Deciphering Spatiotemporal Dynamics of Vegetation Drought Resilience in China

Abstract

Under accelerated global warming, frequent droughts pose mounting threats to vegetation productivity, yet the spatiotemporal patterns and primary controls of drought resilience (DR) in China remain insufficiently quantified. This study aimed to characterize DR trends across Köppen–Geiger climate zones in China from 2001 to 2020 and to identify the dominant drivers and their interactions. We constructed a hazard–exposure–adaptability framework, combining multi-source satellite observations and the station data. A Bayesian-optimized Light Gradient Boosting Machine (LightGBM, version 4.3.0) model was trained under five-fold cross-validation. Shapley Additive exPlanations (SHAP) analysis decomposed each driver’s main and interaction effects on DR. The results indicated that DR was better in tropical regions, whereas arid and polar regions require more attention. From 2001 to 2020, 45.3% of China’s land area saw DR increases, while 36.4% declined. The key drivers influencing DR were temperature, sunlight hours, potential evapotranspiration, and precipitation. Notably, an increase in sunlight hours was often accompanied by a decrease in precipitation, resulting in suboptimal DR in China. When the normalized precipitation fell within the range of 0.12 to 0.65, elevated temperature exhibited an inhibitory effect on DR. Overall, this study established a DR assessment framework, elucidated its spatiotemporal dynamics, and revealed key driver interactions, offering timely insights for ecosystem research and management in the face of climate change.

1. Introduction

Global warming has exacerbated hydrological disasters such as urban flooding and extreme droughts [1,2], causing devastating consequences for urban water supply, agricultural production, economic life, and ecosystems [3,4]. Among these disasters, drought events have become increasingly frequent and severe worldwide due to intensified industrialization and urbanization [5,6]. These trends pose mounting risks to environmental sustainability and food security. China is among the countries most frequently and severely affected by drought disasters [7]. According to the Ministry of Water Resources of the People’s Republic of China, droughts struck 24 provinces across the country in 2022, resulting in a loss of 5.744 billion kilograms of grain and CNY 14.944 billion in economic crops [8].

As a typical hydroclimatic disturbance, droughts increase pressure on water supplies [9], potentially resulting in hazards. As the basis of the Earth’s ecosystem, vegetation not only sustains biodiversity but also plays a crucial role in regulating the atmosphere, water cycle, and carbon cycle [10]. When vegetation is exposed to drought hazards, the exchange of mass and energy with the atmosphere becomes constrained [11,12,13,14,15], thereby limiting its productivity [16,17]. However, plants can also acclimate to drought through short-term physiological adjustments such as enhanced water-use efficiency during photosynthesis and respiration [18]. Vegetation responses to drought are not uniform but are strongly modulated by the prevailing climatic conditions [19]. Climate zones, characterized by distinct combinations of temperature, precipitation, and seasonality provide an effective framework for capturing spatial heterogeneity in drought exposure, ecological structure, and recovery capacity [20]. Studies have shown that vegetation in sub-humid regions tends to exhibit greater resilience to drought compared to that in arid or semi-arid areas, owing to more favorable water availability and ecosystem buffering capacity [21,22]. Therefore, assessing drought resilience (DR) across climate zones not only reflects environmental gradients but also allows for the identification of spatially differentiated mechanisms driving vegetation response and recovery under climate stress.

Resilience can be understood as the ability of an ecosystem to recover to its initial state after a disturbance [23,24,25], and ecosystems with higher resilience have better stability and recovery capacity under drought disturbances [26]. Previous studies have employed a wide range of quantitative methods to evaluate vegetation resilience to drought [27]. Li et al. [28] used standardized precipitation evapotranspiration index (SPEI) to assess ecosystem drought parameters to quantify the resistance and resilience of the ecosystem. The results indicate that forests and woody savannas exhibited greater resistance but lower resilience. Sun et al. [29] quantified global grid-scale vegetation resilience and resistance using leaf area index (LAI) and historical climate data, which provides valuable insights into the dynamic ecological mechanisms of vegetation and contributes to the improvement of ecosystem models. Additionally, the response of vegetation to drought is often analyzed by examining the correlation between vegetation indices and drought indices [30,31]. Liu et al. [32] analyzed the relationship between gross primary productivity (GPP), water use efficiency (WUE), and the anomalous SPEI during the vegetation growing season, successfully identifying periods of vegetation suppression and recovery. The results further revealed that in southwestern China, restored vegetation exhibited greater resistance to drought compared to natural vegetation. These efforts have deepened understanding of vegetation–drought interactions, but most focus on individual indicators or specific processes, lacking a systematic framework for resilience assessment.

Exploring the factors driving changes in vegetation ecosystem DR is beneficial for vegetation management department to formulate policies and measures. Many studies have reported that natural factors, such as climate change and soil situation, etc., impacted vegetation resilience significantly, which might lead to further exacerbation of drought conditions and degradation of vegetation cover [33,34,35]. The heightened frequency of extreme precipitation and thermal events significantly amplifies the intrinsic vulnerability of vegetation to drought stress, and drought resilience increases with elevation and decreases as variability in soil moisture increases [36]. At the same time, land use changes and frequent human activities caused by rapid urbanization will decrease urban ecological resilience [37,38]. In addition, anthropogenic factors also influence drought resilience. Intensified human activity, including social production and residential life, has exerted a significant influence on the integrity and dynamics of vegetation ecosystems. Meanwhile, these anthropogenic influences, when interacting with natural factors, contribute to alterations in DR.

To identify the primary factors driving changes in DR, previous studies have extensively used various mathematical analysis models, including Conditional Permutation Importance [39], Partial Least Squares Regression [33], and Spatial Autocorrelation [40]. Although these models are excellent and sensitive to outliers, they do not fully consider the nonlinear relationship between predictors and predicted variables [41]. In contrast, machine learning methods are becoming increasingly popular due to their greater flexibility [42]. Most of the methods do not require a large number of assumptions and are beneficial for handling patterns in large datasets, especially in complex nonlinear situations [43]. Tree-based machine learning models, including Random Forest, Decision Trees, and Gradient Boosting Trees, are popular nonlinear prediction models [44]. The LightGBM model is well-suited to address contemporary data-intensive challenges in ecological research. Wang et al. [45] applied the LightGBM–SHAP framework to assess environmental drivers of vegetation indices under drought in southern China, identifying vapor pressure deficit as the primary limiting factor.

For drought managers, there is a lack of trust and interpretability in using machine learning models all the world [46]. Currently, most research focuses on the nonlinear relationships of single or multiple factors but does not consider the interactions between these factors [47]. Therefore, in our study, we employed the Shapley Additive exPlanations (SHAP) algorithm to estimate the main effects and interactions of different driving factors, which is a new application in explaining environmental issues [48]. SHAP can explain machine learning models based on game theory [49,50], centered on the fair allocation of contributions by each player towards a common goal [51]. Due to its powerful computational capabilities, the SHAP algorithm is widely used in complex machine learning settings to quantify the contribution of each variable within a model as its powerful computational capabilities [52].

Here, based on the existing risk framework proposed by Ge et al. [33], we developed an assessment framework for drought resilience by quantifying three independent components, hazard, adaptability and exposure, to evaluate DR in different climate zones of China [35]. This framework provides a comprehensive, process-oriented, and multidimensional assessment of resilience. Furthermore, to interpret the driving factors of DR, we employed a Bayesian-optimized LightGBM model combined with the SHAP algorithm, which enables both high predictive performance and interpretability. The primary objectives of this study are to (1) develop a risk-based comprehensive resilience assessment framework to quantify DR, considering hazard, exposure, and adaptability; (2) analyze the spatiotemporal patterns of vegetation DR across different climate zones; and (3) explore the interactions among driving factors that affect DR, providing insights for management strategies.

2. Materials and Methods

2.1. Study Area

China, with its vast territory and diverse climate types, spans five major climate zones according to the Köppen–Geiger classification [53]: Tropical (A), Arid (B), Temperate (C), Cold (D), and Polar (E), as shown in Figure 1. These zones differ in temperature, precipitation, and vegetation structure, which may influence drought occurrence and ecosystem response. Therefore, using primary climate zones provides a consistent and ecologically relevant basis for regional comparison of vegetation drought resilience. The southeastern coastal regions, which belong mainly to the tropical and temperate zones, are characterized by high vegetation productivity and complex land–atmosphere interactions, making them ecologically sensitive to changes in drought patterns.

2.2. Data Sources

2.2.1. Resilience Assessment Datasets

The Standardized Precipitation Evapotranspiration Index (SPEI) is an important indicator for assessing drought under the context of global warming [54]. Compared to other drought indices such as Standardized Precipitation Index (SPI) and Palmer Drought Severity Index (PDSI), the main advantage of SPEI is its ability to identify the combined changes in evapotranspiration and temperature, and it can be represented at different time scales; hence, it is widely used in drought monitoring and forecasting [55]. In this study, we used the high-resolution HSPEI (High-resolution Standardized Precipitation Evapotranspiration Index) dataset (2001–2020) developed by Xia et al. [56], which combines station-based SPEI with remote sensing inputs (GPM, MODIS LST, ERA5, and SRTM) using a Random Forest regression model. Compared with the widely used SPEIbase v2.6, HSPEI improves spatial resolution to 1 km and enhances spatial continuity, making it more suitable for drought analysis at local and regional scales.

For vegetation-related datasets, we utilized long-term satellite observations from MODIS, acquired via the LP DAAC. These data were downloaded, systematically archived, and resampled to a uniform spatial resolution of 1 km to ensure consistency across datasets [14,27,28]. The datasets encompass key parameters, including Gross Primary Productivity (GPP), the Normalized Difference Vegetation Index (NDVI), Vegetation Continuous Fields (VCF), Evapotranspiration (ET), and Land Surface Temperature (LST). These indicators have been widely used as ecological proxies, allowing for consistent, spatially explicit assessments of drought resilience across large regions [14,27,28]. All datasets presented in

Table 1. List of DR assessment datasets with their description, resolution, and source.

Dataset	Units	Description	Resolution	Source
SPEI	/	Standardized Precipitation Evapotranspiration Index	1 km monthly	[56]
GPP	kgC/m2	Gross Primary Productivity	500 m 8-day	MOD17A2H.006
NDVI	/	Normalized Difference Vegetation Index	1 km monthly	MOD13A3.061
VCF	%	Vegetation Continuous Fields	250 m yearly	MOD44B.006
ET	kg/m2	Evapotranspiration	500 m 8-day	MOD16A2.006
LST	K	Land Surface Temperature	1 km monthly	MOD21C3.061

span the temporal range from 2001 to 2020. These data encompass all climate zone types, with MODIS data offering the distinct advantages of continuity, long-term availability, and high resolution. Moreover, the globally harmonized processing algorithm minimizes systematic errors across regions, providing a robust foundation for reliable cross-regional comparisons.

2.2.2. Driving Factor Datasets

To examine the driving effects and interactions of various factors, we collected data for the year 2020 on annual precipitation (PRE), temperature (TEM), gross domestic product (GDP), anthropogenic carbon dioxide emissions (ODIAC), potential evapotranspiration (PET), soil moisture content index (SMCI), vapor pressure deficit (VPD), relative humidity (RH), wind speed (WIN), and sunlight hours (SUN). For variables originally available at monthly resolution, annual values were calculated by averaging monthly data. All datasets used in the driving factor analysis of DR in 2020 (Table 2) were resampled and spatially aligned at a resolution of 1 km to ensure grid-level consistency across variables, and analyzed using the LightGBM model.

The PRE, TEM, and PET data obtained from Peng et al. [57,58,59] have been validated using data from meteorological stations in China and are widely applied in ecological assessments [60]. Gridded GDP data [61] exhibit high accuracy in terms of precision, making them highly significant for model analysis. The ODIAC product [62] is currently widely applied in various research fields, including carbon emissions and carbon dioxide flux inversion. The SMCI dataset, derived from field observations, is characterized by its multi-scale, high-resolution, and high-quality features [63]. It is widely utilized in hydrological, meteorological, and ecological analyses, as well as in modeling applications. The VPD and RH datasets [64], which represent humidity, exhibit root mean square error and mean absolute error values within reasonable ranges, indicating high accuracy and high resolution. The SUN and WIN datasets are interpolated from daily observations of meteorological element stations in China [65].

Table 2. List of driving factor analysis datasets with their description, resolution, and source.

Dataset	Units	Description	Resolution	Source
PRE	0.1 mm	Precipitation	1 km monthly	[57]
TEM	°C	Temperature	1 km monthly	[58]
GDP	USD	Gross Domestic Product	1 km yearly	[61]
ODIAC	Pg	Anthropogenic Carbon Dioxide Emissions	1 km yearly	[62]
PET	mm	Potential Evapotranspiration	1 km monthly	[59]
SMCI	10−3 m3/m3	Soil Moisture Content Index	1 km daily	[63]
VPD	hPa	Vapor Pressure Deficit	1 km monthly	[64]
RH	%	Relative Humidity	1 km monthly
WIN	m/s	Wind Speed	1 km yearly	[65]
SUN	h	Sunlight Hours	1 km yearly

Table 2. List of driving factor analysis datasets with their description, resolution, and source.

Dataset	Units	Description	Resolution	Source
PRE	0.1 mm	Precipitation	1 km monthly	[57]
TEM	°C	Temperature	1 km monthly	[58]
GDP	USD	Gross Domestic Product	1 km yearly	[61]
ODIAC	Pg	Anthropogenic Carbon Dioxide Emissions	1 km yearly	[62]
PET	mm	Potential Evapotranspiration	1 km monthly	[59]
SMCI	10−3 m3/m3	Soil Moisture Content Index	1 km daily	[63]
VPD	hPa	Vapor Pressure Deficit	1 km monthly	[64]
RH	%	Relative Humidity	1 km monthly
WIN	m/s	Wind Speed	1 km yearly	[65]
SUN	h	Sunlight Hours	1 km yearly

2.3. Methods

The method framework mainly consists of two parts, assessment and driving factor analysis, as shown in Figure 2. Drought Hazard (DH), Drought Exposure (DE), and Drought Adaptability (DA) are quantified to assess DR. Additionally, we used optimized machine learning methods and explanatory algorithms to identify the contribution of driving factors.

2.3.1. Assessment of DR

A framework for assessing DR has been established (Figure 2), determined by three independent components. Drought resilience of vegetation can be conceptualized through three equally important dimensions: hazard, exposure, and adaptability. A multiplicative model was used to prevent compensatory effects, ensuring that a low value in any component substantially reduces overall resilience [33,35].

$$DR=1/20\times {{\displaystyle \sum }}_{2001}^{2020}D{H}_i\times D{E}_i\times D{A}_i$$

(1)

where $D{H}_i$, $D{E}_i$, $D{A}_i$ represent drought hazard, drought exposure, and drought adaptability, respectively. In this methodological approach, we focus particularly on the resilience of vegetation ecosystem to meteorological drought.

(1) Drought Hazard ($DH$)

Typically, hazards are understood as the likelihood of drought occurrence, interpreted through frequency, intensity, and spatiotemporal scope [35]. The integration of probability density functions and drought indices [66] allows for a comprehensive consideration of disturbances during droughts. The drought index quantifies the intensity of the drought, while the probability density function characterizes the frequency of occurrence of that specific drought intensity. Thus, $DH$ is determined by the intensity and frequency of drought:

$$D{H}_i={{\displaystyle \int }}_{-\infty }^{-0.5}SPEI\times f\left(SPEI\right)$$

(2)

where $D{H}_i$ represents the vegetation drought hazard at the grid point in year $i$, $SPEI$ indicates the drought intensity at a raster point, as reflected by its specific value, and $f\left(SPEI\right)$ denotes the probability of a specific SPEI value occurring at that raster point in the given year.

An SPEI value of less than −0.5 signifies that precipitation and evapotranspiration conditions have deviated substantially from normal, indicating a certain degree of drought [35,67]. This threshold effectively captures the intensity of mild to moderate droughts, providing a standardized basis for drought classification [68]. Furthermore, SPEI-12, which reflects annual-scale drought conditions, is widely employed to assess drought frequency and intensity, providing a comprehensive evaluation of meteorological drought hazard [35].

(2) Drought Exposure ($DE$)

Exposure relates to the number of entities affected by drought, including human populations and vegetation [69]. Here, drought exposure is determined by the health condition of vegetation ($VHI$) and vegetation density ($VD$):

$$D{E}_i=1/20\times {{\displaystyle \sum }}_{2001}^{2020}V{D}_i\times VH{I}_i$$

(3)

where $V{D}_i$ represents the vegetation density in year $i$, and $VH{I}_i$ represents the Vegetation Health Index in year $i$. $D{E}_i$ represents the vegetation drought exposure in year $i$, with higher $DE$ values indicating greater exposure resilience.

The vegetation density of woody and herbaceous plants can be quantified using tree $VCF$ and non-tree $VCF$, respectively, which are then integrated to represent the overall vegetation density status:

$$V{D}_{m,i}=VC{F}_{tree,i}+VC{F}_{non-tree,i}$$

(4)

where $V{D}_{m,i}$ represents the vegetation density in month $m$ of year $i$; $VC{F}_{tree,i}$ and $VC{F}_{non-tree,i}$ represent the continuous vegetation fields of woody and non-woody plants in year $i$.

Furthermore, in calculating $VHI$, we refer to the method of [70] by calculating vegetation condition index (VCI) and temperature condition index (TCI):

$$VC{I}_{i,m}={\displaystyle \frac{NDV{I}_{i,m}-NDV{I}_{min}}{NDV{I}_{max}-NDV{I}_{min}}}$$

(5)

$$TC{I}_{i,m}={\displaystyle \frac{LS{T}_{i,m}-LS{T}_{min}}{LS{T}_{max}-LS{T}_{min}}}$$

(6)

$$VH{I}_{i,m}=\alpha \times VC{I}_{i,m}+\left(1-\alpha \right)\times TC{I}_{i,m}$$

(7)

where $i$ and $m$, respectively, represent the $i^{th}$ year and the $m^{th}$ month; $\alpha$ is a weight factor, generally fixed at 0.5 in practical applications [71].

(3) Drought Adaptability ($DA$)

Adaptability refers to the degree or time required for vegetation recovery [21], reflecting the vegetation’s capacity to adapt to drought. $DA$ can be characterized by its drought sensitivity $\left(DS\right)$ and ecohydrological resilience ($ER$) [33,72].

$$D{A}_i=D{S}_i\times E{R}_i$$

(8)

A higher $D{A}_i$ value indicates that the vegetation is less susceptible to drought stress, and vice versa.

Here, we use the drought resistance traits of vegetation [73] to characterize $DS$ of the vegetation ecosystem. Pearson correlation analysis is performed between $NDVI$ and $SPEI$ at different time scales as follows:

$$\begin{gathered} R_{i,j}=cor \left(NDV{I}_{i,m}, SPE{I}_{i,m,j}\right), 2001 \le i \le 2020, 1\le m\le 12, \\ j=1, 3, 6, 9, 12, 24 \end{gathered}$$

(9)

$$D{S}_i=max\left( R_{i,j} \right)$$

(10)

where $cor$ is the Pearson correlation coefficient; $i$ represents the year, ranging from 2001 to 2020; $m$ represents the month, ranging from 1 to 12; $j$ is the time scale of the drought index $SPEI$; $NDV{I}_{i,m}$ represents the $NDVI$ index for month $m$ of year $i$; $SPE{I}_{i,m,j}$ represents the $SPEI$ index for month $m$ of year $i$ at a $j$ month scale.

Furthermore, we use $ER$ [74] to characterize the dynamic changes in vegetation response to drought. To quantify the terrestrial ecosystem’s response to drought climate, ecosystem water use efficiency ($eWUE$) is widely applied in the coupling of water and carbon cycles [75], defined as the carbon uptake rate per unit of water loss:

$$eWUE={\displaystyle \frac{GPP}{ET}}$$

(11)

where $GPP$ represents the total primary productivity of vegetation, and $ET$ represents total evapotranspiration.

$ER$ is measured using a dimensionless resilience index, defined by Sharma and Goyal [76]. The water use efficiency of vegetation under adverse conditions was quantified as the ratio of $eWUE$ during the driest month to the mean monthly $eWUE$:

$$ER={\displaystyle \frac{eWU{E}_d}{eWU{E}_m}}$$

(12)

where $eWU{E}_d$ represents the water use efficiency in the driest month, and $eWU{E}_m$ represents the annual average water use efficiency. An $ER$ greater than or equal to 1 indicates that the ecosystem demonstrates resilience by enhancing eWUE despite the impact of drought [76]. Conversely, an $ER$ value below 0.8 signifies a severe deficiency in the vegetation resilience.

(4) Normalization

To ensure data consistency, the $DH$, $DS$, $ER$, $DA$, $DE$, and $DR$ of each year are normalized to dimensionless values between 0 and 1 for assessment:

$$t\prime ={\displaystyle \frac{t-t_{min}}{t_{max}-t_{min}}}$$

(13)

where $t\prime$ represents the normalized value of the variable, while $t_{max}$ and $t_{min}$ denote the maximum and minimum values of the variable across all grid points within the study region and the specified timeframe, respectively.

2.3.2. Analysis of DR Spatiotemporal Patterns

(1) Mann–Kendall (M-K) trend test

The M-K trend test is less sensitive to outliers than parametric statistics and usually does not require normality or linearity assumptions [77], and is often used to evaluate hydrometeorological, drought, and flood time series [78]. The method for calculating the Mann–Kendall test is as follows:

$$S={{\displaystyle \sum }}_{\theta =1}^{n-1}{{\displaystyle \sum }}_{i=\theta +1}^n\mathrm{sgn}\left(x_i-x_{\theta }\right)$$

(14)

where $x$ is $DR$, $\theta$ is the serial number of the time series, $n$ is the length of the time series, and $sgn$ is the symbolic function:

$$\mathrm{sgn}\left(x_i-x_{\theta }\right)=\left\{\begin{array}{c}1 x_i-x_{\theta }>0\\ 0 x_i-x_{\theta }=0\\ -1 x_i-x_{\theta }<0\end{array}\right.$$

(15)

When $n\ge 8$, the statistic $S$ roughly follows the normal distribution law, and its mean $ES=0$ without considering the existence of data points with the same value in the time series. The variance $Var\left(S\right)$ is calculated by the following formula:

$$Var\left(S\right)={\displaystyle \frac{n\left(n-1\right)\left(2n+5\right)-{{\displaystyle \sum }}_{i=1}^n{t}_i\left(t_i-1\right)\left(2t_i+5\right)}{18}}$$

(16)

where $t_i$ is the number of time series groups in group $i$.

The statistics $S$ and variance $Var\left(S\right)$ are used to perform the normal approximation Z test:

$$Z=\left\{\begin{array}{c}{\displaystyle \frac{S-1}{\sqrt{Var\left(S\right)}}} S>0\\ 0 S=0\\ {\displaystyle \frac{S+1}{\sqrt{Var\left(S\right)}}} S<0\end{array}\right.$$

(17)

The positive (negative) value of $Z$ indicates the increasing (decreasing) trend, while $\left|Z\right|$ indicates the significance of the time series trend: when $\left|Z\right|$ is greater than or equal to 1.64, 1.96, and 2.58, it indicates that it passes the significance test of 90%, 95%, and 99% confidence, respectively.

(2) Spatial autocorrelation analysis

Spatial autocorrelation analysis can be used to describe the correlation degree and correlation of the same variable between a certain spatial region and adjacent regions [79,80]. To assess the spatial pattern of vegetation drought resilience, we used the local Moran’s I for spatial autocorrelation analysis:

$$\left\{\begin{array}{c}Local Mora{n}^{’}s I=\left(x_i-\overline{X}\right){\displaystyle \frac{{{\displaystyle \sum }}_{j=1,j\ne i}^n{w}_{i,j}\left(x_j-\overline{X}\right)}{{S_i}^2}}\\ \overline{X}={\displaystyle \frac{{{\displaystyle \sum }}_{i=1}^n{x}_i}{n}}\\ {S_i}^2={\displaystyle \frac{{{\displaystyle \sum }}_{j=1,j\ne i}^n\left(x_j-\overline{X}\right)}{n-1}}\end{array}\right.$$

(18)

where $n$ is the number of spatial units, $x_i$ is the observed value at location $i$, $\overline{X}$ is the mean value of all observations, $w_{i,j}$ is the spatial weight matrix, and ${S_i}^2$ is the local variance.

To determine the statistical significance of spatial clustering, we set the p-value threshold at 0.05, as commonly used in spatial analysis. Regions with p < 0.05 were classified as significant clusters (high-high, low-low), while those with p ≥ 0.05 were classified as non-significant [81]. This p-value threshold was consistently applied across all spatial analyses to ensure result comparability.

2.3.3. Analysis of DR Driving Factors

(1) LightGBM Model

LightGBM is a decision-tree-based machine learning method [82] commonly used for regression and classification problems [83]. As an improved method of the XGBoost algorithm [84], LightGBM utilizes a histogram-based decision tree algorithm, transforming weak learners into strong learners, thereby having lower memory consumption, more efficient parallel training, and more accurate results [85,86].

In this study, a LightGBM model was utilized to investigate the influence of each factor on DR. The Python (version 3.8.18) implementation of the LightGBM library is used. The dataset is split into a training set and a test set at an 80% and 20% ratio, respectively, to simulate and predict DR. Additionally, a five-fold cross-validation method and extrapolation validation are used to evaluate the accuracy. In the five-fold cross-validation, the training set is randomly divided into five parts, four of which are used for training and one reserved for validation. Here, we use four metrics to quantify model predictions: R², RMSE (Root Mean Square Error), MSE (Mean Square Error), and MAE (Mean Absolute Error).

(2) Model hyperparameter settings and evolution

A dataset comprising 1,912,707 samples was compiled from the 2020 Chinese vegetation resilience data to build five distinct machine learning models. 80% of the samples were allocated for model training, with the remaining 20% reserved for validation. BayesianOptimization method was utilized from the bayes_opt library (Version 1.4.3) [87] in Python to perform hyperparameter tuning. Optimal hyperparameters for each model were identified through cross-validation using the training data [88,89]. The model performance was then evaluated by recording the MAE, MSE, and RMSE.

Due to numerous hyperparameters in LightGBM [90], traditional grid search and random search methods can be slow or may miss the combination of parameters that maximize results. In contrast, Bayesian optimization is a global optimization algorithm based on Bayesian theorem [91], which can quickly find the optimal solution with fewer function evaluations based on historical experience [92]. The main application scenario for Bayesian optimization is as follows:

$$X^{*}=\mathrm{argmax}f\left(x\right), x \in \left\{x_1,x_2, \cdots , x_n \right\}$$

(19)

where $x$ is a combination of hyperparameters and $f\left(x\right)$ is the objective function of the hyperparameters, with the principle of Bayesian optimization to determine $X$.

Based on the acquisition function, the Bayesian optimization algorithm can further obtain the next sampling point within the space of hyperparameters. In this study, the Expected Improvement is used as the sampling function, represented as below:

$$\left\{\begin{array}{c}\alpha \left(x\right)=\left[\mu \left(x\right)-q^{+}\right]\mathsf{\varnothing }\left(Z\right)+\sigma \left(x\right)\phi \left(z\right)\\ x_{t+1}=\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{a}\mathrm{x}f\left(x\right)\end{array}\right.$$

(20)

where $\alpha \left(x\right)$ is the objective function; $\mu \left(x\right)$ is the mean; $\sigma \left(x\right)$ is the standard deviation; $q^{+}$ is the current maximum value of the objective function; $\mathsf{\varnothing }\left(Z\right)$ is the cumulative distribution function of the Gaussian distribution; $\phi \left(z\right)$ is the probability density function of the Gaussian distribution; $x_{t+1}$ is the optimized set of hyperparameters.

The hyperparameter settings optimized through Bayesian optimization are presented in

Table 3. Detailed hyperparameters values of the LightGBM model.

Model	Hyperparameters	Before Optimization	After Optimization
LightGBM	learning_rate	0.02	0.25
	colsample_bytree	0.9	0.7347
	reg_alpha	0	0
	max_depth	6	19
	min_child_samples	20	32
	min_gain_to_split	0	0
	n_estimators	250	250
	num_leaves	50	242
	subsample	0.8	0.72
	reg_lambda	0	0

, illustrating the key parameter adjustments that led to enhanced model performance.

(3) Model interpretation

Compared to traditional linear statistical models, the LightGBM model offers superior generalization performance and higher accuracy, but it mainly remains a black box mode lack of strong interpretability [48]. Therefore, we have employed SHAP to interpret our model, aiming to determine the contribution of potential driving factors to changes in DR. The SHAP aims to fairly distribute players’ contribution when they collectively achieve a specific outcome. The Shaply value for each feature $X_j$ is given by

$$Shapely\left(X_j\right)={{\displaystyle \sum }}_{S⊑N\setminus \left\{j\right\}}{\displaystyle \frac{k!\left(p-k-1\right)!}{p!}}\left(f\left(S\cup \left\{j\right\}\right)-f\left(S\right)\right)$$

(21)

where $p$ is the total amount of features; $N\setminus \left\{j\right\}$ is a set of all combinations of the features except $X_j$; $S$ is one feature set in $N\setminus \left\{j\right\}$; $f\left(S\cup \left\{j\right\}\right)$ is the model prediction with features in $S$ as well as feature $X_j$; $f\left(S\right)$ is the model prediction with features in $S$. The Shaply value of one feature is its marginal contribution to model prediction averaged all possible models with different combinations of various features.

SHAP contributes to generate locally addictive feature attribution:

$$\widehat{y_i}=sha{p}_0+shap\left(X_{1i}\right)+shap\left(X_{2i}\right)+\dots +shap\left(X_{pi}\right)$$

(22)

where $\widehat{y_i}$ is the model prediction value for the observation $i$; $sha{p}_0=E\left(\widehat{y_i}\right)$ is the mean prediction across all observation; $shap\left(X_{ji}\right)$ refers to the SHAP value of the $j^{th}$ feature for observation $i$ which presents the marginal contribution of the feature to the prediction.

3. Results

3.1. Patterns of DR Independent Components

Consistent with previous studies [33,93], regions with higher DH are concentrated in northwestern China, including Xinjiang and Inner Mongolia, while areas with lower DH are more concentrated in the three northeastern provinces, Sichuan, the Beijing-Tianjin-Hebei region, and the Yangtze River Delta urban agglomeration (Figure 3a). In studies of DH, urban agglomerations tend to exhibit higher concentrations of DH. Significant differences in DH are observed across various climate zones (Figure 3b). The arid climate zone faces the most severe drought threats, followed by polar and continental regions, while tropical and temperate regions experience the lowest DH.

Overall, DE and GPP exhibit similar spatial patterns, with higher prevalence in the southeast and lower prevalence in the northwest (Figure 3c). In southeastern China, where the vegetation is dense and species-rich, exposure resilience is relatively high. Conversely, vegetation in the northwest region showed a higher likelihood of being exposed to drought events. Vegetation in arid and polar climate zones is more vulnerable to drought threats, exhibiting lower exposure resilience, while vegetation in tropical and temperate regions shows good exposure resilience (Figure 3d).

Here, we quantify DA as a composite of DS and ER, where higher DA values reflect stronger short-term physiological recoverability to drought. Here, adaptability is determined by both DS and ER, with higher DA values indicating greater vegetation adaptability to drought. Vegetation on the Tibetan Plateau area exhibits better DA, whereas the northern regions, compared to the south, exhibit greater variability and generally lower adaptability (Figure 3e). Although the overall average DA levels across the five climate zones do not differ significantly, there is substantial variability in vegetation recoverability within arid and polar climate zones (Figure 3f).

Vegetation in the southern regions is more sensitive to changes in drought (Figure 4a). In terms of climate zones, vegetation in arid and polar climates shows stronger DS (Figure 4b).

ER reflects the capacity of vegetation to couple water and carbon [74], with average levels observed in the southern regions, while better water–carbon coupling abilities are found in central and northeastern China (Figure 4c). From a climatic perspective, vegetation in tropical and temperate zones exhibits concentrated and stable water–carbon coupling capacities (Figure 4d).

3.2. Spatiotemporal Patterns of DR

As illustrated in Figure 5a, the DR Level is relatively high across most regions of China (above the average value of 0.448), with pronounced concentrations in the southern coastal areas, while the northwestern inland regions exhibit comparatively lower levels. As shown in Figure 5b, the DR is strongest in tropical and temperate climate zones, but relatively weak in arid and polar regions.

We analyzed the temporal trends of DR using the M-K method (Figure 6a,b). Overall, between 2001 and 2020, DR in China shows signs of improvement, with approximately 45.3% of areas exhibiting enhancement, and 36.4% potentially facing continued deterioration in vegetation conditions. Specifically, 2.7% of the regions exhibit a significant increasing trend, while 1.1% of the regions display a significant decreasing trend. In most cases, the trend in DR changes is opposite to its baseline values. Regions in the west, which initially had high resilience, have shown a decrease, while areas with lower resilience, such as the southern coast and the northeast, have exhibited gradual improvements (Figure 6a). Notably, the Shaanxi and Henan regions, which already possess poor resilience, continue to deteriorate, highlighting the need for targeted planning by relevant management authorities in these vulnerable areas (Figure 6b).

Trends in DR vary across different climate zones (Figure 6c). In temperate regions, there appears to be no significant trend of increase or decrease in DR. Cold climate regions, however, show a significant increasing trend (p < 0.05). DR in the tropical zones also appears to be improving. Conversely, a worsening trend is observed in the arid and polar regions.

In China, the spatial distribution of DR exhibits a clustered pattern (Figure 6d). At the prefecture-level, high-value clusters are found in the Xinjiang and Heilongjiang regions in the northwest, while concentrated low-value clusters are observed in southern Yunnan, Sichuan, the Yangtze River Delta, and parts of Henan and Fujian.

3.3. Modeling and Analysis of Driving Factors

To quantify the impact of natural and anthropogenic factors on the vegetation drought resilience of China, we employed machine learning models to study the driving mechanisms of vegetation Drought Resilience. After normalizing vegetation drought resilience and the potential driving factors, the model’s accuracy was further enhanced through training. As shown in Figure S1, the performance of the LightGBM model optimized using Bayesian optimization (MAE: 0.015, R²: 0.792, RMSE: 0.026, MSE: 0.001) was superior to that of the unoptimized original LightGBM model (MAE: 0.020, R²: 0.611, RMSE: 0.032, MSE: 0.001). Due to the automatic nature of the LightGBM model, which does not require assumptions about the model’s specifications and parameters, the accuracy is satisfactory and can be used to predict the vegetation drought resilience of China.

Additionally, identifying the contribution ratio of each factor during the DR prediction process is also essential. SHAP values are a novel method for model explanation, which can be combined with different machine learning models for model interpretation [42,94], fulfilling the needs for flexible modeling and visualizing complex geographic phenomena. Overall, all climatic factors significantly impact vegetation drought resilience. The variable importance of SHAP values estimated by the LightGBM model is shown in Figure 7a, with the drivers of Chinese vegetation DR ranked from most to least impactful as TEM (+0.0122) > SUN (+0.0098) ≥ PET (+0.0098) > PRE (+0.0061) > RH (+0.0041) > SMCI (+0.0030) > WIN (+0.0029) > VPD (+0.0021) > GDP (+0.0015) > ODIAC (+0.0009). As the input variables were normalized before training, the resulting SHAP values are also relatively small in magnitude.

Figure 7b depicts a summary plot of SHAP estimates for all major features and their interaction effects, arranged from top to bottom by importance, with the y-axis representing feature values and the x-axis displaying SHAP values for a given instance. Feature values are indicated by color, with red representing high values and blue representing low values. Overlapping points are jittered along the y-axis [95] to better understand the distribution of SHAP values for each feature. Thanks to the global explanation provided by Shapley, we can discern how these driving factors contribute positively or negatively to changes in DR. It is evident that VPD shows particularly significant variability in SHAP values, with the distribution trend of higher feature values (darker red) aligning with higher negative SHAP values, indicating VPD’s negative feedback on DR. Similarly, the distribution trend of lower feature values (darker blue) aligns with higher positive SHAP values, also proving VPD’s negative feedback on DR. Thus, the higher the feature value of VPD, the more significant the importance of the SHAP negative value. VPD has a higher marginal negative effect on the probability prediction of DR, while TEM, PET, SUN, and RH exhibit similar patterns. Among all the driving factors, only VPD displays SHAP values symmetrically distributed around zero, indicating negative influences on DR, while the others show predominantly unidirectional effects.

We focused on pairwise interaction effects as provided by SHAP’s interaction values, which offer interpretable insights into nonlinear dependencies between variables while maintaining computational feasibility. Figure 8a–j illustrate the marginal effects and the most influential pairwise interactions of each driving factor on DR as captured by the LightGBM model. The vertical dispersion of each single factor indicates the influence of another factor on that factor, hence we color it according to the degree of influence of another factor. The x-axis represents an individual independent driving factor, the y-axis represents the corresponding SHAP value for that factor, and the gradient of colors (red for high values, blue for low values) represents the interactive effects brought by another driving factor. No single linear trend exists between all driving factors and DR, which further validates the conclusions depicted in Figure 7b.

Consistent with previous findings [33], temperature (TEM), sunlight hours (SUN), and precipitation (PRE) are identified as the most significant driving factors. The most critical driver, TEM, shows a nonlinear relationship with DR, positively correlated at a lower temperature and contributing positively when sunlight hours get longer. However, as temperature rises, DR shows negative feedback, and longer sunlight hours negatively affect DR (Figure 8a). Shorter sunlight hours leads to better vegetation DR (Figure 8b), and increased rainfall positively affects DR. As the duration of sunlight increases, both increased rainfall and longer sunlight hours can negatively impact drought resilience.

Global warming has increased evapotranspiration, which in turn exacerbates drought conditions [96]. In our study, potential evapotranspiration (PET) significantly impacts the DR of Chinese vegetation (Figure 8c). Lower evapotranspiration conditions enhance vegetation’s drought resistance. Nevertheless, increasing precipitation does not exhibit a uniform suppressive trend with respect to DR. (Figure 8d). Both low temperatures with scant rainfall and high temperatures with abundant rainfall help cultivate vegetation DR. However, when PRE ranges between 0.12 and 0.65 in combination with high temperatures, DR becomes unstable and continues to decline.

RH and VPD, as different indicators of atmospheric moisture, show varied impacts on the dynamic changes [97] in vegetation DR. Lower relative humidity paired with reduced temperatures can foster the development of DR (Figure 8e). Within the 0.5–0.9 RH range, high temperatures significantly reduce DR, resulting in weaker ecosystem resilience. When VPD falls within the 0.2–0.4 range, the intensified effect of rainfall further exacerbates VPD’s negative impact on DR, with the maximum negative contribution reaching −0.06.

SMCI determines the amount of water that plant roots can extract. In drought conditions, vegetation can alleviate drought risks by absorbing soil moisture through their roots, hence a conducive soil moisture environment leads to better vegetation DR (Figure 8f). Wind speed also impacts vegetation DR to some extent (Figure 8g). As wind speed increases, vegetation’s resilience to drought slightly deteriorates.

We consider GDP and ODIAC as anthropogenic factors, but these two factors seem to have different mechanisms of action. When GDP exceeds a threshold of 0.2 under moderate PRE, GDP consistently contributes positively to DR (Figure 8i). However, for ODIAC values above 0.4—occurring alongside elevated potential evapotranspiration—ODIAC increasingly exerts a negative effect on DR (Figure 8j).

4. Discussion

4.1. Drivers’ Interaction on DR

Previous research [33,35,97] has largely concentrated on the effects of individual factors on vegetation’s drought resilience, often overlooking the impact of their interactions. Employing a multifactorial approach allows for a deeper understanding of the underlying mechanisms of vegetation drought resilience at the national scale.

High temperature and prolonged sunshine duration can jointly intensify plant water loss by increasing evaporative demand and thermal load, especially under drought [98,99]. This compound stress reduces vegetation resilience and has been shown to impair productivity and recovery capacity in various ecosystems. In our study, areas experiencing high potential evapotranspiration consistently show poor vegetation resilience, regardless of variations in sunlight hours. Increases in precipitation in some areas are often due to extreme precipitation events [100]. Thus, rapidly increasing extreme precipitation events indicate greater climate instability, which may weaken vegetation’s resistance to climate change, thereby reducing its drought resilience.

Strong winds not only potentially cause mechanical damage to vegetation but also more easily carry away moisture from leaf surfaces [101,102]. Soil moisture also plays a crucial role in vegetation productivity and is strongly influenced by precipitation [103]. In this study, we also found that shorter sunlight hours resulted in better soil moisture conditions, thereby facilitating vegetation’s ability to adapt to drought.

As GDP increases and economic levels advance, more funds can be allocated to the construction of city parks and protective forests, enhancing green protection, and consequently increasing vegetation DR [40]. Conversely, as human activities increase carbon dioxide emissions from the burning of fossil fuels, vegetation DR tends to decline. Excessive carbon dioxide emissions exacerbate the greenhouse effect, and as temperatures rise, soil organic carbon (SOC) reserves decrease, leading to further carbon dioxide emissions into the atmosphere [104]. This process forms a positive feedback loop where drought and the greenhouse effect mutually reinforce each other [105]. Rapid changes in extreme climate conditions may result in poor climate change stability in vegetation, thereby reducing its drought resilience.

4.2. Spatial Differentiation and Regional Drivers of DR

Drought is a severe and widespread natural disaster that affects nearly all regions [106,107]. In recent years, drought frequency has notably increased in temperate and continental climate zones. With the intensification of global warming, increasing attention has been directed toward vegetation resilience to climate change [108,109]. However, DR across different climate zones has received relatively little attention. Therefore, establishing a quantitative assessment of DR is crucial for effective vegetation management.

Vegetation in warmer tropical and temperate regions tends to exhibit higher resilience, and vegetation exposed to abundant sunlight is often in better condition. Regions with abundant precipitation and correspondingly high DR are predominantly located in economically advanced eastern coastal cities. This co-occurrence likely reflects that strong economic capacity and well-developed infrastructure help vegetation to recover from drought. In arid climate zones, high potential evapotranspiration coupled with sparse vegetation cover limits the buffering capacity of ecosystems, resulting in persistently low drought resilience. Vegetation in these regions often lacks sufficient biomass and root depth to withstand prolonged water deficits, making recovery from drought events particularly challenging [110]. The cold zone presents the highest internal variability, possibly due to the coexistence of multiple vegetation types such as boreal forests, grasslands, and alpine meadows, each responding differently to drought [111]. In polar regions, vegetation drought resilience remains low despite increasing moisture from glacial melt, due to extreme cold, short growing seasons, and shallow root systems constrained by permafrost [22].

Historically drought-prone areas are likely to remain vulnerable in the future as well, with an expansion of arid zones. In polar regions, glacial melt is increasing due to global warming [112], improving moisture availability. However, vegetation still needs to adapt, resulting in a worsening trend in DR. Conversely, the drought resilience of vegetation along the eastern seaboard has shown a significant upward trend over the past 20 years, driven by industrial progress and increased precipitation [112]. Moreover, as the ice melts earlier in the northeast due to the warming climate, the fertile and moist black soil also creates a more favorable environment for vegetation growth [113], contributing to the DR improvement. In contrast, the western interior, where DR tends to decrease, has been more severely impacted by global warming, with rising temperatures and increased evaporation rates [114].

Vegetation in areas prone to high-intensity droughts often lacks resilience [115,116]. In the southwestern region of China (Sichuan and Yunnan), despite good vegetation coverage and exposure resilience, the DR is suboptimal due to the high risk of drought in the area. Moreover, the vegetation in karst areas exhibits lower resilience compared with the vegetation in the non-karst areas [32]. The unique geography of karst regions puts pressure on vegetation resilience due to the thin soil layer and limited water storage capacity [117,118]. In contrast, areas like the Yangtze River Delta urban agglomerations, despite also facing severe droughts, have improved their adaptability through economic development and financial investments, enabling better adaptation to climate change [40].

4.3. Limitations

It is acknowledged that within the same climate zone, different vegetation types may exhibit varied responses to drought. As this study focuses on climate-zone-level patterns, intra-zone ecological heterogeneity was not explicitly addressed, which may influence the interpretation of DR values in mixed-vegetation areas. Future studies could integrate land cover classification to improve the ecological specificity of DR assessments. To ensure methodological consistency at a national scale, we adopted equal weights for hazard, exposure, and adaptability in constructing the drought resilience. While this approach facilitates more unified comparisons across regions, it may overlook regional variations in the relative importance of each component. Future research could explore applying differential weighting within the hazard–exposure–adaptability framework guided by vegetation performance indicators to better align resilience assessments with local ecological conditions.

In this study, SPEI lags of 1, 3, 6, 9, 12, and 24 months were selected, providing a foundation that could be further expanded in future studies to enhance the calculation of DS. Additionally, it is possible to examine the dynamic impacts of frequent human activities on drought resilience, including changes in land use, urban scale, industrial structure, and economic losses caused by drought, which were not included in this study due to a lack of data. Future work could adopt a more diverse and comprehensive perspective to provide valuable insights for vegetation management.

While our study utilized satellite-based indicators to characterize vegetation resilience, future research could benefit from integrating species-level diversity metrics and field-based structural attributes to further enhance ecological resolution. Although all datasets were resampled to a 1 km resolution for spatial consistency, inherent differences between remote sensing and station-based data may still introduce uncertainties. Future research could explore advanced scaling or uncertainty assessment methods to better address such issues.

5. Conclusions

Combining multi-source satellite observations and the meteorological station data, this study developed a resilience assessment system based on the hazard–exposure–adaptability framework to evaluate the spatiotemporal patterns of DR across different climate zones, while also examining the interactive effects of key driving factors. The findings reveal that despite generally strong vegetation resilience, regions in northern China exhibit persistently low DR levels. In contrast, southeastern China experienced a consistent improvement in vegetation resilience from 2001 to 2020, whereas the northwestern regions displayed a declining trend. Notably, DR in cold climate zones exhibited a significant upward trend.

Additionally, the LightGBM model, optimized using a Bayesian algorithm, effectively simulated and predicted DR patterns, achieving a notable performance improvement (R² = 0.792; RMSE = 0.026). TEM, SUN, and PET are the three most powerful drivers. SMCI, WIN, and GDP showed similar effects, while the remaining factors exhibited opposing impacts. Furthermore, the interaction analysis between drivers provided deeper insights into the spatial distribution and temporal trends of vegetation drought resilience, emphasizing the complex interplay of climatic and environmental variables.