Quantitative Analysis and Nonlinear Response of Vegetation Dynamic to Driving Factors in Arid and Semi-Arid Regions of China

Abstract

Vegetation dynamics are complexly influenced by multiple factors such as climate, human activities, and topography. In recent years, the frequency, intensity, and diversity of human activities have increased, placing substantial pressure on the growth of vegetation. Arid and semi-arid regions are particularly sensitive to climate change, and climate change and large-scale ecological restoration have led to significant changes in the dynamic of dryland vegetation. However, few studies have explored the nonlinear relationships between these factors and vegetation dynamic. In this study, we integrated trend analysis (using the Mann–Kendall test and Theil–Sen estimation) and machine learning algorithms (XGBoost-SHAP model) based on long time-series remote sensing data from 2001 to 2020 to quantify the nonlinear response patterns and threshold effects of bioclimatic variables, topographic features, soil attributes, and anthropogenic factors on vegetation dynamic. The results revealed the following key findings: (1) The kNDVI in the study area showed an overall significant increasing trend (p < 0.01) during the observation period, of which 26.7% of the area showed a significant increase. (2) The water content index (Bio 23, 19.6%), the change in land use (15.2%), multi-year average precipitation (pre, 15.0%), population density (13.2%), and rainfall seasonality (Bio 15, 10.9%) were the key factors driving the dynamic change of vegetation, with the combined contribution of natural factors amounting to 64.3%. (3) Among the topographic factors, altitude had a more significant effect on vegetation dynamics, with higher altitude regions less likely to experience vegetation greening. Both natural and anthropogenic factors exhibited nonlinear responses and interactive effects, contributing to the observed dynamic trends. This study provides valuable insights into the driving mechanisms behind the condition of vegetation in arid and semi-arid regions of China and, by extension, in other arid regions globally.

1. Introduction

Vegetation, as a natural link between the soil, atmosphere, and water environment, is a key indicator for assessing ecosystem stability and resilience [1,2]. It plays an irreplaceable role in maintaining climate homeostasis by regulating surface energy balance, driving biogeochemical cycles, and participating in key ecological processes such as the global carbon cycle [2,3,4]. However, the global vegetation system is under unprecedented pressure due to the increasing intensity and frequency of human activities [5]. In recent decades, rapid urbanization and industrialization have profoundly reshaped the natural environment, triggering changes in landscape patterns and ecological conditions [6]. As the frequency and intensity of human activities continue to increase, vegetation growth faces significant ecological stress. Especially in arid areas, the fragile ecosystems show high sensitivity to environmental changes, and the frequent occurrence of extreme weather events further aggravates the vulnerability of vegetation; some of the vegetation has died extensively due to the decline in adaptive capacity, resulting in a significant reduction of vegetation cover [7,8,9]. The IPCC’s 6th Assessment Report indicates that vegetation degradation has occurred on about 25% of the global land surface, and the degradation of ecosystems in arid zones is particularly serious. This degradation directly threatens the livelihoods of nearly 2 billion people worldwide and may trigger a series of cascading ecological effects [10]. The complex interactions between natural and anthropogenic factors make analyzing the relative contributions of various drivers an important challenge for current research [11,12]. A deeper understanding of the driving mechanisms of vegetation dynamics change is not only of great scientific significance but also related to the realization of ecological security and the Sustainable Development Goals [13].

In the arid and semi-arid regions of China (ASARs), as an important ecological barrier zone, the dynamic change of its vegetation directly affects the realization of the northern sand control and prevention project and the goal of “carbon neutrality”. With the increasingly urgent demand for nature-based solutions (NBSs) in various countries, it has become imperative to elucidate the intrinsic driving mechanism of vegetation change and provide scientific basis for global ecological governance [14,15]. In the field of vegetation dynamics monitoring, the rapid development of remote sensing technology has promoted the development and application of a series of vegetation indices [16,17]. Early indices, such as the Ratio Vegetation Index (RVI) [18], which is sensitive to atmospheric influence and susceptible to soil background interference, have evolved into more sophisticated indices like the Enhanced Vegetation Index (EVI) [19], which introduces atmospheric correction factors and soil conditioning parameters, and then to the Photochemical Reflectance Index (PRI) [20], which provides a new way to study the efficiency of photosynthesis of vegetation; the vegetation indices have been iteratively updated. The normalized difference vegetation index (NDVI) has become a core index for vegetation monitoring due to its ease of calculation and strong indicativeness, but it is susceptible to saturation in areas of high vegetation cover, is sensitive to soil background, has limited accuracy in areas of low cover, and is difficult to effectively distinguish physiological differences between vegetation types [21,22]. The kernel normalized difference vegetation index (kNDVI) optimally reconstructs the NDVI based on the kernel method in machine learning, which can more effectively capture the complex spectral features of vegetation by mapping the original spectral features into a high-dimensional space [23,24,25]. Empirical studies have demonstrated that the kNDVI exhibits significant advantages in forest ecosystem monitoring and biome assessment, providing a promising new approach for accurately assessing ecosystem productivity [26,27].

Amid global climate change, vegetation dynamics, as a central characterization of climate–ecology interactions, and the analysis of its response mechanisms are at the forefront of current ecological research [28,29]. However, existing studies reveal a significant knowledge gap: most of them focus on single meteorological factors such as temperature and precipitation and do not pay enough attention to bioclimatic variables that can comprehensively characterize climate fluctuations, which have been shown to improve prediction accuracy in species distribution models [30]. Previous studies have shown that the process of vegetation dynamics change is a complex phenomenon influenced by multiple factors. Focusing solely on the contribution of a single factor may not fully uncover its underlying mechanisms [31]. Traditional studies rely on linear analysis methods, which tend to ignore the complex nonlinear relationship between drivers and vegetation dynamics and the synergistic effect of variables, restricting the development of precise management strategies [32]. Machine learning algorithms provide a new paradigm for dealing with nonlinear relationships, but their “black-box” characteristics limit their ecological mechanism analysis value, while the SHAP method, based on game-theoretic Shapley’s value theory, can realize the local mechanism analysis of individual features, which provides a key tool to solve this dilemma [33].

The arid and semi-arid region of China, as a typical representative of this type of climate zone, has a vast territory with a variety of topographies, geographies, and natural landscapes. In the arid and semi-arid regions of China, diverse climatic and topographical conditions have fostered a wide variety of vegetation types, each with unique adaptations to local environmental niches. Deserts, with a relatively large coverage area, dominate the arid western basins by virtue of drought-tolerant species, while grasslands form an east–west gradient distribution on the semi-arid plateaus, and alpine vegetation thrives in areas above 3000 m in mountain ranges like the Tien Shan. Cultivated vegetation clusters in irrigated oases, swamps dot the low-lying wetlands, thickets occupy the transitional zones, and sparse broad-leaved forests remain in the humid eastern regions. As shown in Figure 1f, these vegetation types are categorized into 9 main types (denoted 1–9) and other types (99), collectively shaping the region’s complex vegetation pattern. Due to the scarcity of perennial precipitation, ecological fragility, the reduction in vegetation cover, population growth, and soil erosion caused by warming and drying are becoming more and more prominent, which constrain the regional ecological environment and socioeconomic development [34]. Vegetation dynamics in the region directly influence the local ecological environment evolution and can provide a practical basis and reference paradigm for the study of similar ecosystems around the world. The objectives of this study are (1) to synthesize kNDVI of long time series as an important index to reveal the dynamic change law of vegetation; (2) reveal the relative contributions of natural and human activities to vegetation dynamics change; (3) elucidate the nonlinear relationship between vegetation dynamics change and its driving factors; and (4) develop strategic information for preventing vegetation degradation and effective sustainable management in arid and semi-arid regions of China.

2. Materials and Methods

2.1. Study Area

The arid and semi-arid zone of China is located between 68.07′ E and 126.65′ E, 29.28′N and 52.69′ N, extending 4137 km from east to west and 2411 km from north to south, with a total area of about 3.23 × 10⁶ km², accounting for 33% of China’s total land area (Figure 1). Since the study region lies deep in the interior hinterland of the Asian continent, it is weakly influenced by the monsoon system, resulting in a climate characterized by low precipitation. There is a significant difference in average annual precipitation between the east and the west: about 400 mm in the east and less than 100 mm in the west. These arid climatic conditions are partly responsible for the fact that the region’s population constitutes only 4% of the country’s total population. Geographically, the elevation of the region is high in the west and low in the east, spanning the three levels of China’s topographic step from west to east, with the terrain dominated by basins, mountains, and plateaus, and with significant internal elevation differences, specifically including typical geographic units such as the Tarim Basin, the Junggar Basin, the Tien Shan Mountain Range, the Qilian Mountain Range, the Inner Mongolia Plateau, and the Daxing’anling Mountain Range [35]. The complex interplay of topography and geography makes the regional climate more diverse: the average annual temperature ranges from−33 °C to 17 °C, and the difference between daytime and nighttime temperatures sharply reflects the difference between temperate monsoon climate and temperate continental climate. The diverse climatic and topographical conditions have fostered a wide variety of vegetation types, including deserts, grasslands, alpine vegetation, cultivated vegetation, swamps, thickets, broad-leaved forests, etc. As shown in Figure 1f, these are categorized into 9 main types (denoted 1–9) and other types (99). Ranging from coniferous and broad-leaved forests in moister areas to deserts dominating arid western basins, grasslands forming east–west gradients, alpine vegetation in high mountain ranges, and cultivated vegetation in irrigated oases, each type adapts uniquely to local conditions, collectively shaping the region’s complex vegetation pattern.

In recent years, global climate change has significantly impacted the region. The frequency of droughts has increased notably, and the intensification of human activities, along with frequent environmental changes such as natural disasters, has further disrupted the structure and function of vegetation ecosystems. As a result, the region’s ecological environment now faces more severe challenges.

2.2. Data Sources

The datasets used in this study include climatic data, remote sensing imagery, soil attribute data, topographic data, and relevant human activity data (

Table 1. Data sources.

Data Type	Time	Resolution/Data Format	Data Sources
Climate data	2001–2020	1000 m, Tiff	National Qinghai-Tibet Plateau Science Data Center (https://data.tpdc.ac.cn/, accessed on 2 May 2025)
Remote sensing imagery	2001–2020	500 m, Tiff	Earth Data (https://lpdaac.usgs.gov/, accessed on 4 May 2025)
Soil attribute data	2009	1000 m, Tiff	National Qinghai-Tibet Plateau Science Data Center (https://data.tpdc.ac.cn/, accessed on 4 May 2025)
Topographic data	-	30 m, Tiff	NASA SRTM1 v3.0 (https://www.earthdata.nasa.gov/, accessed on 6 May 2025)
GDP	2000–2020	1000 m, Tiff	Resource and Environmental Science Data Center (https://www.resdc.cn/, accessed on 6 May 2025)
Population density	2001–2023	1000 m, Tiff	LandScan Global (https://landscan.ornl.gov/, accessed on 6 May 2025)
Nighttime lighting data	2000–2020	1000 m, Tiff	Resource and Environmental Science Data Center (https://www.resdc.cn/, accessed on 6 May 2025)

). Through pre-processing, all datasets were spatially standardized to a uniform 1 km resolution grid and cropped using the boundaries of arid and semi-arid zones to ensure consistency in the analysis.

The temperature, precipitation, and evaporation data for this study were obtained from the National Tibetan Plateau Science Data Center (https://data.tpdc.ac.cn/, accessed on 4 May 2025). In addition, our study is based on its published 30-year average 1 km resolution historical and future climate dataset for China. This includes 23 bioclimatic variables with “Biovars”, which are mainly based on the China Surface Climate Standard Values (CSCV) dataset, produced by interpolation using ANUSPLIN software. It includes 23 bioclimatic variables (Biovars) calculated on the basis of 5 monthly valued basic climate variables. Variables with a VIF greater than 10 were considered to have significant multicollinearity and were manually removed. Through this screening process, we finally retained six bioclimatic variables (Figure 2).

Soil properties can determine the distribution of vegetation, in which soil water retention capacity and nutrient effectiveness are closely related to vegetation changes [36,37]. Soil physicochemical parameters, including basic saturation (BS), pH, electrical conductivity (ECE), exchangeable sodium salt (ESP), clay content (CLAY), and organic carbon (OC), were obtained from the National Scientific Data Center for the Tibetan Plateau (https://data.tpdc.ac.cn/, accessed on 4 May 2025), a Chinese soil database based on values of 0–30 cm layers from the World Soil Database (v1.1) (2009), for surface modeling with a spatial resolution of 1 km.

Topographic data were obtained from the Digital Elevation Model (DEM) data of the SRTM dataset (https://www.earthdata.nasa.gov/, accessed on 6 May 2025), with a spatial resolution of 30 m. The study used elevation data (DEM), while slope data were calculated for the study area.

Population density and gross domestic product (GDP) are often recognized as the main potential causes of anthropogenic activity intensity [38]. In this study, gross domestic product (GDP) data, the population density (pop), and the nighttime light index (Ntl) were used to represent anthropogenic factors. The GDP data were obtained from the Resource and Environmental Sciences Data Center (https://www.resdc.cn/, accessed on 6 May 2025), and the population density (pop) data were obtained from the global year-by-year population distribution data LandScan Global (2000–2023) (https://landscan.ornl.gov/, accessed on 6 May 2025). The annual nighttime light data (Ntl) were obtained for the period from 2000 to 2020. Annual nighttime light data (Ntl) from 2000 to 2020 were obtained from the China Data Center for Resource and Environmental Sciences (https://www.resdc.cn/, accessed on 6 May 2025). The DMSP-OLS sensor detects the visible-near-infrared (VNIR) brightness of microlights from the Earth’s surface, representing the nighttime lights of city lights and even of small residential areas and traffic. All data are resampled at a spatial resolution of 1 km [39].

Remote sensing data were obtained from NASA’s MOD09GA v061 surface reflectance product (https://lpdaac.usgs.gov/, accessed on 4 May 2025) at a resolution of 500 m. These multi-band images from 2001 to 2020 were used to calculate the kNDVI.

2.3. Methodology

2.3.1. Calculation of kNDVI

Compared to the traditional normalized difference vegetation index (NDVI), the kNDVI has advantages in terms of saturation resistance and bias resistance [40]. When faced with complex phenological cycles across spatial and temporal scales, this index exhibits stronger noise interference resistance and higher stability. The study used the NASA’s MOD09GA v061 surface reflectance product at a resolution of 500 m to compute the kNDVI for arid and semi-arid regions (ASARs), with a spatial resolution of 500 m. By performing pixel-by-pixel calculations on NDVI data using the Google Earth Engine (GEE) platform, a kNDVI image sequence for the period 2001–2020 was ultimately generated.

$$kNDVI=\mathit{tanh}\left({\left({\displaystyle \frac{NIR-Red}{2\sigma }}\right)}^2\right)$$

(1)

Here, σ is a length scale parameter, which can be adjusted to capture the nonlinear sensitivity characteristics of the $NDVI$ to vegetation density. σ is usually defined as 0.5 (near-infrared band + red light band). Based on this, the $kNDVI$ formula can be simplified to:

$$kNDVI=\tanh \left({NDVI}^2\right)$$

(2)

2.3.2. MK and Sen Trend Analysis

The Theil–Sen median method is a robust, nonparametric trend analysis method that is resistant to measurement errors and outliers, making it widely used in long-term time-series trend analysis [41]. Compared to the linear regression model, which requires the time series to meet the assumption of normal distribution, the Theil–Sen median method is more suitable for estimating the long-term trend of vegetation change, and it is often applied in conjunction with the Mann–Kendall nonparametric test. Based on the classification methods of different scholars and the characteristics of the study area, we constructed a classification system for S_kNDVI: a kNDVI pixel value between −0.0005 and 0.0005 was defined as a stable state of vegetation, ≥0.0005 indicated an increasing trend of vegetation, and ≤−0.0005 was judged as a degradation of vegetation. Significant trends were characterized when the Z-score (ZS) was greater than 1.96 or less than −1.96, while trends taking values within the range of −1.96 to 1.96 were categorized as non-significant [42].

The Mann–Kendall test, a nonparametric statistical method, does not require the measurements to satisfy normal distribution nor does it constrain the trend to be linear. This makes it widely used to test the significance of trends in long time-series data [43]. Given the extensive spatial coverage of the study area, accurately determining the spatial distribution of the growing season is challenging. Therefore, the annual mean kNDVI was used as an indicator to quantify changes in ecosystem trends, and the Theil–Sen slope and (M-K) trend test were applied to analyze trends in the kNDVI in ASARs at the pixel level over the period 2001–2020. The median Theil–Sen slope was calculated as follows:

$$\beta =Median\left({\displaystyle \frac{x_j-x_i}{j-i}}\right),1\le i<j\le n$$

(3)

where $x_i$ and $x_j$, respectively, represent the NDVI values in year i and year j, n signifies the length of the time series, and $\beta$ denotes the slope of NDVI change. Specifically, $\beta$ > 0 indicates an upward trend in the NDVI, whereas $\beta$ < 0 indicates a downward trend.

The Mann–Kendall test is calculated as follows:

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

(4)

$$S={\sum }_{i=1}^{n-1}{\sum }_{j=i+1}^{n}sgn\left(x_j-x_i\right)$$

(5)

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

(6)

$$Vars\left(S\right)={\displaystyle \frac{n(n-1)(2n+5)}{18}}$$

(7)

where $sgn$ represents the sign function; $Z$ and $S$ correspond to the standardized and non-standardized test statistics, respectively; and $Vars\left(S\right)$ denotes the variance of $S$. The variable $Z$ can assume any real number within the interval of (−∞, +∞). For two-tailed significance tests, statistical significance is evaluated at the 95% confidence level (significance level, α = 0.05). Specifically, when the absolute value of $Z$ exceeds 1.96, the trend is considered statistically significant.

2.3.3. XGBoost

The Extreme Gradient Boosting (XGBoost) algorithm is an integrated machine learning framework that combines an additive model with a forward stepwise optimization strategy. The algorithm minimizes the prediction residuals of the previous iteration by iteratively constructing a classification tree, gradually reducing the difference between the predicted probabilities and the actual labels, thus improving the prediction accuracy of the model. It has the advantages of fast training speed, robustness to data quality, efficient classification performance, and high tolerance to multicollinearity [44].

$$\mathcal{L}\left(\theta \right)=\sum _{i=1}^N\ell \left(y_i,{\widehat{y}}_i\right)+\sum _{i=1}^K\Omega \left(f_k\right)$$

(8)

Here, $\ell \left(y_i,{\widehat{y}}_i\right)$ denotes the loss function, quantifying the discrepancy between the model’s predicted value ${\widehat{y}}_i$ and the true label $y_i$, while $\Omega \left(f_k\right)$ represents the regularization term, which constrains the model’s complexity to prevent overfitting. The variables N and K correspond to the number of samples and the number of trees, respectively.

2.3.4. SHAP

In the study, the SHAP algorithm is based on the Shapley value theory of cooperative game theory and realizes the transparency of the model results by deeply analyzing the nonlinear decision-making mechanism of the XGBoost model [45]. At the global interpretation level, SHAP combs the decision logic of the model from a holistic perspective and reveals the macro-role of each feature in the model operation; at the local interpretation level, the Shapley value of each feature variable is precisely calculated for specific samples to determine the degree of contribution of each feature to the model output results [46,47]. The Shapley value formula for feature i is shown as follows.

$${\varphi }_i={\sum }_{S\subseteq N\left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(N-\left|S\right|-1\right)!}{n!}}\left[f\left(S\cup \left\{i\right\}\right)-f\left(S\right)\right]$$

(9)

where ${\varphi }_i$ represents the contribution of the feature i, $N$ denotes the set with n features, and $f\left(S\cup \left\{i\right\}\right)$ and $f\left(S\right)$ represent the model results with or without the feature i.

During the construction of the vegetation dynamics trend analysis model, the XGBoost-SHAP model integrates multi-dimensional drivers such as climatic variables, human activity indicators, terrain parameters, and soil characteristics. The dataset was randomly divided into 70% training set and 30% testing set to ensure the independence and representativeness of the model training and testing data. In the parameter setting of the binary classification task, the learning_rate is set to 0.1 to control the learning step length of each iteration; max_depth is set to 15 to limit the depth of the tree to avoid overfitting; the sub-sampling rate adopts 0.7 to reduce the data dependence. The gamma is set to 0.1, the lambda is set to 2, the L2 regularization parameter λ is set to 0.1, and the L1 parameter α is set to 0. The regularization strategy is based on the regularization strategy, and the L1 parameter α is set to 0. The generalization ability of the model is further optimized by the regularization strategy. To ensure the statistical reliability of the results, the model is cross-validated with 10 folds and iterated 150 times, and the results are averaged each time. The model accuracy was evaluated by regression analysis of the predicted and observed values of the training and test sets using R software, and a number of validation metrics, such as accuracy (ACC), precision (PRE), recall (REC), F1 score (FS), and AUC value, were calculated based on the confusion matrix to comprehensively measure the prediction performance and classification effect of the model [48].

3. Results

3.1. Spatial and Temporal Characteristics of kNDVI

3.1.1. Spatial Distribution Characteristics of kNDVI

Figure 3a shows the spatial distribution characteristics of the mean kNDVI values, which are calculated using 20 consecutive years of kNDVI data from the study period. As shown in the figure, the spatial distribution of the kNDVI in ASARs indicates that vegetation cover is higher in the east and south and lower in the center and northwest. The average kNDVI of the whole region is 0.044, with values range from 0 to 0.427. Using the natural breakpoint method, the statistical summary results of the average kNDVI values and the proportion of each category in ASARs for 20 years are shown in Figure 3b. The kNDVI values less than 0.427 are the same as those of 0.044 for the whole region. Among these, areas with kNDVI values less than 0.023 are low vegetation cover level Ⅰ areas, which are mainly distributed in the ecologically fragile areas around the western part of the Inner Mongolia Plateau, the Junggar Basin in Xinjiang, the Tarim Basin, etc.; the areas with kNDVI values between 0.024 and 0.072 are level Ⅱ areas, which are mainly located in the eastern part of Inner Mongolia, the central and western parts of Gansu, the southern part of Ningxia, and the northern slopes of the Tianshan Mountain in Xinjiang. Areas with kNDVI values between 0.073 and 0.230 are classified as levels III and IV, representing regions with high vegetation cover, while values greater than 0.231 correspond to level V areas, with very high vegetation cover, primarily found in Gansu, southern Ningxia, and the surrounding regions of the Daxing’anling Mountain Range.

3.1.2. Trends in Spatial and Temporal Patterns of kNDVI

During the period 2001–2020, kNDVI exhibited interannual fluctuating changes, with the range of kNDVI values varying widely from year to year (Figure 4). A linear trend analysis was performed to characterize the overall change in the ASAR annual mean kNDVI. It showed an overall upward trend, reaching a mean value of 0.0542 in 2019 c. The overall increase from 2001 to 2020 was about 0.016, with an increase rise of 57.89% (p < 0.01). From a spatial perspective, the distribution of the kNDVI in ASAE showed some spatial heterogeneity as a whole. The high values of the kNDVI in the arid and semi-arid zone are mainly concentrated in the northeast and south, while the low values are widely distributed in the central and western regions, which indicates that the vegetation conditions in the east and south are more favorable. The increase in annual kNDVI in the arid and semi-arid regions of China during this period indicates a general improvement in ecological conditions.

3.1.3. Spatial Distribution of kNDVI Trends

We revealed the spatial trends of the kNDVI in the arid and semi-arid regions of China based on the Theil–Sen median and Mann–Kendall trend test methods. The results of the former showed (Figure 5a) that the kNDVI increased significantly around the Daxinganling in the east, in the south, and locally in Xinjiang (S_kNDVI > 0.05%, 26.71%). However, some of the densely vegetated areas in its vicinity showed significant decreases (S_kNDVI < 0.05%). The latter showed significant changes in the west–central and southern regions, mainly around the Tarim Basin, etc., and not in the rest of the region (Figure 5b). Based on the characteristics of the range of values, Figure 5c visualizes the spatial distribution and statistical results of the kNDVI trend in China from 2001 to 2022. The long-term trend analysis of the kNDVI shows that 26.7% of the region exhibits an increasing trend, 4% exhibits a decreasing trend, and 69.3% of the region is regionally stable, which implies to indicate that the overall trend of the arid and semi-arid vegetation is improving, with vegetation greening. Areas with increasing kNDVI trends usually show better average vegetation conditions than stable or decreasing trends.

3.2. Drivers of Vegetation Dynamics Change in the ASAR

3.2.1. Degree of Relative Contribution of Drivers to kNDVI

The growth of the kNDVI can be regarded as the process of vegetation greening. The F1_score of the vegetation type showing a greening trend is 0.9510, the AUC value is 0.9792, the precision index is 0.9521, and the accuracy index is 0.9288, and all of the above parameters exceed the threshold value of 0.8, which indicates that the constructed model has excellent prediction efficiency. Based on this, we used the Extreme Gradient Boosting model (XGBoost) to quantitatively analyze the nonlinear response of natural environmental factors and anthropogenic factors in the process of vegetation dynamics change in the target area.

In this study, the five key drivers influencing vegetation dynamics change were the water content index (Bio 23, 19.6%), the change in land use (LUCC, 15.2%), average annual precipitation (pre, 15.0% contribution), population density (pop, 13.2%),and rainfall seasonality (Bio 15, 10.9%) (Figure 6). In terms of variable categorization, natural factors had the highest combined importance (64.3%), followed by socioeconomic factors (28.6%) and soil variables (3.9%). In contrast, terrain variable had the lowest contribution (3.2%). In addition, the explanatory contributions of gross domestic product (GDP) and nighttime light index (NTL) to human activities, as well as the total clay content (T_clay), basic soil saturation (T_bs), acidity and alkalinity (ph) with soil organic carbon (OC), on soil properties were all less than 1%. In summary, the vegetation dynamics status in arid and semi-arid areas is mainly regulated by the combination of climatic conditions, anthropogenic interventions, and soil properties, with bioclimatic variables (Bio 23), population density (pop), average annual rainfall (pre), and bioclimatic variables (Bio 15) having a particularly significant influence.

3.2.2. Nonlinear Mechanisms of Action of Vegetation Dynamics Change Drivers

The nonlinear dependency plot, constructed based on the Extreme Gradient Boosting with SHapley Additive exPlanation (XGBoost-SHAP) model (Figure 7), reveals a significant nonlinear association characteristic between the degree of vegetation greening and each screening variable. In this plot, the horizontal dashed line indicates SHAP = 0, which is the key threshold for determining whether a factor has a positive or negative effect on vegetation greening [49]. The vertical dashed line effectively divides the data distribution into two categories: cases with SHAP > 0, indicating a positive impact, and cases with SHAP < 0, which indicate a negative impact on vegetation greening.

The water content index (Bio 23) was defined as the annual precipitation (mm) divided by the annual potential evapotranspiration (mm). When the water content index (Bio 23) is in the range of (0.05, 0.15, 0.25–0.7), there is a negative impact on vegetation greening, while the rest of the range is a beneficial impact. In terms of land use change (LUCC), the distribution of unutilized land is not conducive to the phenomenon of vegetation greening. For population density (pop), its increase showed positive SHAP values, indicating that it promoted the greening of vegetation. When the average annual precipitation (PRE) was greater than 350 mm, the corresponding SHAP value was greater than 0, indicating that the greening of vegetation was more likely to improve with increasing precipitation. Precipitation seasonality (Bio 15) is one of the bioclimatic variables that indicates the degree of inter-monthly variability of precipitation and is usually calculated as the coefficient of variation (CV) of precipitation (standard deviation of monthly precipitation/mean annual precipitation $\times$ 100). When the seasonality of precipitation (Bio 15) is in the range of (0, 1.8), its effect is bidirectional; after exceeding 2, it gradually turns into a positive effect. When it exceeds 4.5, it has a negative effect on vegetation greening. Evapotranspiration (ET) at thresholds below 5000 as well as greater than 13,000 shows mainly a hindering effect on vegetation greening and more of a facilitating effect in between. Temperature (tem) has a positive effect on vegetation greening above the 3 °C threshold. Additionally, when the elevation exceeds 3000 m, the promotion effect on vegetation greening gradually changes to an inhibitory effect.

3.2.3. Interaction Mechanisms of kNDVI Drivers

In this study, an interaction plot of SHAP values was generated for the key factors to illustrate their interaction effects on the kNDVI (Figure 8). In this interaction plot, the horizontal axis represents the SHAP interaction value, which reflects the strength and direction of the interaction effect. The farther the point is from the center line, the stronger its effect. The vertical axis represents the different variables. Clustering of points indicates the model’s certainty about the interaction between these two variables. Closer clusters of points indicate more consistent predictions of the interactions between these variables, while dispersed points indicate greater uncertainty in the predictions [50].

Climate variables in ASARs exhibit stronger interactions with each other. Among the factors with strong interactions were Bio 23, Bio 15, and the interaction between precipitation, which positively affected vegetation greening. Evapotranspiration (ET) also showed some interaction with total precipitation (pre) and Bio 23, but the interaction was relatively weak. No significant interaction of temperature (tem) with other factors was detected for vegetation greening. Population density produced relatively strong interactions with Bio 23, Bio 15, and rainfall and showed significant positive and negative interactions. This may be due to the fact that the ASAR, as a typical region in China with a vast land area, sparse population, perennial drought, and water scarcity, the population in the arid zone is highly dependent on limited water resources, and the spatial and temporal distribution characteristics of precipitation will directly determine the water resource availability. In localized areas with abundant precipitation, the population density may show a positive synergistic effect, where sufficient precipitation can support the agricultural/domestic water use and increase the population aggregation effect, which is conducive to the vegetation gain. As the implementation area of the ecological project “Three North Protective Forests”, the utilization rate of precipitation can be increased, and the growth rate of the kNDVI in the natural state can be increased. In the region of very low precipitation, the antagonistic effect of negative interaction may appear, because the reduction in precipitation will increase the competition for water resources, which will inhibit the carrying capacity of the population and lead to a decrease in the sensitivity of the kNDVI to precipitation [51]. Soil electrical conductivity (T_ece) interacts with Bio 23 and Bio 15, and the biophysical mechanisms of precipitation and soil affect the coupling effect of soil salinity–moisture, which influences the efficiency of water use by vegetation and, thus, the process of vegetation dynamics change.

4. Discussion

4.1. Factors Influencing Changes in kNDVI

By quantitatively evaluating the contributions of climate change and human activities to the changes in the kNDVI, this study clearly reveals the significant roles of different factors in the dynamics trend of vegetation, which echoes the study of Cao et al. [52] on the impacts of climate change and human activities on vegetation in arid regions and further confirms the complex challenges of vegetation ecosystems in the context of global change [53]. The significant increase in the kNDVI in the arid and semi-arid regions in the last two decades not only visualizes the dynamic evolution of regional vegetation cover and the positive transformation of ecosystems but also emphasizes the key role of long-term vegetation monitoring and ecological project implementation in promoting ecological restoration [54,55]. The changes in the KNDVI time series reflect the complexity of current global vegetation dynamics: on the one hand, it echoes the global dryland degradation trend reported by Burrell et al. (2020), while on the other hand, it also captures the greening of vegetation in specific regions [56,57], highlighting the effectiveness of regional ecological governance and the complexity of global ecological issues.

It has been found that climate change and human activities jointly influence vegetation changes [58]. Considering the interactions of different drivers, the relative importance of these factors on vegetation dynamics change can be more comprehensively revealed [59]. For vegetation showing a greening trend, natural factors (64.3%) had a significantly higher influence weight than socioeconomic variables (28.6%), and topography (3.2%) and soil properties (3.9%) also played a significant role. However, due to the relative stability of topographic factors and soil properties in the short term, the strength of their driving roles weakened in the dynamically changing ecological processes. It is noteworthy that despite the outstanding contribution of bioclimatic factors as a whole, the influence weight of pop (13.2%) in anthropogenic factors was higher than that of some bioclimatic variables (e.g., Bio 1 and Bio 17), suggesting that anthropogenic activities may still have a critical impact on vegetation succession at local scales [60,61]. Among the topographic factors, the driving effect of DEM (1.4%) and slope (1.8%) on vegetation greening further reinforce the importance of the altitudinal gradient on vegetation evolution, which is consistent with Wang et al.’s view that altitude is a key driver of vegetation transition from greening to browning [62] and also fits with the research results on the differentiated evolution of vegetation along the altitudinal gradient in west–central and warm–temperate regions of China [63].

At the level of climatic factors, this study clarified that 20-year average precipitation and annual average temperature have more critical roles in vegetation dynamics change in arid regions. The average annual precipitation (PRE) acts as a central regulator of vegetation greening through multi-factor interactions, regulating soil water effectiveness and organic matter decomposition processes and then affecting nutrient availability [64]. As a bioclimatic variable reflecting the regional water supply and demand balance, Bio 23 has a particularly significant effect on changes in the kNDVI. In the semi-arid zone (Bio 23 ≈ 0.5~1), vegetation types dominated by grasses and shrubs are highly sensitive to precipitation changes, showing obvious “greening” or “browning” responses; in the arid zone, the kNDVI, on the other hand, mainly reflects short-term precipitation events. In arid regions, the kNDVI mainly reflects the transient changes in vegetation induced by short-term precipitation events, and water stress becomes a key factor limiting vegetation growth [65,66,67]. In addition, temperature seasonality (Bio 3), as an important indicator of interannual variability, seasonal fluctuations, and extreme events of climate elements, showed significant regional differences in its impact on vegetation growth. The finding that extreme high temperatures inhibit vegetation growth in arid regions, while they may act as a facilitator in humid regions [68], provides additional evidence of the complexity of the effects of climatic factors on vegetation in this study.

In northwest China, water supply has long been a core factor restricting vegetation growth and ecological restoration [69]. With the advancement of regional greening projects, although the surface vegetation coverage has increased, an ecological feedback series triggered by greening may further affect the stability of local ecosystems. Vegetation greening affects air temperature and soil moisture through biophysical processes and may exacerbate the occurrence of compound soil drought and high-temperature events in the future [70]. This process is particularly significant in arid and semi-arid regions, and artificial greening may bring additional ecological risks. Studies by Zhao et al. have shown that in arid and semi-arid regions, the increase in afforestation area will enhance transpiration, leading to local water resource shortages and deterioration of soil ecosystems, thereby triggering local droughts [71]. Compared with natural forests, artificial forests show weaker drought resistance and ecological resilience when facing water stress [72]. The greening process has strengthened surface evapotranspiration (ET), intensified water exchange between land and atmosphere, and further aggravated the tension of regional water resources. Especially in the context of precipitation-driven greening, the increase in ET may lead to reductions in soil moisture and surface runoff, thereby causing more severe droughts at the watershed scale [73].

In addition, although the introduction of drought-resistant tree species during the greening process helps vegetation survival, it may also pose a risk of biological invasion, especially in areas with few native species and simple ecological structures [74]. The ecosystems in arid regions are inherently fragile with scarce water resources, and the competition among species mainly revolves around limited water and nutrients. The introduced alien species, with strong adaptability and competitiveness, easily break the original ecological balance. The simplicity of this structure increases the sensitivity of biological groups to external disturbances and may have long-term impacts on the entire ecosystem [75]. For example, on the Alxa Plateau with an annual average precipitation of less than 200 mm, plants have developed unique survival strategies to adapt to extreme conditions, such as deep root systems of vegetation. Once the co-evolutionary relationships among these species are disrupted by invasive alien species, the material cycle and energy flow of the entire ecosystem will be severely disturbed. As pointed out by Gao et al., the simplicity of arid ecosystems makes them less resistant and resilient to alien species invasion, and the long-term cumulative ecological impacts may exceed the restoration capacity of artificial intervention [76].

4.2. Nonlinear Response of Vegetation Dynamics Trends to Driving Factors

In this study, the XGBoost-SHAP model was employed to quantify the nonlinear response of natural and anthropogenic factors to the trend of vegetation greening in the ASAR, which provides a new perspective to understand the evolutionary mechanism of complex ecosystems. In terms of precipitation, a key natural factor, it was found that there was a significant threshold effect between precipitation and vegetation greening: only when precipitation exceeds 350 mm in the current year could the vegetation greening process be effectively promoted. This conclusion was further verified in the spatial distribution: vegetation growth in areas with abundant precipitation, such as around the Daxinganling Mountains in the east, was significantly better than that in the arid zone in the central part of the country [69]. Extreme precipitation events play an important role in this, as such precipitation not only directly replenishes the surface soil moisture but also penetrates into the deeper layers of the arid region through infiltration, which builds a long-term water supply mechanism for vegetation growth [70]. As shown in Li et al.’s study, the supplementation of soil moisture by extreme precipitation in arid regions can effectively enhance ecosystem productivity and carbon accumulation and then promote the greening process of vegetation [77]. It is also noteworthy that under the special climatic conditions of arid regions, high evapotranspiration triggered by increasing temperatures during the growing season can exacerbate water stress, but it also encourages vegetation to adapt to the environment by enhancing photosynthesis efficiency, elevating evapotranspiration rates, and increasing the strength of root water uptake, which results in the transient and sensitive response of desert vegetation to changes in precipitation [78]. In addition, the nonlinear relationship between the bioclimatic variable Bio 23 and vegetation greening is also noteworthy, as higher levels of Bio 23 (Bio 23 > 0.7) often create favorable conditions for vegetation greening, further highlighting the centrality of regional water supply and demand balance in the ecological evolution of vegetation [79].

The complex association between anthropogenic activities and vegetation greening was deeply analyzed in this study. In China, for example, the implementation of large-scale ecological restoration projects has not only effectively promoted local vegetation greening but also positively led the global greening process, which proves that scientific and rational human activities can be an important contribution to vegetation restoration in the context of climate change [80]. Focusing on the ASAR, the relatively low population density makes the intervention of human activities on the natural ecosystem in a controllable range, and the strong self-repairing and regulating ability of the ecosystem can be given full play to, which creates a favorable environment for vegetation greening. However, this law is very different in densely populated areas: once the population density exceeds the critical threshold, the high intensity of human activities will continue to exert pressure on the vegetation survival environment, which will not only inhibit the process of vegetation greening but may also lead to irreversible ecological degradation [81].

The influence of topographic factors on vegetation greening also showed a significant nonlinear threshold effect. Elevation and slope, as key topographic factors, indirectly regulate vegetation growth conditions by influencing soil water distribution, soil thickness, and surface runoff [82]. In the ASAR, the complex and diverse landforms (plains, hills, and mountains) constitute a differentiated environment for vegetation growth. It was found that the promotion effect on vegetation greening gradually changed to inhibition effect when the altitude exceeded 3000 m, which was especially obvious in the western high-altitude region. The relationship between slope change and vegetation greening is a complex one: as the slope increases, human activities are reduced due to the difficulty of development, and the vegetation improves under the relatively relaxed environment of external disturbance. However, in non-urban areas dominated by agricultural cultivation, the effect of slope change on vegetation greening is not significant due to the limited degree of terrain modification by human activities [83]. This nonlinear response between topographic factors and vegetation greening profoundly reveals the intrinsic laws of ecosystem evolution under the interaction of natural geographic conditions and human activities.

4.3. Policy Implications

This study reveals the key driving mechanisms of vegetation greening in arid and semi-arid zones and its nonlinear response characteristics, providing a crucial foundation for the formulation of scientific ecological management policies. It was found that moisture-related factors (e.g., Bio 23 and multi-year average precipitation) contributed significantly more to vegetation greening than temperature indicators, which is highly consistent with the practical experience of ecological restoration in arid zones around the world. For example, the Sahelian region of Africa has effectively enhanced vegetation cover through the promotion of rainwater harvesting and water-saving irrigation techniques, confirming the centrality of moisture management in ecological restoration in arid zones [84]. Therefore, it is recommended to prioritize the implementation of ecological projects in regions with average annual precipitation of 200–600 mm, accompanied by measures to optimize the allocation of water resources, whereas in extreme arid zones (precipitation <50 mm), human intervention should be reduced to avoid wasting resources. The relationship between human activities and vegetation greening shows significant nonlinear characteristics. Moderate population density (e.g., ASAR) can promote vegetation restoration through ecological projects (e.g., Three North Protective Forest), but overexploitation can trigger ecological degradation. This finding is consistent with international experiences, such as Australia’s “Green Corridor” program, which balances agriculture and animal husbandry with ecological protection to achieve sustainable vegetation management [85]. It is recommended to implement the principle of “green by water, green by land” in ecologically fragile areas, strictly limit the planting of high water-consuming vegetation, establish an ecological compensation mechanism, and guide the community to participate in vegetation protection. At the same time, we need to be vigilant about the negative effects of large-scale afforestation, such as the risk of pests and diseases, and we can learn from the experience of the United States “adaptive afforestation” and prioritize the selection of native drought-tolerant species [86].

The threshold effect of topographic factors on vegetation greening provides new ideas for regional differentiated management. Vegetation restoration in high-altitude areas should avoid blind afforestation and adopt a natural restoration-oriented approach, similar to the conservation strategy of alpine meadows in the Swiss Alps [87]. In addition, the interaction between precipitation seasonality (Bio 15) and soil properties suggests that soil improvement needs to be strengthened in areas with high precipitation variability, which can significantly enhance the drought-resistant capacity of vegetation. It is recommended to incorporate the research results into an ecological engineering assessment system, to develop precise restoration programs for different elevations, slopes, and precipitation zones, and to establish a long-term monitoring network to quantify policy effects. Finally, this study emphasizes the synergistic management of climate change and human activities. Lessons learned from China’s Three North Project show that simply pursuing vegetation cover can be counterproductive. Future policies should draw on international “nature-based solutions” (NBSs), such as the European Union’s “Green Infrastructure” program, which integrates vegetation restoration with carbon sink enhancement, biodiversity conservation, and other multi-objectives [88].

In addition, to address long-term ecological challenges such as vegetation browning, we believe that the establishment of participatory mechanisms is indispensable [89]. Public participation can not only enhance the sense of identity and responsibility of all sectors of society for ecological protection work but also ensure the effective implementation of ecological protection policies through community collaborative participation and social supervision, forming a new ecological governance pattern led by the government, social collaboration, and public participation [90]. The findings of this study are of great significance for the development of scientific and effective ecosystem management policies, especially in arid and semi-arid regions that are highly vulnerable to extreme climate events, and can provide important theoretical guidance and practical reference for mitigating the negative impacts of climate change on terrestrial ecosystems and realizing sustainable development of ecosystems.

4.4. Limitations and Future Research

This study uses the annual mean kNDVI to assess vegetation dynamics, a choice motivated by the unique ecological characteristics of arid and semi-arid regions. This approach effectively integrates vegetation conditions across all seasons, capturing both direct and lagged effects of drought (e.g., spring droughts delaying germination into summer or summer droughts suppressing photosynthesis and slowing soil moisture recovery until autumn or the next year) and avoids subjectivity in defining a uniform “critical season” amid spatially varying growing seasons due to complex topography (e.g., shorter seasons in high-altitude areas). However, limitations remain: the annual mean kNDVI cannot fully distinguish the specific impacts of droughts in different seasons and may insufficiently characterize the immediate responses to extreme drought events.

Future research will address these constraints by integrating process mechanism models to explicitly quantify time-lag effects of drought on vegetation and fusing high-er-resolution remote sensing and ground observation data to refine the understanding of drought–vegetation interactions. Additionally, the research scope will be expanded to other arid and semi-arid regions, with cross-regional comparative analyses to reveal the universality and specificity of driving mechanisms behind vegetation dynamics, thereby providing more systematic scientific support for ecological governance in global arid regions.

5. Conclusions

This study systematically investigated the spatiotemporal dynamics of vegetation greening and its driving mechanisms in China’s arid and semi-arid regions from 2001 to 2020. By integrating trend analysis (Mann–Kendall test and Theil–Sen estimator) with interpretable machine learning (XGBoost-SHAP model), we quantified the nonlinear responses of vegetation greening to climatic, anthropogenic, topographic, and soil factors. The results revealed a significant increasing trend in the kNDVI across 26.7% of the study area, with key drivers identified as the moisture index (Bio 23, 19.6%), the change in land use (LUCC, 15.2%), annual precipitation (pre, 15.0%), population density (pop, 13.2%), and precipitation seasonality (Bio 15, 10.9%). Furthermore, threshold effects and interactive mechanisms among these factors were elucidated, providing deeper insights into vegetation dynamics under complex environmental changes.

However, this study still has certain limitations. The Mann–Kendall test and Theil–Sen estimation are mainly applicable to linear or monotonic nonlinear trend analysis, with limited ability to capture non-monotonic complex fluctuating trends. Although the XGBoost-SHAP model can reveal nonlinear relationships, its performance depends on the completeness of input variables and may ignore certain unincorporated potential driving factors, such as the instantaneous impact of extreme climate events, and its explanation of extremely short-term and long-term dynamic responses is insufficient. In addition, grid analysis at the spatial scale may mask the impact of microtopography or small-scale ecological processes on vegetation. Future research will improve and expand on these limitations, mainly including the following: first, integrating high-resolution remote sensing data with ground observation data, optimizing model input variables, and combining high-resolution climate models with remote sensing technology to enhance the explanatory power for complex ecological processes; second, extending the research period to focus on analyzing the lag effects and cumulative impacts of extreme climate events on vegetation dynamics; and third, expanding the research area to other arid and semi-arid regions, conducting cross-regional comparative analyses, and predicting vegetation changes under different scenarios to reveal the universality and particularity of the driving mechanisms of vegetation dynamics, thereby providing more comprehensive scientific support for global arid region ecological governance.