comparison, all three data products were resampled to a 0.25° spatial resolution, resulting in 2783 grids in Xinjiang. Detailed information about each dataset can be obtained from the respective data sources, as listed in

Table 1. Description of reference and global grid datasets used in this study.

Dataset	Record Length	Temporal Resolution	Spatial Resolution	Data Category	Variable	Source of Data
In situ stations	1961–2020	Daily	Point	Station (102)	temperature, precipitation, surface pressure, wind speed, relative humidity, sunshine duration	http://data.cma.cn/ (accessed on 15 January 2024)
CN05.1	1961–present	Daily	0.25°	Gauge-interpolated	temperature, precipitation, surface pressure, wind speed, relative humidity, sunshine duration	http://www.geophy.cn//article/doi/10.6038/cjg20130406 (accessed on 15 January 2024)
CRU	1901–present	Daily	0.5°	Gauge-interpolated	temperature, precipitation, PET	http://www.cru.uea.ac.uk/data/ (accessed on 15 January 2024)
ERA5	1940–present	Hourly	0.25°	Reanalysis	temperature, precipitation, wind speed, radiation, dew point temperature	https://cds.climate.copernicus.eu/cdsapp#!/dataset/ (accessed on 15 January 2024)
NCEP-NCAR	1948–present	Daily	2.5°	Reanalysis	temperature, precipitation, wind speed, radiation, relative humidity, surface pressure	https://psl.noaa.gov/data/gridded/data.ncep.reanalysis.html (accessed on 15 January 2024)

.

2.2.1. Reference Data

The station observation data were sourced from the China National Ground Meteorological Station Daily Meteorological Elements Dataset (V3.0), which includes daily observations from 102 stations in Xinjiang for the period from 1960 to 2020 (Figure 1). The observed elements included the maximum temperature, the minimum temperature, the average temperature, precipitation, the average pressure, the average wind speed, relative humidity, and sunshine duration. The availability rate of each element’s data is generally above 97%, and the accuracy of the data is nearly 100%. For stations with missing data (primarily Turpan Dongkan and Wensu), the missing values were imputed using a stepwise regression analysis method. The specific details are as follows: data from nearby stations within a 150 km radius of the target site (missing data site) were selected. Working based on the complete daily data from 1960 to 2020 for both the target site and the nearby stations, a stepwise regression model was established. If data were still missing, the missing values were interpolated using the long-term average value of the target site on days without missing data from 1960 to 2020. Taking the interpolation of precipitation on 13 January 1960, at Jimunai, we identified stations within 150 km that had no missing data for that day (Stations 1 to 4: Habahe, Burqin, Fuhai, and Hoboksar). The precipitation values at these four stations (×1 to ×4) were 3.6 mm, 2.7 mm, 2.9 mm, and 2.3 mm, respectively. A stepwise regression model was established with the precipitation at the Urumqi Agrometeorological Experiment Station (y) and the precipitation values at the other four stations (×1 to ×4), resulting in the equation of y = 0.254 × 1 + 0.355 × 2 + 0.199 × 3 + 0.112 × 4 + 0.191, where p < 0.05. This yielded an interpolated precipitation value of 2.9 mm for that day. The daily meteorological observation dataset for the 102 stations across Xinjiang from 1960 to 2020, after interpolation, was used to evaluate the ability of GMEs to monitor drought.

The CN05.1 dataset was created by Wu Jia and Gao Xuejie using data from over 2400 national meteorological stations managed by the National Meteorological Information Center [69]. They employed the thin-plate spline function method (ANUSPLIN) and the angular distance weighting method (ADW) to interpolate the data, resulting in a spatial resolution of 0.25° × 0.25° for daily observational data. The observed elements included maximum temperature, minimum temperature, average temperature, precipitation, average pressure, average wind speed, relative humidity, and sunshine duration across China. Due to its dense source station network and strict quality control, CN05.1 was frequently used as a reference dataset for evaluations in China [70,71,72,73,74].

2.2.2. Global Meteorological Estimate Datasets (GMEs)

CRU TS v.4.06 was produced by the Climatic Research Unit (CRU) at the University of East Anglia, covering various variables such as cloud cover, precipitation, temperature, vapor pressure, PET, and frost day frequency on a monthly timescale from January 1901 to December 2021. This dataset was created by normalizing station series to the 1961–1990 period using observed data and then gridding these anomalies to a 0.5° grid using the angular distance weighting (ADW) method. The resulting anomaly grids were converted into actual values using CRU CL v1.0 climatology data [23]. The dataset utilized daily or sub-daily observations from national meteorological services and other institutions, providing global coverage at a 0.5° × 0.5° longitude grid, excluding Antarctica. We selected the temperature, precipitation and PET data from the CRU dataset. The PET data were calculated using the Penman–Monteith formula [55], which was consistent with the method used to calculate PET from station data in this study.

ERA5, the latest global reanalysis dataset produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), was created using a 4D-Var data assimilation system [24]. This system integrates a modern numerical weather prediction model with observational data from various platforms (in situ, radiosonde, and satellite), providing detailed records of the global atmosphere, land surface, and ocean waves from 1950 onwards. Compared to its predecessor, ERA-Interim [75], ERA5 features higher spatial and temporal resolutions, an improved 4D-Var data assimilation scheme, an updated forecast model, and a larger set of output parameters. ERA5 also offers enhanced surface parameterization and cloud precipitation models and incorporates more historical observational data into the assimilation system. The number of assimilated observations increased from an average of approximately 750,000 per day in 1979 to about 24 million per day in January 2019. Over the 40-year period from 1979 to 2019, 946 billion observations were assimilated by the 4D-Var system. We extracted daily values of temperature, precipitation, wind speed, radiation, and dew point temperature from ERA5. The spatial resolution was 0.25°, and the temporal resolution was 1 h.

The NCEP-NCAR Reanalysis 1 dataset is a comprehensive global atmospheric and ocean reanalysis dataset providing daily global reanalysis meteorological elements from 1948 to the present [25]. Developed collaboratively by the National Centers for Environmental Prediction (NCEP) and the National Center for Atmospheric Research (NCAR), this dataset was created through the rigorous quality control of observational data from 1948 onwards, followed by three-dimensional variational assimilation. The assimilation process incorporates observational data from various sources, including surface stations, ships, radiosondes, balloons, and satellites. The reanalysis dataset includes a wide range of atmospheric and oceanic variables. We extracted daily values of temperature, precipitation, wind speed, radiation, relative humidity, and pressure from the NCEP-NCAR Reanalysis 1 data. The horizontal resolution was 2.5° × 2.5°.

To match the 0.25° reference data (CN05.1), all GME data were interpolated to a 0.25° grid using bilinear interpolation, resulting in a total of 2783 grid points in Xinjiang.

2.3. Method Description

2.3.1. Standardized Precipitation Evapotranspiration Index (SPEI)

Commonly used indices for evaluating dryness and wetness variations in Xinjiang include the Palmer Drought Severity Index (PDSI) [6], the Standardized Precipitation Index (SPI) [9], and the Standardized Precipitation Evapotranspiration Index (SPEI) [11]. The SPEI, which integrates the effects of temperature and precipitation on drought and accommodates multiple timescales [18], has become an effective indicator for studying and monitoring dryness and wetness variations during global warming. It has been widely applied in the assessment and prediction of drought in Xinjiang in recent years [11,12,13,14,15,16,17].

In this study, we use the Standardized Precipitation Evapotranspiration Index (SPEI) with a three-month time step (hereafter referred to as SPEI) to analyze the seasonal variation in dryness and wetness characteristics in Xinjiang. The SPEI value for December at a 12-month timescale (SPEI-12) is used to characterize the drought conditions for the given year. The SPEI is obtained by calculating the difference between precipitation and potential evapotranspiration, followed by normalization using a normal distribution. It primarily reflects the extent to which climate conditions deviate from normal conditions in a given year. The potential evapotranspiration is calculated using the Penman–Monteith equation [55], which is adopted as Equation (1).

$$\mathit{PET}={\displaystyle \frac{0.408∆(R_n-G+\gamma \frac{900}{T+273}{u}_2(e_s-e_a)}{∆+\gamma (1+0.34u_2)}}$$

(1)

where PET is potential evapotranspiration (mm/day), R_n is the net radiation at the crop surface [MJ/(m²·day)], G is the soil heat flux [MJ/(m²·day)], $\gamma$ is the psychrometric constant (kPa/°C), T is the average daily air temperature (°C), u₂ is the wind speed at a height of 2 m (m/s), e_s is the actual vapor pressure (kPa), e_a is the actual vapor pressure (kPa), and (e_s − e_a) is the vapor pressure deficit (kPa). The SPEI was obtained by calculating the difference between the cumulative precipitation and PET over three months. The interpolated values were fitted into a log-logistic probability distribution, and then we normalized the cumulative probability density to a standard normal distribution [76]. As shown in

Table 2. Classification of SPEI values into different wet and dry magnitude conditions.

SPEI Value	Classification
>1.5	Severely wet
1 to 1.49	Moderately wet
−0.99 to 0.99	Near normal
−1.99 to −1	Moderate drought
<−1.5	Severe drought

, wet and dry conditions were categorized into classes based on their different strengths.

2.3.2. Evaluation Metrics

Six metrics were selected to evaluate the performance of the three GMEs [22,49]. Methods were categorized into two types. ① The first was continuous statistical metrics: absolute error (AE), the Pearson correlation coefficient (CC), and the root-mean-square error (RMSE) were used to quantitatively assess the accuracy of the dataset. AE was used to indicate the systematic bias of the dataset, with negative and positive values representing underestimation and overestimation, respectively. CC represented the linear correlation between the dataset and the meteorological station observations. RMSE measured the average error of the dataset. ② The second was categorical statistical metrics: the probability of detection (POD), false alarm rate (FAR), and equitable threat score (ETS) were used to evaluate the dataset’s ability to detect drought events. POD measured the probability of correct detection, while FAR measured the probability of false detection. ETS provided a more comprehensive assessment of the dataset’s performance by eliminating the contribution of random hits. Typically, a drought event threshold was set at −1, and a severe drought event threshold was set at −1.5. The formulas for these metrics are as follows:

$$AE={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n(y_i-x_i)$$

(2)

$$CC={\displaystyle \frac{\sum _{i=1}^n(x_i-\overline{x})(y_i-\overline{y})}{\sqrt{\sum _{i=1}^n{(x_i-\overline{x})}^2}\sqrt{\sum _{i=1}^n{(y_i-\overline{y})}^2}}}$$

(3)

$$RMSE=\sqrt{{\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(x_i-y_i)}^2}$$

(4)

$$POD={\displaystyle \frac{H}{H+M}}$$

(5)

$$FAR={\displaystyle \frac{F}{H+F}}$$

(6)

$$ETS={\displaystyle \frac{H-H_e}{H+M+F-H_e}}$$

(7)

$$H_e={\displaystyle \frac{(H+M)(H+F)}{H+M+F+C}}$$

(8)

In these formulas, H is the number of events observed by both the reference data and GMEs, M is the number of events observed by the reference data but not by the GMEs, F is the number of events observed by the GMEs but not by the reference data, and C is the number of non-events observed by both the reference data and GMEs. H_e represents the number of events correctly detected by the evaluated data due to random chance. POD measures the proportion of correctly detected events, equivalent to the hit rate of the evaluated data. A POD value closer to 1 indicates that the GMEs detect more actual precipitation events. FAR measures the proportion of events detected by the GMEs that do not actually occur and is equivalent to the false detection rate. The range is from 0 to 1, with a value closer to 0 indicating that fewer false events are detected by the GMEs. ETS provides the proportion of correctly detected events, considering the probability of random hits. The range is from −1/3 to 1, with a higher ETS value indicating stronger detection capability of the GMEs. In this study, thresholds were set to calculate the categorical statistical metrics for drought occurrence (POD-d, FAR-d, and ETS-d) and severe drought occurrence (POD-s, FAR-s, and ETS-s), respectively.

2.3.3. Visual Analysis

Visual analysis was employed as a quasi-subjective method to compare the ability of data products to capture the magnitude and spatial patterns of significant droughts [61]. This analysis focused on six typical drought years (1962, 1965, 1974, 1977, 1980, and 1997) and compared the results against the reference data.

2.3.4. Attribute Contributions of Precipitation and PET

As the SPEI was derived by normalizing the difference between precipitation and PET, we analyzed the changes in these differences between GMEs and observational data before and after specific transitions. This approach allowed us to quantify the contributions of precipitation and PET to the overestimation or underestimation of SPEI by GMEs. Specifically, we calculated the differences in precipitation and PET changes during periods of SPEI overestimation and underestimation. For instance, we examined the transition in GMEs from the overestimation to the underestimation (or vice versa) of SPEI around 1990. The change in precipitation differences for the period 1991–2020 relative to 1961–1990 was assessed as follows:

$$P_d=\left(P_{G-a}-P_{O-a}\right)-\left(P_{G-b}-P_{O-b}\right)$$

(9)

$P_{G-a}$ denotes the precipitation from GMEs during the post-transition period (1991–2020), while $P_{O-a}$ represents the precipitation observed during the same period. $P_{G-b}$ represents the precipitation from GMEs during the pre-transition period (1961–1990), and $P_{O-b}$ represents the precipitation observed during the same period (1961–1990). When $P_d$ < 0, it indicates that the GMEs underestimated the precipitation during the post-transition period. Similarly, the change in the difference for PET is given as follows:

$${\mathit{PET}}_d=\left({\mathit{PET}}_{G-a}-{\mathit{PET}}_{O-a}\right)-\left({\mathit{PET}}_{G-b}-{\mathit{PET}}_{O-b}\right)$$

(10)

${\mathit{PET}}_{G-a}$ represents the PET from GMEs during the post-transition period (1991–2020), ${\mathit{PET}}_{O-a}$ represents the PET observed during the same period, ${\mathit{PET}}_{G-b}$ represents the PET from GMEs during the pre-transition period (1961–1990), and ${\mathrm{PET}}_{\mathrm{O}-\mathrm{b}}$ represents the PET observed during the same period. The same rule applies to SPEI (${SPEI}_d$), temperature (${TAV}_d$), relative humidity (${RHU}_d$), and wind speed (${WIN}_d$). When ${\mathit{PET}}_d$ > 0, it indicates that the GMEs overestimated the PET during the post-transition period. The contribution of precipitation and PET to SPEI overestimation/underestimation, respectively, is as follows:

$$\mathit{Rcr}\text{-}p={\displaystyle \frac{P_d}{/P_d/+/{\mathit{PET}}_d/}}\times 100\%$$

(11)

$$\mathit{Rcr}\text{-}\mathit{pet}={\displaystyle \frac{{\mathit{PET}}_d}{/P_d/+/{\mathit{PET}}_d/}}\times 100\%$$

(12)

$\mathit{Rcr}\text{-}p$ and $\mathit{Rcr}\text{-}\mathit{pet}$ represent the contributions of GMEs to the overestimation or underestimation of SPEI relative to observational data. The absolute value of Rcr indicates relative importance, while the sign denotes the direction of the bias. Specifically, $\mathit{Rcr}\text{-}p$ < 0 or $\mathit{Rcr}\text{-}\mathit{pet}$ > 0 indicates a contribution to the underestimation of SPEI, whereas $\mathit{Rcr}\text{-}p$ > 0 or $\mathit{Rcr}\text{-}\mathit{pet}$ < 0 indicates a contribution to the overestimation of SPEI.

3. Results

3.1. Comparison of the Ability of GMEs to Represent the Spatiotemporal Characteristics of Drought

From 1961 to 2020, the annual average SPEI for Xinjiang was 0.012, with southern Xinjiang (0.013) being slightly more humid than northern Xinjiang (0.008). The GMEs underestimated SPEI from 1990 to 2010 but overestimated it during most other periods (Figure A1). Consequently, we considered the decadal variation in absolute error (AE) between the datasets and the reference SPEI in our analysis (Figure 2). The GMEs consistently showed significant overestimation of the SPEI for the entire Xinjiang region from 1961 to 1980, with absolute error (AE) values of 0.396, 0.279, and 0.364 for CRU, ERA5, and NCEP-NCAR, respectively. Conversely, they exhibited significant underestimation from 1991 to 2010, with corresponding AE values of −0.537, −0.373, and −0.476. Both CRU and ERA5 demonstrated better performances during the periods 1981–1990 and 2011–2020, with AE values of 0.069 and 0.001 for 1981–1990, and 0.11 and 0.14 for 2011–2020, respectively. NCEP-NCAR, however, showed underestimation and overestimation during 1981–1990 and 2011–2020, with AE values of −0.120 and 0.433, respectively. The spatial distribution of SPEI bias over the past 60 years indicated that regions with larger errors were mainly concentrated in the basins of southern and northern Xinjiang, while the Tianshan Mountains had smaller errors (Figure 2). Overall, the GMEs exhibited better performances in representing SPEI for northern Xinjiang compared to southern Xinjiang, with ERA5 showing the smallest absolute error across most periods. When analyzing the other two continuous statistical indicators, the GMEs demonstrated higher correlations and smaller errors in northern Xinjiang, particularly in the Tianshan Mountain region, while they performed less effectively in the basin areas of southern Xinjiang. Compared to NCEP-NCAR (CC = 0.27, RMSE = 0.76), CRU and ERA5 had higher correlations (0.48 and 0.52, respectively) and lower errors (0.62 and 0.61, respectively). However, the suitability of these datasets varied by region: CRU performed better in northern Xinjiang, while ERA5 showed stronger representation capabilities in southern Xinjiang (Figure 3). Specifically, in northern Xinjiang, CRU had a higher correlation coefficient (CC) of 0.78 compared to ERA5’s 0.68, whereas in southern Xinjiang, ERA5 had a higher CC of 0.46 compared to CRU’s 0.36. Additionally, CRU exhibited a lower root-mean-square error (RMSE) of 0.39 in northern Xinjiang compared to ERA5’s value of 0.48, while ERA5 had a lower RMSE of 0.67 in southern Xinjiang compared to CRU’s value of 0.72.

3.2. Comparison of GMEs in Detecting Drought Events

Figure 4 presents the differences in three categorical statistical indices—probability of detection (POD), false alarm ratio (FAR), and equitable threat score (ETS)—for detecting drought (Figure 4(a1–c3)) and severe drought (Figure 4(d1–f3) from 1961 to 2020 across four datasets. The results indicated that the three GMEs showed similar spatial distributions of these indices, with northern Xinjiang performing significantly better than southern Xinjiang. Both CRU and ERA5 demonstrated stronger drought detection capabilities, whereas NCEP-NCAR consistently exhibited the weakest detection ability across all categorical indices. In terms of the spatial distribution of the probability of detection (POD), the high-value areas for CRU and ERA5 were primarily located in the western part of northern Xinjiang and the eastern periphery of Xinjiang, while the low-value areas were mainly in the basin regions of southern Xinjiang (Figure 4(a1–f1)). CRU displayed superior detection capabilities in northern Xinjiang, with POD values of 0.61 and 0.48 for detecting drought (POD-d) and severe drought (POD-s), respectively, compared to values of only 0.39 and 0.24 in southern Xinjiang. ERA5 exhibited relatively consistent spatial variability in terms of probability of detection (POD) across Xinjiang, demonstrating a certain level of accuracy throughout the region. In northern Xinjiang, the POD values for drought events (POD-d) and short-term drought events (POD-s) were 0.52 and 0.39, respectively, while in southern Xinjiang the values were 0.40 and 0.25, respectively. Compared to CRU, ERA5 showed better drought detection capabilities in the basin areas of southern Xinjiang. The false alarm ratio (FAR) analysis indicated a higher rate of false alarms for all three datasets across Xinjiang, particularly in the basin areas of southern Xinjiang (Figure 4(a2–f2)). CRU had the lowest false alarm rates in northern Xinjiang, with FAR-d and FAR-s values of 0.41 and 0.53, respectively. In contrast, ERA5 had the lowest false alarm rates in southern Xinjiang, with FAR-d and FAR-s values of 0.59 and 0.76, respectively. The equitable threat score (ETS) results indicated that CRU had better representation capabilities in northern Xinjiang, with ETS-d and ETS-s values of 0.36 and 0.29, respectively. In southern Xinjiang, CRU and ERA5 had similar detection capabilities, with ETS-d values of 0.15 and 0.17, and ETS-s values of 0.10 and 0.11, respectively. CRU performed better in eastern southern Xinjiang, while ERA5 excelled in the western and central basin areas of southern Xinjiang (Figure 4(a3–f3)).

3.3. Representation of Drought Characteristics by GMEs in Xinjiang During Typical Drought Years

To assess the skill of the studied data products in capturing spatial patterns and severity of droughts, we analyzed six major drought events in Xinjiang (1962, 1965, 1974, 1977, 1980, and 1997). The SPEI-12 values for December were used to generate maps for each dataset and drought year, facilitating visual comparisons (Figure 5). As depicted in Figure 5(a1–a6), more than 30% of Xinjiang’s area experienced drought during these years, with variations in spatial distribution and severity. For instance, the drought in 1965 exhibited the highest severity, affecting over half of Xinjiang’s area with mild and severe droughts; 54.6% experienced mild drought and 23.3% experienced severe drought, while other regions remained unaffected by drought (Figure 5(a2)). The drought events in 1962, 1974, 1977, and 1997 covered the majority of northern Xinjiang, accounting for 89.4%, 82.1%, 81.6%, and 84.5% of the area, respectively (Figure 5(a1,a3,a4,a6)). Specifically, the severe droughts of 1962 and 1974 were concentrated in most areas north of the Tianshan Mountains, while in 1997, severe drought affected most of the western region of northern Xinjiang. During these years, most areas of southern Xinjiang experienced near-normal conditions, except for 1977, when drought events affected the northwest region (23.0%). In the other three years, minor drought and wet events were observed in various parts of southern Xinjiang. Unlike the other five major drought years, the drought in 1980 was concentrated in the basin areas of southern Xinjiang, with nearly all areas of northern Xinjiang experiencing near-normal conditions (Figure 5(a5)).

The representation of these six major drought events by GMEs was inconsistent. Despite differences among the data products, all GMEs captured the drought in northern Xinjiang in 1962 (Figure 5(a1–d1)). CRU effectively captured the wet event at the southern boundary in 1962, while ERA5 and NCEP-NCAR did not accurately reflect the “near-normal” conditions in the western part of southern Xinjiang. Notably, all data products accurately represented the absence of drought in the basin areas of southern Xinjiang. None of the GMEs accurately simulated the drought in 1965. Only ERA5 depicted the widespread occurrence of drought in Xinjiang, albeit with a tendency to overestimate its severity in a significant portion of the region compared to the reference data (Figure 5(a2–d2)). In contrast, both CRU and NCEP-NCAR underestimated the extent of the drought in Xinjiang, with NCEP-NCAR even erroneously indicating widespread wet conditions. CRU and ERA5 captured the drought events in northern Xinjiang in 1974, but both underestimated its severity (Figure 5(a3–d3)). Additionally, all three GMEs amplified the severity and geographical continuity of wet conditions in western and southern Xinjiang. The GMEs performed poorly in simulating the droughts of 1977 and 1980, underestimating the severity of drought-affected areas. In 1977, they erroneously indicated wet conditions in some parts of southern Xinjiang. However, CRU and ERA5 accurately reflected the absence of drought in northern Xinjiang in 1980 (Figure 5(a4–d4,a5–d5)). All GMEs effectively captured the drought in northern Xinjiang in 1997. However, none of the three products accurately represented the “near-normal” conditions in most parts of southern Xinjiang in the same year. Both CRU and NCEP-NCAR significantly overestimated the geographical continuity of the drought and incorrectly depicted severe drought in most parts of southern Xinjiang (Figure 5(a6–d6)).

Overall, CRU and ERA5 exhibited similar performances in capturing the six major drought events in Xinjiang, demonstrating notably superior accuracy compared to the NCEP-NCAR. Both CRU and ERA5 accurately simulated drought conditions in northern Xinjiang but showed greater uncertainty in southern Xinjiang. Specifically, they effectively represented the absence of drought in northern Xinjiang in 1980. In the five major drought events in northern Xinjiang, CRU and ERA5 accurately simulated the droughts in 1962, 1974, and 1997, while underestimating the severity of the drought in 1965 and overestimating it in 1977. In southern Xinjiang, except for accurately representing the absence of drought in 1962, CRU and ERA5 failed to accurately reproduce the drought and near-normal conditions in other drought years compared to the reference data.

3.4. Analysis of Error Sources in Drought Characteristics Based on GMEs

Figure 6 illustrates the variations in meteorological elements (precipitation, potential evapotranspiration, temperature, wind speed, and relative humidity) used to compute the SPEI. For each meteorological element, there was a consistent pattern of overestimation or underestimation by the GMEs throughout most of the study period (Figure 6). Moreover, the three GMEs exhibited varying levels of accuracy in reproducing different meteorological elements. Specifically, CRU demonstrated the highest accuracy in simulating precipitation and potential evapotranspiration, while the accuracy of mean temperature varied depending on the location. CRU accurately simulated precipitation in Xinjiang, albeit with a slight underestimation across almost the entire region (relative bias, RB = −26.7%). Conversely, ERA5 (RB = 42.1%), and NCEP-NCAR (RB = 176.5%) generally overestimated precipitation in Xinjiang. ERA5 significantly overestimated precipitation in the Tianshan region, with deviations exceeding 400 mm in some areas. NCEP-NCAR underestimated precipitation in the Tianshan region while significantly overestimating it north of the Tianshan and in southern Xinjiang (Figure A2(a1–d1)). All three GMEs underestimated PET in Xinjiang (Figure 6c). Specifically, CRU slightly overestimated PET in the Tianshan region (average of 63.3 mm), while underestimating PET in other areas (by −133.1 mm), with eastern Xinjiang experiencing more significant underestimation (up to −700 mm). In contrast, ERA5 overestimated PET in the Xinjiang basin (by 324.6 mm) and underestimated PET in mountainous regions (by −255.9 mm). NCEP-NCAR underestimated PET across almost the entire Xinjiang region (by −233.1 mm) (Figure A2(a2–d2)). For temperature, ERA5 performed relatively well in northern Xinjiang, while CRU exhibited a better performance in southern Xinjiang. However, both datasets overestimated temperatures in Xinjiang, displaying a consistent spatial distribution in terms of temperature deviations. These deviations were characterized by underestimation in mountainous regions and overestimation in basins, with the center of overestimation located in the basin of northern Xinjiang and the western regions of southern Xinjiang. In contrast, NCEP-NCAR (AE = −1.76 °C) severely underestimated temperatures across nearly the entire Xinjiang region (Figure A2(a3–d3)). Due to the lack of wind speed and relative humidity data from CRU, its optimal performance could not be assessed. Based on the available data, it was found that both ERA5 and NCEP-NCAR significantly underestimated wind speeds in Xinjiang. ERA5 provided relatively high-quality relative humidity data, whereas NCEP-NCAR tended to overestimate relative humidity in the region (Figure 6d,e).

The reasons for biases in GMEs when simulating the SPEI remained unclear. This was particularly true for the consistent shift from overestimation to underestimation observed around 1990. Since SPEI is derived from the difference between precipitation and PET followed by normalization, it was crucial to analyze the changes in discrepancies between GMEs and observational data before and after this transition to determine the causes of SPEI underestimation or overestimation. For PET, when the SPEI discrepancy in the latter 30 years was greater than that in the earlier 30 years (${\mathit{PET}}_d$ > 0), GMEs tended to underestimate SPEI in the later period. Similarly, a reduction in precipitation discrepancy in the latter 30 years ($P_d$ < 0) also led to an underestimation of SPEI by GMEs during the later period. Figure 7 revealed the consistent spatial distribution of ${SPEI}_d$ across the three GMEs (Figure 7(a1–c1)). The GMEs generally underestimated the SPEI for the latter 30 years in most regions, with the most severe underestimation concentrated in the southern Xinjiang basin. Among the methods, CRU exhibited the most significant underestimation, with an overall ${SPEI}_d$ of −0.61 across Xinjiang, while NCEP-NCAR performed relatively better, with an ${SPEI}_d$ of −0.35. To elucidate the contributions of precipitation and PET to SPEI changes on a spatial scale, Figure 7(a2–c3) presented the spatial distributions of $\mathit{Rcr}\text{-}p$ and $\mathit{Rcr}\text{-}\mathit{pet}$. The absolute value of $\mathit{Rcr}$ indicated relative importance, with the sign denoting the direction of underestimation or overestimation. For CRU, the underestimation or overestimation of SPEI across Xinjiang was predominantly driven by changes in PET. This trend of SPEI underestimation persisted for the most recent 30 years across all regions, except for the eastern and southern borders of Xinjiang. ERA5 exhibited a similar pattern to CRU, but precipitation primarily drove SPEI changes in the Tianshan region. In contrast, NCEP-NCAR displayed a more complex situation where areas dominated by either precipitation or PET lacked spatial continuity, with most regions being jointly influenced by both factors.

Since SPEI values primarily reflected relatively wet or dry periods within the study timeframe, the accurate simulation of overall trends in meteorological elements by the GMEs masked some inherent systematic errors. This study identified three representative grid sites (Figure A3) based on the magnitude of SPEI errors (large/small). Site A, located in the Tianshan mountainous region, had small SPEI errors, while sites B and C, located in the northern and southern basins of Xinjiang, respectively, exhibited larger SPEI errors. To reduce randomness, we also extracted data from eight additional surrounding sites for each representative site and used the average values from these nine locations to analyze the changes in local meteorological elements. For the Tianshan mountainous region (Station A), the three GMEs demonstrated relatively good performances in representing SPEI. However, ERA5 and NCEP-NCAR exhibited significant errors in the meteorological elements (Figure 2 and Figure 6). Specifically, ERA5 showed the substantial overestimation of precipitation and relative humidity, along with the significant underestimation of PET and wind speed. Similarly, NCEP-NCAR significantly overestimated temperature and underestimated wind speed. In contrast, CRU provided a more accurate simulation of precipitation and PET in the Tianshan region, making it the optimal dataset for this area, although the errors in precipitation and PET were still non-negligible. For the basin regions represented by Stations B and C, the GMEs underestimated SPEI before 1990 and overestimated it in the subsequent period, with this phenomenon being more pronounced in southern Xinjiang (Figure 7(a1–c1) and Figure 8(b1,c1)). NCEP-NCAR exhibited significant errors across all meteorological elements, with substantial overestimations and underestimations of precipitation and PET, rendering it unsuitable for application in Xinjiang. CRU and ERA5 demonstrated stronger capabilities in simulating low precipitation in basin regions, where errors were primarily driven by PET. PET is mainly influenced by temperature, wind speed, and relative humidity. CRU significantly overestimated temperatures in the northern Xinjiang basins but accurately simulated temperature trends (Figure 8(b4)). Unfortunately, CRU did not provide wind speed and relative humidity data, leaving considerable uncertainty regarding the internal causes of PET errors. ERA5 demonstrated strong capabilities in representing both the magnitude and trends in temperature and relative humidity. However, it significantly overestimated wind speeds in basin regions and failed to capture the actual wind speed trends. By analyzing the temporal variations in PET and other meteorological elements, we found a corresponding relationship between significant changes in observed PET and wind speed data. For instance, the substantial decrease in the observed wind speed over the last 30 years corresponded to a similar decrease in PET. However, the GMEs failed to accurately simulate the sustained decrease (or significant increase) in wind speed trends. We concluded that errors in wind speed were the main factors causing the GMEs’ SPEI to shift from overestimation to underestimation. Similar results appeared in the basin regions and across Xinjiang in both ERA5 and NCEP-NCAR datasets (Figure 6 and Figure 8).

4. Discussion

The spatial matching of observational data with GMEs was imperative for accurate evaluation. Two widely adopted methods facilitated this process. The first method involved upscaling meteorological station data, wherein interpolation techniques were employed to scale point-scale station data to a spatial resolution of 0.25° [22,27]. It is noteworthy that the sparse network of measurement stations may inadequately represent regional meteorological parameters, introducing errors that could potentially influence assessment outcomes. Consequently, alternative research explored downscaling based on GMEs, wherein point data were extracted directly using coordinates from meteorological stations. This approach aimed to mitigate potential errors arising from the limited representation of the region’s meteorological features in station data [46,48]. We evaluated GMEs using CN05.1 data, which were derived from the interpolation of over 2400 national-level stations across China, including the 102 stations used in this study. Given the relatively sparse density of observational stations in Xinjiang, we employed downscaled interpolation based on GMEs to validate the accuracy of the GMEs. Figure A4 illustrates the SPEI values generated by GMEs at point scale compared to the station data. It was observed that the temporal variations in SPEI at a point scale aligned with the results obtained at a grid scale. The high correlation (CC = 0.98) between CN05.1 data and station data further substantiated the reliability of CN05.1 as reference data. At the point scale, GMEs exhibited a nearly perfect reproduction of the spatial distribution of SPEI deviations compared to the reference data (Figure A5), detecting overestimation during 1961–1980 and underestimation during 1991–2010. When assessing the capability of GMEs to capture major drought events in Xinjiang, we additionally presented the SPEI distribution of station data during specific drought years (Figure 5(e1–e6)), revealing consistency between CN05.1 and station data in depicting drought and severe drought occurrences. Both spatial matching interpolation methods yielded identical outcomes.

To facilitate the better utilization of GMEs, a deeper analysis of potential sources of errors was warranted. Regarding both the spatiotemporal simulation of SPEI and drought monitoring capabilities, GMEs demonstrated superior performances in characterizing northern Xinjiang compared to southern Xinjiang, possibly due to the lower density of stations in the south. Generally, datasets deemed accurate exhibited strong capabilities in representing both the magnitude of and trend in observed variables. While CRU demonstrated good performance among the three GMEs, uncertainty arose due to the low and uneven distribution of observational stations in Xinjiang, which affected the accuracy of data interpolation based on point data. Moreover, in the latter half of the 20th century, there was a reduction in the number of observational points in China. For instance, the average number of stations used to generate CRU precipitation data for China (Xinjiang) decreased from approximately 380 (21) in the 1960s to less than 270 (15) in the 21st century [23]. Therefore, improvements are necessary for CRU to better simulate meteorological elements, addressing issues such as the underestimation of precipitation and PET in Xinjiang and the overestimation of temperature. Additionally, reanalysis data were used to accurately simulate the trends in temperature, precipitation, and PET, with results that were consistent with previous research findings [27,41,43,48,49].

However, numerical errors could not be overlooked. Particularly concerning precipitation, reanalysis data tended to overestimate rainfall frequency compared to observational data, leading to a significant overestimation of precipitation in Xinjiang [48]. Previous studies indicated that atmospheric models tend to release convective instability too easily, resulting in excessive convective triggering in simulations using cumulus parameterization schemes. This tendency led to an overestimation of the frequency of moderate- to low-intensity precipitation events [77,78,79]. The selection of microphysical schemes and land surface models, as well as the ability to resolve terrain-induced drag and valley circulation in complex terrains, also affected precipitation simulations in the region [48,80,81]. It is worth noting that reanalysis data may not accurately reflect changes in near-surface wind speeds in Xinjiang and cannot accurately simulate the phenomena of the stilling and recovery of surface winds, which is consistent with previous studies [26,31]. Possible reasons for this discrepancy may include the treatment of wind speed as a non-model-predicted variable, where different model diagnostic methods (based on assumptions about the stability of the model’s lower layers) may have introduced systematic biases [82]. Differences in underlying surface elevation and inappropriate representations of atmospheric boundary layer processes in the model could also have resulted in assimilation data being unable to effectively capture surface wind speeds [83]. Moreover, the assimilation data’s consideration of surface roughness and the decadal changes in the ocean–atmosphere coupling system, which significantly contributed to the weakening of wind speeds in the Northern Hemisphere, were less accounted for [31,84]. These factors thus became important contributors to wind speed errors.

In arid regions, the performance of datasets has a profound impact on drought management. High-quality drought index datasets provide accurate meteorological, soil moisture, and precipitation data, which can effectively reflect the intensity, extent, and duration of drought. These datasets support early drought prediction, enabling governments and relevant agencies to issue warnings and develop emergency plans, thereby reducing the damage drought causes to agriculture, ecology, and the economy [85]. Additionally, datasets play a crucial role in water resource management, particularly in arid regions facing water scarcity [86]. By analyzing hydrological and climate data, water reservoir scheduling, irrigation system distribution, and groundwater extraction strategies can be optimized to ensure water supply [87,88]. Moreover, agricultural production in arid regions depends on accurate climate data, allowing farmers to adjust planting schedules, crop varieties, and irrigation plans, thus mitigating drought-related losses and ensuring food security [89]. In terms of ecological restoration and management, long-term meteorological and ecological data can be used to assess the effectiveness of restoration efforts and help to prevent land desertification [17,90]. As climate change intensifies, the frequency and severity of droughts may change, and the performance of datasets determines the ability to predict future drought patterns, providing a scientific basis for adaptive policies.

However, all climate products possess inherent uncertainties, primarily arising from data quality issues, periods of data unavailability, and/or the insufficient spatial coverage of observations [68]. Several studies have assessed the performance of drought datasets in arid regions. Turco et al. evaluated the quality of long-term continuous climate data based on the Standardized Precipitation Index (SPI) for meteorological drought monitoring [91]. Subsequently, a new global land grid dataset was developed through the application of an ensemble method used in weather/climate prediction studies. Zhang et al. constructed a daily evapotranspiration deficit index (DEDI) based on daily actual evapotranspiration (AET) and potential evapotranspiration (PET), as determined from high-resolution ERA5 reanalysis data, and systematically assessed the ability of DEDI to identify intense drought events and characterize the evolution of drought [92]. DEDI was able to accurately detect the occurrence and evolution of an extreme drought event in Yunnan Province, Southwest China, in 2019. In addition, some global datasets have been developed. For example, the 1961–2015 Xinjiang regional SPEI drought index dataset provides data support for analyzing the long-term trends of and variability in droughts in Xinjiang at various time scales, and can also serve as a foundation for exploring the impacts of drought events on ecological hydrological environments and agricultural production [93]. Zhang et al. developed the Microwave Integrated Drought Index (MIDI), which, at a one-month scale, showed statistically significant correlations with SPI/SPEI across a broad bioclimatic zone from drought to wet conditions [94]. Liu et al. created a global SPEI dataset based on precipitation from the ERA5 dataset and potential evapotranspiration from the Singer dataset, providing important support for studying the spatiotemporal dynamics of drought events [93]. The aforementioned research findings enable the more comprehensive capture of the uncertainties associated with meteorological droughts, offering new opportunities for global drought monitoring, climate change impact assessments, and decision support.

In Xinjiang, long-term drought datasets are scarce. Before utilizing GME products for specific research or operational activities, comprehensive evaluations were deemed necessary. The performance assessment results provided crucial information and guidance for selecting the most appropriate products for specific applications. Xinjiang has exhibited a climatic characteristic of “wet–dry transition” over the past 60 years [68], highlighting the importance of effective long-term drought monitoring. Consequently, this study focused on evaluating a limited number of products with long-term records (CRU, ERA5, and NCEP-NCAR) to assess their drought monitoring capabilities. In the future, the exploration of more high-resolution products—including gauge-interpolated, satellite-estimated, reanalysis, and multi-source merged data—can be undertaken to enhance their application in short-term drought detection. Additionally, the drought event detection method employed in this study was based on a single SPEI threshold. Future research could benefit from using multi-threshold operations to provide deeper insights into the diagnosis of drought events.

5. Conclusions

We evaluated the drought monitoring performance of three Global Meteorological Estimates (GMEs)—CRU, ERA5, and NCEP-NCAR—in Xinjiang. These products differed in spatial resolution, record length, and data construction techniques. We used CN05.1 data with a spatial resolution of 0.25° and in situ records from 102 stations as the ground truth. Continuous and categorical statistical metrics were employed to assess the performance of the GMEs in simulating annual and seasonal droughts in Xinjiang.

(1)

The results indicated that all GMEs significantly overestimated the SPEI before 1990 and substantially underestimated it after 1990. Among the GMEs, ERA5 exhibited the best overall performance in terms of absolute error (AE) throughout Xinjiang. However, CRU performed better in terms of correlation coefficient (CC) and root-mean-square error (RMSE) in northern Xinjiang, while ERA5 demonstrated stronger representation capabilities in southern Xinjiang. Categorical statistical metrics yielded similar results, with CRU excelling in drought detection in northern Xinjiang and ERA5 showing superior drought detection capabilities in the southern basin regions.

(2)

Differences were observed among the three GMEs in representing the geographical distribution and severity of drought events. CRU and ERA5 performed well in capturing the six major droughts in Xinjiang and were notably superior to NCEP-NCAR. Both CRU and ERA5 exhibited relatively accurate simulations of drought conditions in northern Xinjiang. They accurately represented the droughts in 1962, 1974, and 1997, and correctly reflected the absence of drought in northern Xinjiang in 1980. However, they underestimated the severity of the drought in 1965 and overestimated it in 1977. However, considerable uncertainty was observed among the three GMEs in the southern Xinjiang region.

(3)

Additionally, the GMEs exhibited varying levels of accuracy in reproducing different meteorological elements. In terms of precipitation and PET, CRU demonstrated the highest quality, despite the presence of underestimation in Xinjiang. For temperature, ERA5 performed relatively well in northern Xinjiang, while CRU performed better in southern Xinjiang. Both datasets exhibited underestimation in mountainous areas and overestimation in basin areas. However, ERA5 and NCEP-NCAR did not capture wind speed satisfactorily, with both significantly underestimating wind speeds in Xinjiang. ERA5 demonstrated better quality in relative humidity data, whereas NCEP-NCAR overestimated relative humidity in Xinjiang. Among the GMEs, differences were observed in the leading factors contributing to the transition of SPEI from overestimation to underestimation around 1990. CRU and ERA5 were mainly influenced by changes in PET, whereas NCEP-NCAR was influenced by both precipitation and PET. Wind speed played a significant role in driving the differences in PET variation. The inability of the GMEs to accurately simulate the sustained decrease (or notable increase) in surface wind speed led to the transition of SPEI from overestimation to underestimation around 1990.

In conclusion, our findings suggest that ERA5 and CRU are valuable resources for studying long-term wet–dry variations in Xinjiang, effectively enabling the accurate monitoring of drought events. However, both datasets have certain limitations, which users should carefully consider when applying them.