Drought Driving Factors as Revealed by Geographic Detector Model and Random Forest in Yunnan, China

Abstract

Yunnan Province, as a critical ecological security barrier in China, has long been highly susceptible to drought events. Characterizing the spatiotemporal distributions of drought and identifying its driving factors is crucial. Due to the complexity of drought occurrence, linear correlation analysis alone is insufficient to quantify drought drivers and their interactions. This study used the Standardized Precipitation Evapotranspiration Index (SPEI) as a drought indicator to analyze the spatiotemporal trends of drought across Yunnan and its six major river basins. The geographic detector model (GDM) and random forest (RF) were utilized to quantify the impacts of meteorological, topographical, soil, and human activities on drought, as well as the interactions among these factors. The results showed that 63.61% of the study area exhibits a significant drying trend (p-value < 0.05), with the Jinsha River Basin (JSRB) experiencing the highest frequency of extreme drought events. Precipitation (PRE), temperature, potential evapotranspiration (PET), vapor pressure deficit (VPD), and relative humidity (RH) were identified as the primary controlling factors of drought, with factor interactions displaying nonlinear enhancement effects. PRE plays a dominant role in driving drought across Yunnan, whereas elevation primarily influenced drought severity in the JSRB, Lancang River Basin (LCRB), and Nujiang River Basin (NJRB). The RF-based SPEI prediction model demonstrated superior performance in simulating short-term drought (SPEI_1, R² > 0.931, RMSE < 0.279), particularly in the JSRB (R² = 0.947 RMSE = 0.228). These findings provide a scientific basis for regional water resource management applications and drought early warning systems, offering a robust framework for understanding and mitigating drought impacts in ecologically sensitive regions.

1. Introduction

Drought is a complex natural disaster characterized by its slow onset, prolonged duration, extensive spatial impact, and high frequency of occurrence, making it one of the most devastating natural hazards worldwide [1,2,3,4]. With the intensification of climate change and the increasing frequency of extreme weather events, drought risks have escalated significantly, leading to a notable rise in their unpredictability and uncontrollability [5,6]. Therefore, the objective quantification of regional drought driving factors and their mechanisms is of significant importance for the prevention, monitoring, and management of droughts.

Droughts are typically categorized into meteorological, agricultural, hydrological, and socioeconomic droughts [7]. Meteorological drought is primarily triggered by a deficiency in precipitation (PRE) [8]. Agricultural drought manifests as a sustained decline in soil moisture over time, persisting at lower levels [9,10], while hydrological drought is closely associated with reductions in surface water, groundwater resources, and streamflow [4].

Socioeconomic drought further encompasses the economic, social, and environmental impacts caused by drought or water scarcity [11]. To quantify drought conditions, researchers commonly employ multiple drought indices for comprehensive monitoring and analysis. Among these, the assessment of meteorological drought primarily relies on the Standardized Precipitation Evapotranspiration Index (SPEI) [12,13,14,15]. The SPEI not only retains the simplicity and multi-temporal scalability of the Standardized Precipitation Index (SPI) but also incorporates potential evapotranspiration (PET), which reflects the balance between water supply and demand at the surface, into its calculations [16,17,18]. This makes it one of the most widely used drought indices. The SPEI can be calculated across multiple time scales, each corresponding to distinct research objectives [18]. For instance, SPEI_1 (monthly scale) captures short-term meteorological drought with high sensitivity to monthly PRE and evapotranspiration variations, making it suitable for short-term water balance analysis [19]. SPEI_3 (seasonal scale) is commonly used to monitor seasonal drought and assess water availability during the growing seasons [13,19]. In contrast, SPEI_6 and SPEI_12 provide insights into medium- and long-term climatic drought trends and patterns [16]. The multiscalar nature of SPEI allows for a comprehensive assessment of drought evolution across different temporal resolutions [18]. Agricultural drought is assessed through the Standardized Soil Moisture Index (SSMI) [20,21], while hydrological drought is characterized using the Standardized Runoff Index (SRI) [22,23]. Since agricultural and hydrological droughts are often directly or indirectly triggered by meteorological drought [21,22,23,24], and socioeconomic drought reflects the cumulative long-term effects of meteorological drought [25], meteorological drought serves as the root cause and driving force behind other types of drought. Given this causal relationship, this study focuses on meteorological drought as the research subject, utilizing the SPEI as the drought index, to thoroughly investigate the driving mechanisms of meteorological drought and its regional heterogeneity.

Previous research on the driving factors of drought has achieved significant progress. Wang et al. [26] revealed the impact of teleconnection factors on drought in major river basins of China through wavelet coherence analysis; Ma et al. [27] employed attribution analysis to quantify the contributions of PRE and PET to drought in Southwest (SW) China; Song et al. [28] employed wavelet coherence analysis to investigate the impact of global climatic events on spring droughts in SW China; Wang et al. [29] utilized linear correlation analysis to investigate the influence of atmospheric circulation anomalies on extreme drought variations in Yunnan; Lan et al. [30] analyzed the spatiotemporal characteristics of drought in Yunnan from 1961 to 2020 and its relationship with teleconnection factors based on the SPEI; Qin et al. [31] elucidated the predominant role of reduced relative humidity in atmospheric drought in Yunnan through partial correlation analysis. These studies have provided a significant foundation for understanding the driving mechanisms of regional droughts. However, the processes leading to drought occurrence are exceptionally complex, involving a multitude of factors such as meteorological conditions, topography, soil properties, and human activities, along with their interactions [7,32,33]. Traditional correlation analysis methods, including linear and partial correlation analyses, are inadequate for comprehensively revealing the intensity of these factors and their interactive mechanisms across different regions. To address these gaps, this study employs a multi-source data fusion approach, utilizing geographic detector model (GDM) to quantify the spatial heterogeneity characteristics of various driving factors and their interaction intensities [34]. In addition, based on a comprehensive analysis of driving factors, this study constructed and integrated a drought prediction model, aiming to provide a scientific basis for early warning and response strategies for regional droughts.

Yunnan is located in SW China, characterized by complex topography and diverse climatic conditions, serving as both a vital ecological zone and a key agricultural region [35,36,37]. However, the region is facing severe drought threats, such as the extreme droughts in 2009–2010 and 2019, which caused significant economic losses and ecological stress [38,39,40]. Therefore, investigating the driving mechanisms of droughts in Yunnan is crucial for addressing drought risks.

Based on the above context, this study aims to reveal the spatiotemporal evolution patterns and driving mechanisms of meteorological drought in Yunnan. Specifically, we calculated the SPEI at different time scales using data from 108 weather stations in Yunnan, analyzing drought trends in Yunnan and its subregions from 1979 to 2023. By integrating the GDM and RF, we identified the main controlling factors of drought variability and their spatial heterogeneity and further developed a drought prediction model to provide a scientific basis for regional drought prevention and control.

2. Methods

2.1. Study Area

The study area is Yunnan Province, SW China, covering an area of approximately 394,100 square km², with a unique geographical location bordering Vietnam, Laos, and Myanmar (Figure 1). The province has a diverse and complex topography, with the highest elevation (ELE) in the northwest reaching 5620 m, and the lowest point in the southeast at 79 m (Figure 1b). The landscape features mountainous terrain and valleys in the west, while the eastern region is characterized by karst landforms. Yunnan has six major river basins: the Jinsha River Basin (JSRB; the ELE ranges from 249 to 5620 m), Yuanjiang River Basin (YJRB; the ELE ranges from 79 to 3318 m), Nanpan River Basin (NPRB; the ELE ranges from 165 to 2852 m), Lancang River Basin (LCRB; the ELE ranges from 486 to 5461 m), Nujiang River Basin (NJRB; the ELE ranges from 464 to 4784 m), and Irrawaddy River Basin (IRRB; the ELE ranges from 268 to 5075 m). There are 108 weather stations in Yunnan, of which 37 are located in the JSRB, 20 in the LCRB, 23 in the NPRB, 6 in the NJRB, 6 in the IRRB, and 16 in the YJRB. The province’s large but variable water resources and complex topography make it an ideal region for studying drought and its driving factors.

2.2. Data Description

Four primary data categories are used in this study: climate, topography, soil, and human activity. Climate data include PRE, maximum temperature (Tmax), minimum temperature (Tmin), mean temperature (Tmean), PET, relative humidity (RH), and vapor pressure deficit (VPD). VPD represents the degree of atmospheric moisture deficit, with a higher VPD indicating higher atmospheric evaporative demand [31]. The calculation of VPD follows the method outlined in [31]. This study employed the Hargreaves method to calculate PET, which has been validated as an effective approach and has been widely applied [15,41,42,43]. PRE serves as the primary input of water, while temperature, PET, RH, and VPD influence water output. Topographic data include ELE, aspect (ASC), and slope (SLO). ASC and SLO were extracted from ELE in ArcGIS 10.8. Topographic features influence the spatial distribution of climatic variables [31,44]. In this study, DEM SRTM data were selected. Soil data include soil type (ST) and soil moisture (SM). ST influences water retention, permeability, and plant water availability, while SM serves as a direct drought indicator, reflecting actual water availability and aiding in drought identification and intensity assessment [45,46]. Human activity data include land use cover/use (LULC), population density (PD), and human footprint (HF). LULC reflects direct human modifications to the environment, while PD and HF influence regional water resource distribution [47,48]. These factors all play important role in the hydrological cycle. Most of the data were resampled to a spatial resolution of 250 m using ANUSPLIN. As a spatial interpolation method, ANUSPLIN demonstrates robust reliability and precision in downscaling applications [49]. As discrete data, LULC (land use/land cover) was resampled to a uniform resolution of 250 m using the nearest neighbor allocation method. The workflow of this study is illustrated in Figure 2. All data were categorized into nine classes using the natural breaks classification method in ArcGIS 10.8 (Figure 3).

Table 1. Variables used in this study.

Variable Number	Indexes	Abbreviations	Time Coverage	Data Types	Sources
1	Precipitation (mm)	PRE	1979–2023	Observed	http://data.cma.cn/
2	Maximum temperature (°C)	Tmax	1979–2023	Observed	http://data.cma.cn/
3	Minimum temperature (°C)	Tmin	1979–2023	Observed	http://data.cma.cn/
4	Mean temperature (°C)	Tmean	1979–2023	Observed	http://data.cma.cn/
5	Potential evapotranspiration (mm)	PET	1979–2023	Derived	-
6	Relative humidity (%)	RH	1979–2023	Observed	http://data.cma.cn/
7	Vapor pressure deficit (kpa)	VPD	1979–2023	Derived	-
8	Elevation (m)	ELE	-	Raster	https://www.resdc.cn/
9	Aspect	ASP	-	Raster	-
10	Slope (°)	SLO	-	Raster	-
11	Soil types	ST	-	Vector	https://www.resdc.cn/
12	Soil moisture (cm3/cm3)	SM	2000–2020	Raster	https://www.gleam.eu/
13	Land use/cover	LULC	1985–2020	Raster	http://doi.org/10.5281/zenodo.4417809 (accessed on 16 January 2025)
14	Population density (population/km2)	PD	2000–2020	Raster	https://www.resdc.cn/
15	Human footprint	HF	2000–2020	Raster	https://www.resdc.cn/

provides specific information on the variables used in this study.

2.3. Drought Index

In this study, we selected the SPEI as the drought index. SPEI is a drought index based on PRE and PET. It describes the deviation of wet and dry conditions by standardizing the difference between PET and PRE [18]. The calculation method for SPEI is as follows:

Calculate net positive PRE:

$$D_j=P_j-{PET}_j$$

(1)

$D_j$ represents the net PRE for month j in mm; $P_j$ and ${PET}_j$ denote the average PRE and PET for month j, respectively, in mm.

Calculate the accumulated difference between the PRE and the PET at different time scales (1-month, 3-month, 6-month, and 12-month scales). The accumulated difference ($x_{i,j}^k$) at the k-month time scale is calculated using

$$\left\{\begin{array}{cc}{x}_{i,j}^k={\sum }_{l=13-k+j}^{12}{D}_{i-1,l}+{\sum }_{l=1}^j{D}_{i,l}& ifj<k\\ x_{i,j}^k={\sum }_{l=j-k+1}^i{D}_{i,l}& ifj\ge k\end{array}\right.$$

(2)

where $x_{i,j}^k$ is the accumulated difference between the PRE and the PET at the k-month time scale in the j-th month of the i-th year; D_i,l is the monthly difference between the PRE and the PET in the l-th month of i-th year.

Normalize the $x_{i,j}^k$ data sequence. Because there may be negative values in the original data sequence $x_{i,j}^k$, the SPEI uses the three-parameter log-logistic probability distribution suggested by [18]. For the data sequences of all timescales, the accumulative function of the log-logistic probability distribution F(x) is as follows:

$$F(x)={\left[1+{\left({\displaystyle \frac{\alpha }{x-\gamma }}\right)}^{\beta }\right]}^{-1}$$

(3)

where α, β, and γ are scale, shape, and position parameters, respectively, which can be calculated using the equations proposed by [18].

p is the probability of a definite $x_{i,j}^k$ value:

$$p=1-F\left(x\right)$$

(4)

If p ≤ 0.5,

$$\omega =\sqrt{-2Inp}$$

(5)

$$SPEI=\omega -{\displaystyle \frac{c_0+c_1\omega +c_2{\omega }^2}{1+d_1\omega +d_2{\omega }^2+d_3{\omega }^3}}$$

(6)

If p > 0.5,

$$\omega =\sqrt{-2In(1-p)}$$

(7)

$$SPEI={\displaystyle \frac{c_0+c_1\omega +c_2{\omega }^2}{1+d_1\omega +d_2{\omega }^2+d_3{\omega }^3}}-\omega$$

(8)

In this study, we calculated the SPEI for different time scales, including 1-month (SPEI_1), 3-month (SPEI_3), 6-month (SPEI_6), and 12-month (SPEI_12) periods, to comprehensively assess the drought characteristics across various time scales.

2.4. Analysis of Drought Trends

The Sen’s Slope Estimator is a non-parametric method used to quantify trend intensity in time series data, making it particularly suitable for analyzing datasets in fields such as climatology, ecology, and hydrology [50]. Its primary advantage lies in its independence from distributional assumptions, making it effective for handling time series data that deviate from normal distributions or contain outliers and extreme values [51]. Similarly, the Mann–Kendall trend test is a non-parametric statistical method that does not require assumptions of independence or normality, making it effective for detecting trends in time series data [52]. The drought and wetness types were categorized based on the outcomes of the trend analysis (

Table 2. Criteria for determining drought and wetness types.

|Z|	Types of Drought and Wetness
≥0	Wetting
≥1.96	Significant wetting
≥2.58	Extremely significant wetting
<0	Drying
≤−1.96	Significant drying
≤−2.58	Extremely significant drying

).

$$Slope=median\left({\displaystyle \frac{y_j-y_i}{x_j-x_i}}\right)$$

(9)

where i and j represent different pairs of data points with i < j, and (x_i, y_i) and (x_j, y_j) are the coordinates of these data points. Positive values of S denote growing or increasing trends in the data, while negative values suggest declining or decreasing trends. A value of zero indicates the absence of a discernible trend in the data [31]. This interpretation of slope is commonly used in change point analysis and trend detection to assess the direction and magnitude of trends in time series data [31].

2.5. Geographical Detector Model (GDM)

GDM is a statistical method used to analyze the relationships between different variables in geographical spatial data, primarily to reveal the spatial distribution characteristics of geographical phenomena and the influence of their driving factors [34,53]. It detects the spatial relationship between different factors and the target variable, determines the contribution of each factor to the target variable, and then evaluates the relative importance of these factors [34,53], and it is calculated as follows:

$$q=1-{\displaystyle \frac{{\sum }_{h=1}^L{N}_h{\sigma }_h^2}{{N\sigma }^2}}$$

(10)

where h = 1,…, represents the stratification (classification) of the variables; $N_h$ and N denote the number of cells in the h-th stratum and the entire region, respectively; ${\sigma }_h^2$ and ${\sigma }^2$ are the variances of the dependent variable Y within the h-th stratum and across the whole region, respectively, and dependent variable in stratum h and the whole region, respectively. The value of q lies in the range of [0, 1], with higher values of q indicating stronger explanatory power, and lower values indicating weaker explanatory power.

Interaction detection determines whether the interactions between factors enhance or reduce the explanatory power of the dependent variable or if the factors influence the dependent variable independently of one another.

Table 3. Interactivity types of variables.

Judgment Criteria	Interactivity Types
q(X1 ∩ X2) < Min ((q(X1), q(X2))	Nonlinear—weakened
Min ((q(X1), q(X2)) < q(X1 ∩ X2) < Max ((q(X1), q(X2))	Nonlinear—weakened by one
q(X1 ∩ X2) > Max ((q(X1), q(X2))	Mutually enhanced
q(X1 ∩ X2) = q(X1) + q(X2))	Independent effect
q(X1 ∩ X2) > q(X1) + q(X2))	Nonlinear—enhanced

shows the specific interactivity types.

In this study, the 15 variables listed in Table 1 and Figure 3 are used as independent variables in the model, with SPEI as the dependent variable. Based on these inputs, the model outputs the results of factor detection and interaction detection.

2.6. Random Forest (RF)

RF is an ensemble learning method that solves classification and regression problems by integrating multiple decision trees. The core idea is to use bootstrap sampling to extract subsamples from the dataset to construct multiple decision trees. During prediction, the results are aggregated through majority voting (for classification tasks) or averaging (for regression tasks), thereby improving predictive accuracy and effectively reducing overfitting [54,55]. Random forest selects a random subset of features at each decision tree split, a characteristic that not only enhances the model’s generalization ability but also provides valuable feature importance assessments, helping to identify key influencing factors [7].

This study combines the GDM and RF methods to analyze the driving factors of drought and develop predictive models. GDM reveals the explanatory power of different variables on the spatial distribution characteristics of drought, identifies key driving factors and their interactions, and provides the spatiotemporal patterns of drought occurrences. On the other hand, random forest, with its powerful classification and regression capabilities, handles the nonlinear relationships between drought-related variables, identifies the importance of factors, and constructs precise predictive models. By combining both methods, the model not only deepens the understanding of the geographical characteristics of drought but also improves the accuracy and interpretability of predictions.

This study employed an integrated approach combining the GDM and RF algorithm to investigate drought-driving mechanisms and develop predictive frameworks. The GDM method quantifies the explanatory power of environmental variables on the spatial heterogeneity of drought patterns, identifies dominant drivers through factors and interaction detection, and delineates spatiotemporal drought dynamics. The RF algorithm leverages its robust capacity in handling high-dimensional data and nonlinear relationships to rank variable importance, optimize feature selection, and construct high-precision drought prediction models.

The dataset was randomly split into a training set (70%) and a validation set (30%) to ensure the robustness of the model. To improve model performance, grid search was used to optimize key parameters. The model’s predictive performance was evaluated using multiple metrics, including the coefficient of determination (R²), root mean square error (RMSE), Mean Squared Error (MSE), standard deviation (SD), and correlation coefficient (COR), to describe the model’s fitting effect and predictive capability.

3. Results

3.1. The Spatiotemporal Characteristics of SPEI

The results indicate a decreasing trend in SPEI values across all time scales in Yunnan Province from 1979 to 2023 (Figure 4). Among the time scales, the downward trend is most pronounced for SPEI_6 (p-value < 0.01, slope = −0.09/10 a) and SPEI_12 (p-value < 0.01, slope = −0.09/10 a), indicating that these two time scales are more suitable for long-term drought monitoring. In contrast, SPEI_1 (p-value < 0.05, slope= −0.04/10 a) and SPEI_3 (p-value < 0.01, slope= −0.09/10 a) exhibit more noticeable fluctuations, suggesting that the meteorological conditions in the study area are prone to short-term variability, leading to more frequent drought events at monthly and seasonal scales. Additionally, drought fluctuations in Yunnan have become more pronounced in the 21st century (Figure 4).

Most river basins in Yunnan exhibit a drying trend in SPEI_12 (p-value < 0.05), with 63.61% of the area showing a significant drying trend and 15.64% (p-value < 0.01) showing an extremely significant drying trend (Figure 5). The differences in drought trends among river basins are noteworthy: the JSRB has the largest proportion of area experiencing an extremely significant drying trend (31.52%, p-value < 0.01), while the NPRB has the largest proportion of area experiencing a significant drying trend (88.27%, p-value < 0.05). Additionally, the proportions of area with extremely significant drying trends are 19.38% (p-value < 0.01) in the YJRB and 18.06% (p-value < 0.01) in the LCRB, while the proportions of area with significant drying trends are 71.56% (p-value < 0.05) in the YJRB and 67.46% p-value < 0.05) in the NJRB.

Overall, the JSRB, LCRB, and NJRB exhibit higher proportions of extreme drought areas compared to other basins, whereas the NPRB and YJRB have higher proportions of significant drought areas. Further analysis of intra-basin variations shows that the JSRB demonstrates the most pronounced differences in wet and dry conditions, particularly in the upstream and downstream regions, where extreme droughts are more prevalent (Figure 5). Similarly, the upstream regions of the LCRB, NJRB, and IRRB show significantly higher proportions of extreme drought areas compared to their midstream and downstream regions. These findings highlight the substantial spatial disparities in drought trends across different river basins, warranting further investigation and attention.

3.2. Drought Driving Factors Analysis Based on GDM

3.2.1. Factors Detection

In terms of the explanatory power of various driving factors of drought in Yunnan Province, PRE (q = 0.24, p-value < 0.01) had the highest explanatory power (Figure 6). Following this was VPD (q = 0.11, p-value < 0.01), RH (q = 0.09, p-value < 0.01), and temperature factors, which have relatively high explanatory power. Among these, Tmin (q = 0.05, p-value < 0.01), Tmean (q = 0.05, p-value < 0.01), and Tmax (q = 0.04, p-value < 0.01) show lower explanatory power. Additionally, PET, (q = 0.02, p-value < 0.01) and SM (q = 0.02) have weak explanatory power. Overall, the results show that climatic factors strongly influence drought, with water-related factors being more important than temperature factors.

Figure 6b–g illustrates that drought across different river basins is primarily driven by PRE, with an explanatory power ranging from 0.22 to 0.48, strongly surpassing other factors. Moisture-related variables (RH, VPD, and PET) and SM also exhibit substantial influence, with SM playing a larger role in the LCRB. Temperature factors (Tmin, Tmean, and Tmax) have stronger impacts in LCRB and YJRB, but their influence is weaker in other regions. Topographic factors, including ASC, ELE, and SLO, show limited impact overall. In summary, climatic factors dominate drought dynamics, while the contributions of other factors are relatively minor.

3.2.2. Interaction Among Factors

The interaction analysis results indicated that most factors exhibit nonlinear enhancement, where the combined effect exceeds the sum of their individual influences (Figure 7). Notably, the interaction between PRE and temperature-related factors (Tmean, Tmax, and Tmin) yields q-values exceeding 0.50 (p-value < 0.01) across all river basins (Figure 7), suggesting that the combined influence of PRE and temperature amplifies drought impacts. In particular, the interaction intensity between PRE and temperature factors surpasses 0.70 in the LCRB and YJRB basins. With the exception of temperature, all other factors exhibit nonlinear enhancement in their interactive effects with PRE (p-value < 0.01), which further underscores the dominant role of PRE in drought processes.

In contrast, the interaction effects between topographical factors and climate factors are relatively weak, but in some areas (such as the YJRB), they still show some degree of influence, suggesting that topography affects the distribution of climate conditions and water resources, thereby indirectly influencing the occurrence of drought. The interaction between human activity factors and climate factors is also weak, but in LCRB and YJRB, human activities indirectly affect drought occurrence through land use changes and other factors. The above results indicate that the interaction effects between PRE and temperature are the primary drivers of regional drought, particularly in LCRB and YJRB, where their enhancement effects greatly amplify drought risks.

3.3. Drought Driving Factors Analysis Based on RF

3.3.1. Driving Factors Analysis

The random forest analysis revealed the contributions of various factors to changes in SPEI (Figure 8). Consistent with the findings from GDM, PRE emerged as the most critical drought-driving factor across all basins, explaining between 36.65% and 48.95% of SPEI variations. Its significance was particularly evident in the NJRB (46.84%), NPRB (48.95%), and IRRB (42.77%), highlighting that PRE variability was the primary determinant of drought severity in these regions. Temperature-related factors and PET followed as the next most important contributors. Other factors, such as RH, had relatively minor overall impacts on SPEI variations but showed notable effects in specific basins. Such as, RH accounted for 6.97% of the variability in JSRB.

Factors such as SLO, VPD, and SM generally had lower overall contributions but exhibited notable at the basin scale (Figure 8). For example, SLO explained 2.30% and VPD 4.30% of the variability in NPRB. Although secondary factors such as LULC, PD, and HT had lower overall importance in the analysis, they still influence specific patterns of SPEI variation in certain basins. Furthermore, the random forest analysis emphasized the critical role of ELE in influencing SPEI changes in Yunnan, particularly in regions with significant topographical variation (Figure 8a,b,e,g), where ELE was identified as a key factor.

3.3.2. Drought Modeling Based on Driving Factors

The GDM and RF indicated that climate factors, especially PRE, are the primary drivers of SPEI variation in Yunnan (Figure 6 and Figure 8). Other climate factors, topography, and soil factors follow in importance. The SPEI prediction model based on these factors performs best at the monthly scale (SPEI_1), with the lowest RMSE (0.160) and MSE (0.026), highest R² (0.977), and highest COD (0.989), demonstrating strong accuracy and stability (Figure 9). As the time scale increases, RMSE rises, and R² and COD decrease, but they remain relatively high. By eliminating redundant variables and focusing on key factors, the prediction accuracy of SPEI, especially for short-term droughts, is improved (Figure 9).

This study revealed river basins differences in predictive performance by comparing the performance of different river basins across multiple evaluation metrics (Figure 10). In SPEI_1, the JSRB and NPRB stand out with high R² values (JSRB: 0.947, NPRB: 0.970) and COR values (JSRB: 0.973, NPRB: 0.985), along with low RMSE (JSRB: 0.228, NPRB: 0.177) values. The R² values for LCRB, YJRB, NJRB, and IRRB all exceed 0.90, indicating excellent model fitting and minimal prediction errors at the SPEI_1 scale. In contrast, the YJRB and NJRB exhibit poorer performance in SPEI_3 and SPEI_6 (Figure 10b,c), particularly in SPEI_3, where their R² values are only 0.758 and 0.640, respectively, and their RMSE values are relatively high, suggesting limitations in predictive ability. At the SPEI_12 scale (Figure 10a), all river basins exhibited R² values exceeding 0.87, with the highest predictive performance observed in JSRB and the lowest in YJRB. Meanwhile, the LCRB and IRRB demonstrate intermediate performance, with moderate R² values and a lower RMSE.

4. Discussion

4.1. Multifactor Driven Drought Patterns

From 1979 to 2023, the intensification of drought in Yunnan and its river basins has been consistent with patterns of global climate change [41,56]. This study indicated that PRE is the most critical factor influencing drought across various river basins of Yunnan (Figure 6 and Figure 8). Based on this, we analyzed the trends of PRE changes in Yunnan Province from 1979 to 2023 (Figure 11). The results reveal a declining trend in PRE across various river basins of Yunnan. Among these, the JSRB exhibited the largest decrease in PRE, with nearly 100% of its area affected, followed by the LCRB, where 97.17% of the area experienced a decline in PRE. In other basins, the area with reduced PRE exceeded 93.00%. In contrast, 7.33% of the NPRB showed an increasing trend in PRE, a phenomenon that differs from trends observed in other basins. There has been a slight increase in PRE in the NJRB, IRRB, and YJRB as well. This modest rise in PRE may be one of the factors contributing to the relatively lower severity of drought in these river basins, particularly in the NPRB (Figure 4). This finding fully corroborates the results of the present study, as well as the research conducted by Ma et al. [27] in Yunnan demonstrating that insufficient PRE is the primary factor contributing to drought conditions in Yunnan. In addition, several studies have demonstrated that although the overall PRE in Yunnan has exhibited a declining trend, alterations in PRE intensity and the onset timing of rainfall have been observed, which may increase the likelihood of extreme drought occurrences [57,58,59]. In addition to PRE changes, the study conducted by Qin et al. [31] also revealed that Yunnan Province has experienced a significant upward trend in Tmean, a continuous increase in VPD, and a decreasing trend in RH. The co-occurring changes of these climatic factors have exacerbated the severity of droughts in this region. In contrast to the findings of this study, some research has identified PET, Tmean, and SM as the primary driving factors of drought evolution [7,57,60]. This discrepancy may be attributed to the heterogeneity of climatic characteristics and underlying conditions across different regions.

Yunnan’s large ELE differences and diverse topographical features have shaped the distribution patterns of climatic variables [31,44]. Therefore, ELE is an indispensable factor in drought-related studies in Yunnan and elsewhere. In this study, the RF analysis also highlighted the influence of ELE on drought (Figure 12). By analyzing the ELE area proportions and the q-values of driving factors across different regions, we found that PRE is the dominant factor universally, but its influence is modulated by other factors in different ELE intervals. In the 0–1000 m ELE range, regions such as YJRB and IRRB have relatively large area proportions, where drought is primarily driven by PRE, RH, and SM. Human activities also play a significant role in this ELE range. Low-ELE areas are typically densely populated, meaning that human activities, such as land use conversion and agricultural irrigation, may alter the values of climatic factors, thereby affecting drought [61,62]. In the ELE range above 3000 m, JSRB has a significantly larger area proportion compared to other regions. In this range, the influence of mean temperature and relative humidity on drought surpasses that of PRE, and the importance of SM also increases. This is likely due to the presence of snow and glaciers in high-ELE areas, where snowmelt serves as a critical water source [63,64]. The seasonality and stability of snowmelt processes partially mitigate the impact of insufficient PRE [65,66]. Additionally, influenced by freeze–thaw cycles and topographic characteristics, soils in high-ELE regions typically exhibit higher water retention capacity [65], further alleviating the direct impact of PRE deficits. In other ELE ranges, PRE remains the dominant factor, with no significant changes observed in the influence of other factors. Overall, PRE is a universal driver of drought across ELE, but the probability of drought being influenced by human activities increases in low-ELE areas, while the influence of Tmean, RH, and SM on drought becomes more pronounced in high-ELE areas. Furthermore, regions such as JSRB, LCRB, NJRB, and IRRB span a wide range of ELEs. Therefore, drought management strategies must account for ELE-induced variations, emphasizing the need for adaptive, region-specific, and stratified management approaches tailored to the unique conditions of each region.

Although human activities are not the dominant factor driving drought, they still exert influence on its occurrence and development [57,67]. In JSRB, the changes in LULC are particularly notable, with 13.60% of forests being con-verted into Farmland, Impermeable surfaces, Shrub, and Grasslands (Figure 13). This transformation reflects the dual influence of urban expansion and agricultural development. Notably, the extent of cropland conversion in the JSRB exceeds that in YN, confirming that the JSRB experiences the most intensive agricultural expansion within the province. While human activities do not directly trigger droughts, they can exacerbate drought severity [67,68]. Such high-intensity land use changes may degrade soil water retention capacity and enhance evapotranspiration, thereby accelerating the occurrence and progression of droughts [69,70]. Compared to JSRB, the land use changes in LCRB, YJRB, NJRB, IRRB, and NPRB are smaller, primarily characterized by the gradual conversion of grasslands and shrubs to forests (Figure 13). This trend is likely closely associated with the implementation of ecological restoration policies in recent years, such as the Natural Forest Protection Program and the Grain for Green Project [71,72]. The implementation of these policies has significantly enhanced the drought resistance capacity of those areas [72].

4.2. Practical Implications of Regional Drought Monitoring Model

The disparities in predictive accuracy across regions indicate that, in drought monitoring and early warning practices, model parameters should be optimized according to the climatic, hydrological, and geographical characteristics of each region to enhance the accuracy and stability of predictions [54,73]. As examples, JSRB and LCRB demonstrated superior model performance at the SPEI-1 and SPEI-12 scales, with R² values exceeding 0.90, indicating that the model effectively captures the drought characteristics of these river basins [74,75,76]. In contrast, the predictive accuracy in the NJRB and IRRB are lower. Specifically, at the SPEI_3 scale, the R² value for NJRB is below 0.70, and the RMSE is the highest, indicating that the drought processes in this region are more complex [74,75,76]. The predictive capability for NPRB falls between the two, possibly reflecting a moderate level of climate variability and the hydrological system’s response to drought events [74,75,76].

For river basins with strong predictive capabilities shown here, such as JSRB and LCRB, the model can be further refined to support more detailed drought risk assessments [77]. On the other hand, for river basins with lower predictive performance, such as NJRB and IRRB, it is necessary to incorporate multi-source data (such as vegetation indices and atmospheric circulation) to improve models [78,79], enhancing their ability to simulate drought processes under complex environmental conditions. In addition, the performance differences among models at different time scales (SPEI_1, SPEI_3, SPEI_6, SPEI_12) also highlight the imbalance between short-term and long-term predictive capabilities. For example, at shorter time scales (SPEI_1, SPEI_3), the predictive ability is weaker in some regions, indicating more variation in short-term drought events. In contrast, at longer time scales (SPEI_6, SPEI_12), the overall predictive performance improves, suggesting that long-term trends are more readily captured by the model [76]. Therefore, in practical applications, it is essential to consider the predictive differences across regions and integrate the results from different time scales to support more beneficial water resource management and agricultural production decisions. This approach will aid in the development of accurate drought early warning systems for other regions.

4.3. Limitations and Future Work

Although this study has revealed the impacts of climate, topography, soil, and human activities on drought, the sparse distribution of weather stations and the low resolution of existing data make it difficult to fully capture the drought characteristics and driving factors at smaller scales, especially in areas with complex terrain or localized microclimate conditions. To address this limitation, deep learning techniques, as a promising approach for developing efficient models, can be used to perform downscaling of reanalysis data, enabling more accurate evaluation of drought characteristics and their driving factors in complex or smaller-scale regions [76,77]. In the future, combining traditional meteorological monitoring, remote sensing technologies, ground observation data, and numerical simulation methods can help develop multi-source information fusion drought monitoring models, further improving the accuracy and comprehensiveness of drought monitoring and forecasting.

5. Conclusions

The main conclusions are as follows: (1) From 1979 to 2023, Yunnan exhibited a drying trend at various temporal and spatial scales. Temporally, SPEI values at different scales showed a declining trend, indicating intensified drought conditions. Spatially, 63.61% (p-value < 0.05) of Yunnan experienced significant drying, with JSRB showing the highest proportion of extremely severe drought areas (31.62%, p < 0.05) compared to other basins, and NPRB having the largest proportion of drying (88.27%), followed by YJRB (71.56%, p < 0.05). (2) Drought in Yunnan is primarily controlled by meteorological factors, with nonlinear interactions among these factors exacerbating drought conditions, especially the interactions between PRE and other variables. Furthermore, in river basins with large ELE heterogeneity (JSRB, LCRB, and NJRB), ELE emerges as a critical factor contributing to drought intensification. (3) RF-based SPEI prediction model was demonstrated high accuracy and stability at short term scales (SPEI_1), characterized by low RMSE, high R² (0.977), low RMSE (0.160), and strong COR (0.989). Among the different river basins, JSRB exhibited the best predictive performance. These findings have important implications for ecosystem managements in response to regional drought.