Unveiling Drivers and Projecting Future Risks of Desertification Vulnerability in the Mongolian Plateau

Abstract

Desertification presents a significant ecological challenge in arid and semi-arid regions, posing a severe threat to regional ecological security and sustainable development. This study introduces an integrated framework for desertification vulnerability assessment, combining the MEDALUS model with the XGBoost algorithm, to evaluate desertification dynamics across the Mongolian Plateau from 2000 to 2020 and project future trends under four Shared Socioeconomic Pathways (SSPs) for 2030. The findings are as follows: (1) Between 2000 and 2020, desertification vulnerability was most pronounced in the southern and western regions of the plateau, with lower vulnerability observed in the northern and eastern areas. High-vulnerability zones expanded over time, highlighting the need for targeted and prioritized management efforts. (2) Climate factors—particularly temperature, wind speed, and precipitation—emerged as the dominant drivers of desertification, followed by soil characteristics and vegetation (NDVI). The influence of human activities on desertification became increasingly significant, stressing the need for improved land management and sustainable practices. (3) Future risks show that desertification vulnerability in the Mongolian Plateau will intensify under high-emission scenarios (SSP3-7.0, SSP5-8.5), with significant expansion of high vulnerability areas. Lower-emission scenarios (SSP1-2.6, SSP2-4.5) may reduce some impacts, but high vulnerability will persist, highlighting the need for urgent climate mitigation and adaptation efforts.

1. Introduction

Desertification is one of the most critical manifestations of global ecosystem degradation, representing a significant threat to regional ecological security and sustainable development [1,2,3]. The Mongolian Plateau, home to the largest contiguous temperate grassland ecosystem in Eurasia [4,5], serves as a vital ecological transition zone that connects the East Asian monsoon domain with the arid regions of Central Asia [6,7]. In recent decades, this region has experienced a marked decline in ecological stability, driven by the compounded effects of climate change and anthropogenic pressures [8,9,10]. Currently, over 72% of the land on the Mongolian Plateau is affected by desertification to varying degrees [11,12], making the region a critical hotspot for investigating desertification dynamics in dryland systems globally.

Traditional desertification assessment approaches often rely on a single driving mechanism, limiting their ability to capture the complex, coupled nature of human-environment systems. For instance, models such as the China Land Desertification Assessment System (CLDAS) focus primarily on meteorological influences [13,14,15], while remote sensing-based models like the Revised Universal Soil Loss Equation (RUSLE) emphasize soil erosion [15,16]. These models, though valuable, often overlook the spatiotemporal heterogeneity, multiple data sources, and integrated analysis of driving factors that are essential for a comprehensive desertification assessment [17,18]. For example, Chen et al. utilized Google Earth Engine and machine learning techniques to identify desertification drivers in Qilian Mountain National Park [19], and Vieira et al. highlighted the importance of social factors in analyzing vulnerability patterns in Brazil’s Northeast [20], while Vieira et al. emphasized the role of social factors by analyzing vulnerability patterns in Brazil’s Northeast [19]. The Mediterranean Desertification and Land Use (MEDALUS) model, developed by the European Commission in the early 1990s, is widely used to identify Environmentally Sensitive Areas (ESAs) to desertification by integrating climate, soil, vegetation, and land management indices [21]. These indices are combined using a geometric mean to assess sensitivity levels. While this model has been applied in various regions, including the Souss River Basin and Niger River Basin, its utility is limited by the lack of consideration for dynamic interactions and weight adjustments among the driving factors [22,23]. The integration of Geographic Information Systems (GIS) enhances spatial analysis, supporting land management and policy-making efforts to combat desertification [24,25].

Recent advancements in machine learning, particularly the Extreme Gradient Boosting (XGBoost) algorithm, offer significant potential for ecological modeling [26,27,28]. XGBoost is recognized for its ability to handle high-dimensional feature selection, adapt to nonlinear relationships, and provide high predictive accuracy [29]. When combined with SHapley Additive exPlanations (SHAP), XGBoost can also offer enhanced interpretability, quantifying the relative contributions of individual predictors and improving the transparency and scientific credibility of model results [30,31]. This combination not only improves prediction accuracy but also facilitates a better understanding of complex ecological processes [32,33].

This study proposes an integrated desertification vulnerability assessment framework by coupling the MEDALUS model with XGBoost. This framework addresses the traditional limitations of desertification models, particularly in terms of factor interaction and dynamic weighting. By focusing on the Mongolian Plateau from 2000 to 2020, the study analyzes spatiotemporal desertification patterns using multi-source remote sensing and meteorological data. The XGBoost-SHAP model further reveals nonlinear driving mechanisms and forecasts desertification vulnerability under different Shared Socioeconomic Pathways (SSPs) for 2030, offering valuable insights for China–Mongolia ecological cooperation and advancing the methodology for arid land degradation assessment.

2. Materials and Methods

2.1. Study Area

Located in central Eurasia (37°22′–53°20′N, 87°43′–126°04′E), the Mongolian Plateau spans approximately 2.6 million km², encompassing Mongolia and extensive regions of western and central Inner Mongolia, China [34]. This topographically enclosed highland, averaging ~1580 m in elevation with a west-high, east-low configuration, is bounded by the Altai, Greater Khingan, Sayan–Khangai, and Helan–Yinshan mountain ranges [35]. The region experiences a strongly continental temperate climate characterized by extreme seasonal temperature variations and pronounced precipitation seasonality. Over 70% of annual precipitation occurs between June and August, exhibiting a steep east-west gradient that decreases from 200–400 mm (predominantly rainfall) in the east to arid conditions (predominantly snowfall) in the west [36,37]. This climatic pattern drives distinct vegetation zonation, transitioning from meadow steppe through dry steppe to desert. Driven by climate variability and anthropogenic pressures, the plateau faces widespread and intensifying land degradation. Desertification affects approximately 72% of Mongolia’s territory, while Inner Mongolia contains sandy lands accounting for approximately one-fifth of China’s national total. These changes significantly impair regional ecological stability and functioning, altering carbon cycling, hydrological regimes, and land–atmosphere energy exchanges [38]. Given its position as a climate-sensitive ecotone between the East Asian monsoon and Central Asian arid systems, the Mongolian Plateau serves as a critical region for diagnosing ecological vulnerability and elucidating desertification mechanisms under global change using remote sensing techniques. An overview of the study area is presented in Figure 1.

2.2. Data Sources

Table 1. Data sources.

Data Type	Data Source	Description	Resolution	Website
Digital Elevation Model (DEM)	Shuttle Radar Topography Mission (SRTM) Dataset	Elevation data for the Mongolian Plateau	30 m	https://earthexplorer.usgs.gov/ (accessed on 2 March 2025)
Slope Information	Generated via ArcGIS 10.2	Slope data derived from DEM	30 m	N/A
Meteorological Variables	European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5-Land	Temperature, precipitation, and wind speed data	31 km	https://www.ecmwf.int/en/era5-land (accessed on 6 March 2025)
Climate Projections	WorldClim	Climate projections for future scenarios	30 s (~1 km)	https://worldclim.org/ (accessed on 13 March 2025)
Vegetation Index (NDVI)	MOD09A1 Product (Google Earth Engine)	NDVI computation for vegetation coverage	500 m	https://earthengine.google.com/ (accessed on 17 March 2025)
Soil Index (NDSI)	MOD09A1 Product (Google Earth Engine)	NDSI computation for soil conditions	500 m	https://earthengine.google.com/ (accessed on 21 March 2025)
Land Cover Classification	Moderate Resolution Imaging Spectroradiometer (MODIS) MCD12Q1 v061 Dataset	Land cover classification data	500 m	https://lpdaac.usgs.gov/ (accessed on 5 March 2025)
Soil Characteristics	Harmonized World Soil Database (HWSD) v1.2 (Food and Agriculture Organization of the United Nations)	Soil property data including sand content, organic matter, and texture	1 km	http://www.fao.org/geonetwork/srv/en/main.home (accessed on 10 March 2025)
Demographic Data	WorldPop	Population density data	100 m	https://www.worldpop.org/ (accessed on 18 March 2025)
Spatial Boundary	Ministry of Natural Resources of China	Spatial boundary delineation of the research area	30 m	http://bzdt.ch.mnr.gov.cn/ (accessed on 23 March 2025)
Digital Elevation Model (DEM)	Shuttle Radar Topography Mission (SRTM) Dataset	Elevation data for the Mongolian Plateau	30 m	https://earthexplorer.usgs.gov/ (accessed on 4 March 2025)
Slope Information	Generated via ArcGIS 10.2	Slope data derived from DEM	30 m	N/A
Meteorological Variables	European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5-Land	Temperature, precipitation, and wind speed data	31 km	https://www.ecmwf.int/en/era5-land (accessed on 7 March 2025)
Climate Projections	WorldClim	Climate projections for future scenarios	30 s (~1 km)	https://worldclim.org/ (accessed on 12 March 2025)
Vegetation Index (NDVI)	MOD09A1 Product (Google Earth Engine)	NDVI computation for vegetation coverage	500 m	https://earthengine.google.com/ (accessed on 19 March 2025)
Soil Index (NDSI)	MOD09A1 Product (Google Earth Engine)	NDSI computation for soil conditions	500 m	https://earthengine.google.com/ (accessed on 25 March 2025)
Land Cover Classification	Moderate Resolution Imaging Spectroradiometer (MODIS) MCD12Q1 v061 Dataset	Land cover classification data	500 m	https://lpdaac.usgs.gov/ (accessed on 14 March 2025)
Soil Characteristics	Harmonized World Soil Database (HWSD) v1.2 (Food and Agriculture Organization of the United Nations)	Soil property data including sand content, organic matter, and texture	1 km	http://www.fao.org/geonetwork/srv/en/main.home (accessed on 9 March 2025)
Demographic Data	WorldPop	Population density data	100 m	https://www.worldpop.org/ (accessed on 16 March 2025)
Spatial Boundary	Ministry of Natural Resources of China	Spatial boundary delineation of the research area	30 m	http://bzdt.ch.mnr.gov.cn/ (accessed on 27 March 2025)

provides an overview of the various data sources used in this study, including data types, sources, descriptions, spatial resolutions, and access information, which are crucial for the spatial and environmental analysis conducted in this research.

2.3. Vulnerability Assessment of Desertification Based on the MEDALUS Model

Based on the framework and evaluation criteria of the MEDALUS (Mediterranean Desertification and Land Use) model, this study establishes a Desertification Sensitivity Index (DSI) by selecting 12 representative indicators across four dimensions: Soil Quality (SQI), Vegetation Quality (VQI), Climate Quality (CQI), and Management Quality (MQI) [39,40]. Each parameter is graded on a sensitivity scale from 1 to 4. The corresponding quality indices are computed, and a comprehensive Desertification Vulnerability Index (DVI) is synthesized by integrating the Management Quality Index (MQI), Vegetation Quality Index (VQI), Soil Quality Index (SQI), and Climate Quality Index (CQI) through an entropy-weighted geometric mean, where weights are derived via the Entropy Weight Method. Finally, the DVI values are categorized using the Natural Breaks (Jenks) classification approach.

2.3.1. Soil Quality Index (SQI)

The SQI evaluates topographical impacts and edaphic processes in desertification development [41]. In this study, SQI is evaluated using three indicators: slope, sand content, and soil organic matter content. The Normalized Difference Soil Index (NDSI) is a remote sensing-derived index that effectively reflects soil characteristics, particularly moisture content and surface coverage, by utilizing the spectral differences between the red and shortwave infrared bands. All indicators are classified using the Natural Breaks method, with the classification criteria shown in

Table 2. Classification of Soil Quality Index and its corresponding weights.

Indicator	Grade	Description					Weight
		2000	2005	2010	2015	2020
NDSI	1	<−0.18	<−0.15	<−0.19	<−0.16	<−0.15	1.00
	2	−0.18~−0.08	−0.15~−0.06	−0.19~−0.09	−0.16~−0.07	−0.15~−0.06	2.00
	3	−0.08~−0.01	−0.06~−0.01	−0.09~−0.01	−0.07~0.00	−0.06~−0.01	3.00
	4	>−0.01	>0.01	>−0.01	>0.00	>0.01	4.00
Sand (%)	1	<25					1.00
	2	25~40					2.00
	3	40~65					3.00
	4	>65					4.00
Slope (°)	1	<5					1.00
	2	5~15					2.00
	3	15~30					3.00
	4	>30					4.00
OM (%)	1	>3					1.00
	2	2~3					2.00
	3	1~2					3.00
	4	<1					4.00

. The SQI is calculated as follows:

$$\mathrm{S}\mathrm{Q}\mathrm{I}=(\mathrm{S}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}\times \mathrm{S}\mathrm{a}\mathrm{n}\mathrm{d}\times \mathrm{O}\mathrm{M}{\times \mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I})}^{1/4}$$

where $\mathrm{S}\mathrm{l}\mathrm{o}\mathrm{p}\mathrm{e}$ is terrain slope; $\mathrm{S}\mathrm{a}\mathrm{n}\mathrm{d}$ is soil sand content; $\mathrm{O}\mathrm{M}$ is soil organic matter content; and $\mathrm{N}\mathrm{D}\mathrm{S}\mathrm{I}$ is normalized soil condition index.

2.3.2. Vegetation Quality Index (VQI)

Phytogenic stabilization has a key impact on the progression of land degradation [42]. This study evaluates the Vegetation Quality Index (VQI) from three aspects: erosion protection capability, drought resistance, and vegetation cover as indicated by the Normalized Difference Vegetation Index (NDVI). NDVI, a widely used remote sensing indicator, measures vegetation health by comparing the difference between near-infrared and red reflectance. NDVI values are classified using the Natural Breaks method, with the classification criteria detailed in

Table 3. Classification of Vegetation Quality Index and its corresponding weights.

Indicator	Grade	Description					Weight
		2000	2005	2010	2015	2020
NDVI	1	<−0.13	<−0.13	<−0.13	<−0.14	<−0.15	1.00
	2	0.13~0.23	0.13~0.23	0.13~0.22	0.14~0.25	0.15~0.26	2.00
	3	0.23~0.36	0.23~0.34	0.22~0.34	0.25~0.39	0.26~0.39	3.00
	4	>0.36	>0.34	>0.34	>0.39	>0.39	4.00
Invasion Protection	1	Coniferous Forest, Alpine Vegetation, Others					1.00
	2	Farmland, Grassland, Meadow					2.00
	3	Shrubland					3.00
	4	Desert					4.00
Drought Resistance	1	Coniferous Forest, Others					1.00
	2	Farmland, Alpine Vegetation, Shrubland					2.00
	3	Grassland, Meadow					3.00
	4	Desert					4.00

. The VQI is computed as follows:

$$\mathrm{V}\mathrm{Q}\mathrm{I}=(\mathrm{D}\times \mathrm{P}\times \mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}{)}^{1/3}$$

where $\mathrm{D}$ is drought resistance; $\mathrm{P}$ is invasion protection; and $\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}$ is vegetation health.

2.3.3. Climate Quality Index (CQI)

Desertification patterns are evaluated through climate impact measurements in the CQI [43]. This study analyzes CQI through three indicators: annual mean temperature, precipitation, and wind speed. Classification is performed using the Natural Breaks method, with criteria provided in

Table 4. Classification of Climate Quality Index and its corresponding weights.

Indicator	Grade	Description					Weight
		2000	2005	2010	2015	2020
Precipitation (mm)	1	<75.42	<72.72	<76.02	<85.41	<84.47	1.00
	2	75.42~138.86	72.72~146.66	76.02~130.80	85.41~150.74	84.47~161.68	2.00
	3	138.86~205.83	146.66~227.64	130.80~194.96	150.74~223.32	161.68~240.87	3.00
	4	>205.83	>227.64	>194.96	>223.32	>240.87	4.00
Temperature (°C)	1	<−2.22	<−1.95	<−2.63	<−2.04	<−1.75	1.00
	2	−2.22~1.99	−1.95~2.31	−2.63~1.90	−2.04~1.91	−1.75~2.18	2.00
	3	1.99~5.94	2.31~6.13	1.90~6.17	1.91~5.95	2.18~6.12	3.00
	4	>5.94	>6.13	>6.17	>5.95	>6.12	4.00
Wind Speed (m/s)	1	<2.21	<2.86	<3.47	<4.73	<2.21	1.00
	2	1.00~2.11	2.11~2.81	2.81~3.44	3.44~4.58	1.00~2.11	2.00
	3	1.16~2.31	2.31~3.02	3.02~3.70	3.70~5.12	1.16~2.31	3.00
	4	>1.08	>2.15	>2.79	>3.44	>1.08	4.00

. The CQI is calculated as follows:

$$\mathrm{C}\mathrm{Q}\mathrm{I}=(\mathrm{T}\mathrm{E}\mathrm{M}\mathrm{P}\times \mathrm{P}\mathrm{R}\mathrm{E}\mathrm{P}\times \mathrm{W}\mathrm{I}\mathrm{N}\mathrm{D}{)}^{1/3}$$

where $\mathrm{T}\mathrm{E}\mathrm{M}\mathrm{P}$ is annual average temperature; $\mathrm{P}\mathrm{R}\mathrm{E}\mathrm{P}$ is annual precipitation; and $\mathrm{W}\mathrm{I}\mathrm{N}\mathrm{D}$ is annual average wind speed.

2.3.4. Management Quality Index (MQI)

As an evaluative framework, the MQI correlates with anthropogenic drivers of ecological deterioration in semi-arid regions [41]. In this study, MQI is evaluated using two factors: population density and land use intensity, with classification criteria presented in

Table 5. Classification of Management Quality Index and its corresponding weights.

Indicator	Grade	Description	Weight
Population Density (persons·km−2)	1	<2	1.00
	2	2–5	1.60
	3	5–10	2.20
	4	10–20	2.80
	5	20–40	3.20
	6	>40	4.00
Land Use Intensity	1	Forest, Shrubland, Wetland, Bare Land	1.00
	2	Grassland, Water Bodies	2.50
	3	Built-up Land, Farmland	4.00

. The $\mathrm{M}\mathrm{Q}\mathrm{I}$ is calculated as follows:

$$\mathrm{M}\mathrm{Q}\mathrm{I}=(\mathrm{P}\mathrm{D}\times \mathrm{L}\mathrm{U}{)}^{1/2}$$

where $\mathrm{P}\mathrm{D}$ is the population density and $\mathrm{L}\mathrm{U}$ is the land use type.

2.3.5. Desertification Vulnerability Index (DVI)

The synthesis of the Desertification Vulnerability Index (DVI) employs multi-index aggregation techniques, integrating the Management Quality Index (MQI), Vegetation Quality Index (VQI), Soil Quality Index (SQI), and Climate Quality Index (CQI) through the Entropy Weight Method (EWM), thereby mitigating fluctuations [44]. The $\mathrm{D}\mathrm{V}\mathrm{I}$ is calculated as follows:

$$\mathrm{D}\mathrm{V}\mathrm{I}={\left(\mathrm{C}\mathrm{Q}{\mathrm{I}}^{{\mathrm{w}}_{\mathrm{C}\mathrm{Q}\mathrm{I}}}\times \mathrm{V}\mathrm{Q}{\mathrm{I}}^{{\mathrm{w}}_{\mathrm{V}\mathrm{Q}\mathrm{I}}}\times \mathrm{S}\mathrm{Q}{\mathrm{I}}^{{\mathrm{w}}_{\mathrm{S}\mathrm{Q}\mathrm{I}}}\times \mathrm{M}\mathrm{Q}{\mathrm{I}}^{{\mathrm{w}}_{\mathrm{M}\mathrm{Q}\mathrm{I}}}\right)}^{1/4}$$

In the formula, ${\mathrm{w}}_{\mathrm{C}\mathrm{Q}\mathrm{I}}$, ${\mathrm{w}}_{\mathrm{V}\mathrm{Q}\mathrm{I}}$, ${\mathrm{w}}_{\mathrm{S}\mathrm{Q}\mathrm{I}}$ and ${\mathrm{w}}_{\mathrm{M}\mathrm{Q}\mathrm{I}}$ are the weights determined using the Entropy Weight Method. MQI is the Management Quality Index; $\mathrm{C}\mathrm{Q}\mathrm{I}$ is the Vegetation Quality Index; $\mathrm{S}\mathrm{Q}\mathrm{I}$ is the Soil Quality Index; and VQI is the Vegetation Quality Index.

2.4. Coefficient of Variation

The coefficient of variation (CV), or relative standard deviation, serves as a statistical indicator to evaluate the extent of dispersion in a dataset [45]. It effectively reflects the temporal stability of spatial data over a time series. The Coefficient of Variation (CV) is particularly useful for evaluating the temporal fluctuations in desertification sub-indices, as it provides a measure of the relative variability over time. By calculating the CV, we can assess the degree of change in the sub-indices, helping to identify regions where desertification is more dynamic or undergoing significant shifts. This approach is crucial for understanding the temporal evolution of desertification vulnerability and provides insights into the stability or instability of ecological conditions across different time periods. The calculation formula is as follows:

$${\mathrm{C}}_{\mathrm{V}}={\displaystyle \frac{\mathsf{\sigma }}{\mathsf{\mu }}}$$

The formula for calculating the standard deviation $\mathsf{\sigma }$ is:

$$\mathsf{\sigma }=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{x}}_{\mathrm{i}}-\overline{\mathrm{x}})}^2}$$

In the formula, ${\mathrm{C}}_{\mathrm{V}}$ represents the coefficient of variation; $\mathsf{\sigma }$ represents the standard deviation; $\mathsf{\mu }$ represents the mean; ${\mathrm{x}}_{\mathrm{i}}$ represents the true value of the observed data; $\overline{\mathrm{x}}$ represents the mean of the observed data; $\mathrm{n}$ represents the number of observations; $\mathrm{i}$ represents the time sequence number; and the size of ${\mathrm{C}}_{\mathrm{V}}$ reflects the degree of data fluctuation. The larger the ${\mathrm{C}}_{\mathrm{V}}$ value, the greater the dispersion of the data, indicating poorer stability; conversely, it indicates smaller dispersion and better stability.

2.5. Pearson Correlation Analysis

To explore the linear relationships between the driving factors of desertification, this study adopts the Pearson Correlation Coefficient for quantitative analysis. Pearson’s r assesses how strongly two continuous variables are linearly related, with values between −1 and 1, where 1 and −1 represent perfect correlations and 0 implies no linear link [46]. The formula is:

$$\mathrm{r}={\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}} \left({\mathrm{x}}_{\mathrm{i}}-\stackrel{\mathrm{‾}}{\mathrm{x}}\right)\left({\mathrm{y}}_{\mathrm{i}}-\stackrel{\mathrm{‾}}{\mathrm{y}}\right)}{\sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}} {\left({\mathrm{x}}_{\mathrm{i}}-\stackrel{\mathrm{‾}}{\mathrm{x}}\right)}^2}\cdot \sqrt{{\sum }_{\mathrm{i}=1}^{\mathrm{n}} {\left({\mathrm{y}}_{\mathrm{i}}-\stackrel{\mathrm{‾}}{\mathrm{y}}\right)}^2}}}$$

In this formulation, ${\mathrm{x}}_{\mathrm{i}}$ and ${\mathrm{y}}_{\mathrm{i}}$ denote the measured values of two distinct variables for the $\mathrm{i}$-th observational unit, where $\stackrel{\mathrm{‾}}{\mathrm{x}}$ and $\stackrel{\mathrm{‾}}{\mathrm{y}}$ represent the arithmetic means of their respective variables, and $\mathrm{n}$ indicates the total sample size. The numerator in the expression corresponds to the covariance between the variables, while the denominator constitutes the product of their standard deviations. Through the construction of a correlation coefficient heatmap, the interrelationships among driving factors across variables can be intuitively visualized, thereby elucidating underlying correlation patterns in multivariate datasets. To explore the relationships between the variables used in the DVI index, the Pearson Correlation Coefficient was applied to parameters such as vegetation cover (NDVI), climate variables (temperature and precipitation), soil conditions, and human activity factors. This analysis identified significant correlations among these variables, helping to reveal their combined impact on desertification. A correlation heatmap was also created to visually represent the strength and direction of these relationships, facilitating the identification of key interactions driving desertification.

2.6. Driven Interpretation and Prediction Based on XGBoost and SHAP Models

To investigate the driving mechanisms of desertification vulnerability in the study area, This study employs a hybrid modeling framework that integrates Extreme Gradient Boosting (XGBoost) with Shapley Additive Explanations (SHAP) [47]. This combined approach not only provides robust predictive performance but also enhances interpretability by quantifying the contribution of each influencing factor, thereby identifying their relative importance in driving desertification vulnerability.

2.6.1. Construction of the XGBoost Model

XGBoost is a powerful and accurate algorithm that is especially good at modeling complex relationships between different variables. In this study, we used the XGBoost model to identify the main factors driving desertification vulnerability, focusing on four key groups of environmental factors: climate (e.g., temperature, precipitation), vegetation (e.g., NDVI), soil (e.g., sand content), and topography (e.g., elevation, slope). The XGBoost model works by minimizing the difference between the predicted values and the actual values, while also adding a regularization step to keep the model from becoming too complicated and overfitting the data, improving its ability to make accurate predictions on new data [48]. The objective function is formulated as follows:

$$\mathrm{O}\mathrm{b}\mathrm{j}=\sum _{\mathrm{i}=1}^{\mathrm{n}} \mathrm{L}\left({\mathrm{y}}_{\mathrm{i}},{\widehat{\mathrm{y}}}_{\mathrm{i}}\right)+\sum _{\mathrm{k}=1}^{\mathrm{K}} \mathsf{\Omega }\left({\mathrm{f}}_{\mathrm{k}}\right)$$

where $\mathrm{L}\left({\mathrm{y}}_{\mathrm{i}},{\widehat{\mathrm{y}}}_{\mathrm{i}}\right)$ denotes the loss function measuring the discrepancy between the observed and predicted values, and $\mathsf{\Omega }\left({\mathrm{f}}_{\mathrm{k}}\right)$ represents the complexity penalty of the $\mathrm{k}$ tree.

2.6.2. SHAP for Feature Importance Analysis

To enhance the interpretability of the model outputs, the SHAP method was introduced to quantify the marginal contributions of individual driving factors. Rooted in cooperative game theory, SHAP leverages Shapley values to fairly attribute each feature’s impact on the model’s predictions. This method offers not only a prioritized ranking of feature importance but also indicates whether each feature contributes positively or negatively to the prediction outcome [49].

For each feature $\mathrm{i}$, the Shapley value ${\mathsf{\varphi }}_{\mathrm{i}}$ is calculated as:

$${\mathsf{\varphi }}_{\mathrm{i}}=\sum _{\mathrm{S}\subseteq \mathrm{N}\setminus \left\{\mathrm{i}\right\}} {\displaystyle \frac{\left|\mathrm{S}\right|!\left(\right|\mathrm{N}|-|\mathrm{S}|-1)!}{\left|\mathrm{N}\right|!}}[\mathrm{v}(\mathrm{S}\cup \{\mathrm{i}\})-\mathrm{v}(\mathrm{S})]$$

where $\mathrm{N}$ is the set of all features, $\mathrm{S}$ is a subset that does not include feature $\mathrm{i}$, and represents the model prediction given a specific set of features [50]. Through SHAP analysis, this study not only identifies the relative importance of each driving factor but also improves the robustness and transparency of the results, making it suitable for causal analysis in complex environmental systems.

2.6.3. Projection of Desertification Vulnerability in 2030

To further explore the spatiotemporal evolution of desertification risk, this study applies the XGBoost model to predict desertification vulnerability in the year 2030 based on historical data from 2000, 2005, 2010, 2015, and 2020, including temperature, precipitation, and soil sand content. By constructing a time series model that integrates climate and soil variables with corresponding vulnerability data, the XGBoost framework effectively captures the influence of environmental drivers on the evolution of desertification. This enables reliable prediction of the spatial patterns of desertification vulnerability in 2030. The prediction model serves as a useful tool to guide future strategies in land resource management and ecological restoration.

3. Results

3.1. Spatiotemporal Evaluation of Desertification Vulnerability Sub-Indices

As depicted in Figure 2, the desertification vulnerability sub-indices across the Mongolian Plateau demonstrated overall stability from 2000 to 2020, though certain localized areas exhibited either increasing or decreasing trends, reflecting spatial heterogeneity. High-value zones for Vegetation, Climate, and Soil Quality Indices (VQI, CQI, SQI) gradually expanded toward the southwest, signaling an area of increasing desertification vulnerability. Conversely, low-value areas in the northeast exhibited minimal change, remaining spatially stable throughout the period. This pattern suggests that the northern regions may be less susceptible to desertification pressures, possibly due to more favorable climatic conditions and land management practices. In contrast, the Management Quality Index (MQI) exhibited notable spatial variability, with intensified management pressure observed in southeastern regions. This variation can be attributed to higher anthropogenic activities such as agriculture, grazing, and urbanization in those areas, which exert additional stress on the ecosystem.

The coefficient of variation (CV) analysis, as shown in Figure 3, further emphasizes the considerable spatial heterogeneity in desertification vulnerability across the plateau. The CV for VQI, CQI, and SQI showed higher values in the central and northeastern regions, indicating heightened sensitivity to interannual fluctuations in both climate conditions and land use. This increased variability underscores the dynamic nature of the ecosystem, where changes in vegetation, climate, and soil quality are more pronounced and responsive to environmental shifts. In these regions, the landscape appears to be more susceptible to fluctuations in weather patterns, such as droughts or changes in precipitation, which can exacerbate desertification. Meanwhile, the MQI exhibited the most significant fluctuations in the southeast, reflecting the region’s vulnerability to anthropogenic disturbances, including land degradation and unsustainable land management practices, which exacerbate desertification processes. This finding highlights the critical role of human activities in influencing desertification patterns in certain areas of the plateau.

3.2. Spatiotemporal Evolution Analysis of Desertification Vulnerability in the Mongolian Plateau

3.2.1. Spatiotemporal Distribution Characteristics of Desertification Vulnerability

From Figure 4, it is evident that between 2000 and 2020, desertification vulnerability in the Mongolian Plateau exhibited significant spatiotemporal variations. Despite the expansion of low vulnerability areas in certain regions, areas categorized as high and extremely high vulnerability still occupied the majority of the land, indicating the widespread ecological fragility of the region.As shown in Figure 5, in Inner Mongolia, the area of low vulnerability increased from 128.3 × 10³ km² in 2000 to 152.1 × 10³ km² in 2015, before slightly declining to 144.8 × 10³ km² in 2020. The area of moderate vulnerability exhibited a continuous upward trend, suggesting a transition of some high vulnerability regions to a lower vulnerability category, which may be related to ecological restoration and improvements in land management. The area of high vulnerability fluctuated significantly, peaking at 368.8 × 10³ km² in 2010, while the area of extremely high vulnerability consistently remained high, exceeding 550 × 10³ km² throughout the study period. In contrast, most of Mongolia remained predominantly in the high vulnerability category. Although low and moderate vulnerability areas slightly increased after 2010, the high and extremely high vulnerability areas continued to dominate, with their combined area exceeding 950 × 10³ km² annually. The area of extremely high vulnerability reached 427.8 × 10³ km² in 2020, the highest value during the study period for Mongolia, indicating an intensification of land degradation processes, likely linked to climate aridification and human disturbances such as overgrazing. A direct comparison reveals that Inner Mongolia showed a more noticeable improvement in low vulnerability areas, while Mongolia continued to face greater desertification pressure, with limited reductions in high-risk areas. This comparison may reflect differences in the implementation of ecological policies, land use management, and investments in restoration projects. At the scale of the entire Mongolian Plateau, the area of low vulnerability increased by approximately 37 × 10³ km² over the 20-year period. However, extremely high vulnerability areas remained dominant, reaching 973.7 × 10³ km² in 2020. These results highlight the long-term susceptibility of the region to desertification and reveal a spatial mismatch between localized ecological improvements and the overall vulnerability pattern.

3.2.2. Transition Patterns of Desertification Vulnerability

From 2000 to 2020, desertification vulnerability in the Mongolian Plateau exhibited significant spatiotemporal variations (Figure 6). Overall, desertification pressure intensified during most periods, particularly between 2000 and 2005 and 2010 and 2020, with notable expansions in high vulnerability (High) and very high vulnerability (Very High) areas, particularly with the transition from high vulnerability to very high vulnerability regions. This change was driven primarily by climate change, frequent droughts, and human activities, which led to ecosystem degradation and the acceleration of desertification. During the 2000–2005 period, desertification issues began to emerge, particularly with significant transitions from low vulnerability (Low) to moderate vulnerability (Moderate) and high vulnerability areas. During this period, the high vulnerability and very high vulnerability areas increased by approximately 1.22 × 10⁵ km² and 6.51 × 10⁴ km², respectively, indicating the initial stages of desertification and its intensification. In the 2005–2010 period, although the area of high vulnerability and very high vulnerability continued to expand, there were also signs of partial recovery, especially in the very high vulnerability regions, where area reduction was observed, reflecting the positive effects of ecological restoration and improved management. However, from 2010 to 2015, desertification once again intensified, with notable transitions from moderate vulnerability to high vulnerability and very high vulnerability regions. The area of very high vulnerability continued to expand, increasing by approximately 1.00 × 10⁵ km² during this period. Between 2015 and 2020, desertification pressure peaked, particularly in the transition from moderate vulnerability to high vulnerability and very high vulnerability regions, which reached a historical maximum of 1.12 × 10⁵ km². Although the net increase in very high vulnerability areas slowed in some regions, indicating some recovery, the overall trend of desertification expansion remained significant. Specifically, the low vulnerability areas showed almost no transition to very high vulnerability areas, suggesting that although recovery efforts have yielded some positive results, the overall desertification issue remains largely unaddressed.

In summary, the spatial transition patterns of desertification vulnerability in the Mongolian Plateau reveal a long-term trend of increasing desertification, particularly with the continuous expansion of high vulnerability and very high vulnerability areas. This highlights the profound impact of climate change and human activities on the ecosystem. However, the partial recovery and management improvements in certain regions demonstrate that appropriate ecological restoration measures can stabilize or recover some areas. Given the increasingly severe desertification situation, future efforts should focus on strengthening ecological protection and restoration, especially in high vulnerability and very high vulnerability areas, with effective governance strategies essential to halting the expansion of desertification in the context of climate change.

3.3. Analysis of Driving Factors of Desertification Vulnerability

3.3.1. Correlation Among Driving Factors

As shown in Figure 7, in the desertification vulnerability analysis of the Mongolian Plateau, correlation analysis of driving factors highlights the strong relationships between ecological changes and factors such as climate, soil conditions, land use, and population density. The correlation heatmap reveals significant interaction between climate factors (temperature and precipitation) and ecological changes, with a positive correlation between temperature and precipitation observed in 2000 and 2020 (correlation coefficient of 0.66). This indicates that climate change, particularly rising temperatures and fluctuating precipitation, is a major driver of ecological changes. Increased temperatures lead to higher evaporation rates, causing soil dryness, while reduced precipitation impacts vegetation water supply, destabilizing plant growth and ecosystems. Climate change acts as a catalyst for desertification, reducing vegetation cover and intensifying soil erosion. Soil factors, such as organic matter (OM) and sand content, show a significant positive correlation with NDVI, especially in 2005 and 2015, suggesting that soil quality significantly influences ecosystem health. Higher organic matter improves moisture retention and promotes vegetation growth, while increased sand content can degrade soil structure and accelerate desertification. These changes in soil conditions, in conjunction with climate change, contribute to a vicious cycle of desertification. Land use change and population density exhibit weaker correlations, though land use showed stronger correlations with NDVI and sand content in 2015. Expanding farmland and urbanization alter the ecological structure and hydrological conditions, intensifying soil erosion and reducing vegetation cover. Higher population density increases resource demand, leading to further land overuse and hindering ecosystem recovery. Topographic factors, such as DEM and slope, show weak correlations, indicating limited direct effects on desertification vulnerability. However, topography can influence vegetation indirectly by affecting water flow and soil erosion. Steep slopes promote water runoff, leading to soil degradation, while low-lying areas may encourage plant growth, though excessive water can cause soil salinization.

Overall, ecological changes on the Mongolian Plateau are driven by multiple factors, with climate change being the primary driver. Soil conditions and land use changes also contribute to desertification, while topographic factors have localized, indirect effects. To address desertification, integrated management focusing on climate regulation, soil conservation, and sustainable land use is essential for regional sustainability.

3.3.2. Identification of Key Drivers and SHAP-Based Interpretation

Figure 8 illustrates the temporal evolution of key drivers influencing desertification vulnerability across the Mongolian Plateau from 2000 to 2020. Climate-related variables, particularly temperature, wind speed, and precipitation, exerted the strongest influence. Temperature consistently ranked as one of the most critical drivers, with its impact peaking in 2005 (0.2668) and remaining significant in 2020 (0.2154). This variation is primarily attributed to temperature’s role in exacerbating evaporation, soil moisture depletion, and vegetation stress. Wind speed maintained its significant effect, particularly in 2000 (0.1853) and 2020 (0.1936), due to its close association with wind erosion processes. Precipitation had a substantial impact on desertification vulnerability in the early period (0.1809 in 2000), but its influence diminished post-2005, with a value of 0.1355 in 2020, indicating a reduced effect on soil moisture retention and vegetation growth. Temperature and wind speed, as primary climatic factors, played a dominant role in desertification dynamics. The rise in temperature not only intensified soil moisture loss but also exacerbated vegetation stress, particularly under drought and high-temperature conditions, thus significantly driving the progression of desertification. Wind speed, through facilitating wind erosion, made soil particles more susceptible to being displaced, leading to soil degradation and increased ecosystem vulnerability. The reduction in precipitation, especially after 2005, directly affected soil water supply, making it difficult for vegetation to thrive under arid conditions, thereby accelerating land degradation. Soil characteristics also played a significant role in shaping desertification vulnerability. Sand content reached its peak in 2015 (0.1885), while organic matter (OM) exhibited an increasing influence over time, with a value of 0.1242 in 2020. Areas with higher sand content exhibit poorer moisture retention and are more vulnerable to wind and water erosion, further exacerbating desertification. Vegetation indices, such as NDVI (e.g., 0.1185 in 2010) and NDSI (around 0.1287 in 2020), primarily reflect the ecosystem’s response to climatic and soil conditions, with their contribution gradually increasing over time. Topographic factors, such as slope, had a relatively minor influence, with values of 0.0229 in 2000 and 0.0304 in 2015. Similarly, land use and population density exhibited moderate effects, with land use contribution rising from 0.0326 in 2005 to 0.0620 in 2020, and population density decreasing from 0.0796 in 2000 to 0.0639 in 2020. These results underscore the dominant role of temperature, wind speed, and precipitation in desertification dynamics, while the significance of soil characteristics and vegetation indices also increased over time. To mitigate desertification risks, urgent implementation of climate-resilient strategies is necessary, including temperature regulation, the establishment of windbreaks, soil stabilization through sand fixation and organic matter enhancement, as well as vegetation restoration in highly vulnerable areas.

As illustrated in Figure 9, the SHAP analysis offers deeper insights into the complex relationships driving desertification vulnerability. Temperature emerges as a key factor, with elevated temperature values consistently aligning with high SHAP values, underscoring its significant positive contribution to vulnerability. Wind speed also plays a crucial role, exhibiting a strong correlation with increased SHAP values, particularly evident in the years 2005 and 2020, suggesting its exacerbating effect on land degradation during these periods. Precipitation, conversely, shows a negative relationship with SHAP values: areas experiencing low rainfall are associated with higher SHAP values, indicating that drought conditions exacerbate vulnerability. On the other hand, increased rainfall appears to mitigate the impact of desertification by reducing SHAP values. Sand content is another influential factor, with high levels correlating strongly with elevated SHAP values, further emphasizing its role in soil degradation and desertification processes. Organic matter (OM) shows an inverse relationship with SHAP values, with higher OM content contributing to lower SHAP values, reaffirming its role in promoting soil resilience and enhancing land productivity. The Normalized Difference Soil Index (NDSI) reveals a positive correlation with SHAP values, highlighting areas with exposed soil as degradation hotspots, where the absence of vegetation and protective ground cover leads to increased vulnerability. Interestingly, the Normalized Difference Vegetation Index (NDVI) demonstrates a negative relationship, where higher vegetation coverage correlates with lower SHAP values, indicating that healthy vegetation plays a protective role against desertification. Population density, particularly from 2015 onward, shows an increasing influence on desertification vulnerability. High-density areas are linked with higher SHAP values, suggesting the growing impact of anthropogenic pressures, such as overgrazing, land-use changes, and urban expansion, which contribute to land degradation. Topographic factors, including slope and Digital Elevation Model (DEM) values, show limited influence on desertification vulnerability, with SHAP distributions remaining relatively small and clustered. This suggests that while topography may influence local variations in land degradation, its effect is less pronounced compared to climate factors and human-induced pressures.

Overall, the analysis emphasizes that desertification vulnerability is driven by a complex interplay of climatic, ecological, and anthropogenic factors. Climate variables, particularly temperature and wind speed, are dominant contributors, while soil and vegetation act as vital buffers. However, human activities are emerging as significant threats to land stability, highlighting the need for targeted management strategies to mitigate the accelerating impacts of desertification.

3.4. Projection of Desertification Vulnerability Based on Climate and Soil Interactions

Based on analysis, temperature, precipitation, and soil sand content were identified as key drivers of desertification vulnerability. Using data from 2000 to 2020, this study projected vulnerability for 2030 under four SSP scenarios: SSP126, SSP245, SSP370, and SSP585 (Figure 10 and

Table 6. Desertification vulnerability areas (km²) across severity levels for 2020 and projected values under SSP scenarios for 2030.

Scenario	Low	Moderate	High	Very High
2020	202,984.92	648,072.18	844,236.86	1,00,9801.68
2030 SSP1-2.6	182,363.69	510,806.02	813,161.59	1,202,990.63
2030 SSP2-4.5	189,153.36	498,963.81	814,290.73	1,211,657.27
2030 SSP3-7.0	196,470.53	530,552.21	796,415.13	1,190,627.28
2030 SSP5-8.5	145,623.17	445,377.27	881,623.97	1,241,440.74

). The SSP scenarios assess the potential impacts of different socioeconomic development paths on climate change and the environment. SSP126 represents a path with significant progress in reducing greenhouse gas emissions, focusing on green energy and environmental protection, limiting global temperature rise to below 2 °C. SSP245 follows a more moderate approach to climate change mitigation, where emissions are reduced moderately, and global temperature rise may exceed 2 °C but is managed through adaptation measures. SSP370 depicts a fragmented world with limited global cooperation and rapid increases in emissions, leading to global temperatures rising by more than 3 °C. Finally, SSP585 represents a high-emission path driven by fossil fuels, where emissions remain high, causing significant climate change impacts and global temperatures rising by more than 4 °C. The model demonstrated good predictive performance (R² = 0.8947, RMSE = 0.1411). The results revealed increasing vulnerability and shifting spatial patterns from SSP126 to SSP585. Under SSP126, northern and northeastern regions maintain low to moderate vulnerability, while high-emission scenarios (SSP370, SSP585) show widespread high vulnerability, particularly in southern and transitional zones, with a noticeable south-to-north resilience gradient. Table 6 compares the structural distribution of vulnerability in 2030 to 2020. All scenarios show an expansion of vulnerable areas, with the Very High category increasing sharply—reaching 1.24 million km² under SSP585—indicating severe ecological degradation. Meanwhile, the Low and Moderate categories decline, indicating a loss of ecologically secure zones. These findings underscore the urgency of adopting low-carbon development pathways, enhancing land-use sustainability, and strengthening ecological restoration efforts.

As of 2020, the plateau exhibits significant vulnerability, with the very high vulnerability areas (10,098.02 km²) representing the largest portion of the affected land. Additionally, high vulnerability regions cover 844,236.86 km², underscoring the severe desertification challenges currently faced. Under the SSP1-2.6 scenario, which represents a low-carbon, green development pathway, there is a slight improvement, with low vulnerability areas increasing to 182,363.69 km². However, high vulnerability areas still expand to 1,202,990.63 km², reflecting the long-term impacts of climate change and environmental degradation, even under mitigation efforts. Similarly, the SSP2-4.5 scenario, which suggests moderate mitigation policies, results in a similar trend, with a slight improvement in low vulnerability regions (189,153.36 km²), but high vulnerability areas remain at 1,211,657.27 km². In contrast, the SSP3-7.0 scenario, characterized by rapid emission increases and limited global cooperation, exacerbates desertification, with high vulnerability areas reaching 1,190,627.28 km². Finally, the SSP5-8.5 scenario, which represents a high-emission and fossil-fuel-driven pathway, results in the most severe projections, with high vulnerability areas expanding further to 1,241,440.74 km², highlighting the significant threat of climate change and anthropogenic activities. These findings indicate that while low-carbon development scenarios can mitigate some desertification effects, high-emission pathways will exacerbate land degradation, especially in the southern and western regions, emphasizing the urgent need for targeted environmental policies and climate adaptation strategies.

4. Discussion

4.1. Ecological Interpretation of Driving Mechanisms

This study applied the XGBoost model combined with SHAP analysis to identify the key drivers of desertification vulnerability on the Mongolian Plateau. Temperature and wind speed were found to be the primary climatic factors, with average SHAP values ranging from 0.215 to 0.265 and 0.185 to 0.194, respectively, which is consistent with the findings of previous studies [51]. These results indicate an intensifying aridification trend closely linked to the strengthening Siberian high-pressure system. Elevated temperatures have increased evapotranspiration, reducing soil moisture availability and inhibiting vegetation growth [52]. Concurrently, intensified wind speeds have aggravated wind erosion, leading to an approximately 15% annual increase in sand transport, which degrades soil structure and vegetation cover [53]. This interaction creates a negative feedback loop: vegetation degradation accelerates soil deterioration, which further hampers vegetation recovery and exacerbates desertification. Soil properties, particularly sand content and organic matter, also significantly influence ecological vulnerability. Higher sand content correlates positively with increased degradation risk, whereas elevated organic matter improves soil stability and resilience. Vegetation health, represented by NDVI, negatively correlates with desertification risk, emphasizing its critical role as an ecological buffer.

4.2. Future Vulnerability Trends and Policy Implications

Forecasts under four Shared Socioeconomic Pathways (SSPs) scenarios for 2030 demonstrate that high-emission scenarios (SSP370 and SSP585) will substantially increase desertification vulnerability in central and southwestern regions, heightening ecological risks [54,55]. In contrast, the low-emission sustainable development pathway (SSP126) predicts relatively stable ecological conditions, illustrating the benefits of low-carbon and environmentally friendly development strategies [56]. The high predictive accuracy of the model (R² = 0.8837, RMSE = 0.1401) validates the effectiveness of integrating machine learning with remote sensing data to capture the complex drivers and spatiotemporal evolution of desertification. These findings underscore the urgency to implement climate-adaptive policies, improve land management, and promote ecological restoration [57]. Recommended actions include establishing shelterbelts to reduce wind erosion, adopting water-saving irrigation and vegetation restoration techniques to enhance water-use efficiency, regulating grazing practices scientifically to prevent overexploitation, and strengthening cross-border cooperation between Mongolia and Inner Mongolia for coordinated governance [58]. Pursuing sustainable, low-carbon development pathways is critical to mitigating climate impacts and preserving regional ecological security.

4.3. Limitations and Future Directions

Despite significant advances, this study has several limitations. Socioeconomic variables such as land-use change dynamics, policy enforcement intensity, and economic development levels were not fully integrated, which may limit the comprehensiveness of vulnerability assessments [59]. Additionally, future projections were primarily based on historical trends of climate and soil sand content, excluding possible effects of extreme events or abrupt policy changes. The spatial resolution and temporal coverage of the datasets constrained the capacity to detect fine-scale variations and micro-regional changes. Future research should focus on integrating multi-scale, multi-source data fusion leveraging GIS technologies to enable dynamic monitoring and precise early warning of desertification [60]. Incorporating social-ecological system dynamics into integrated modeling frameworks will further improve the understanding and prediction of complex desertification processes. Moreover, long-term monitoring of policy effectiveness and ecological restoration outcomes is essential to guide sustainable ecosystem management and land use planning in arid and semi-arid regions [34].

5. Conclusions

This study developed a comprehensive framework for assessing desertification vulnerability in the Mongolian Plateau, utilizing the MEDALUS model alongside multi-source remote sensing and geospatial data. The assessment incorporated four key dimensions: management, vegetation, soil, and climate quality. By applying the XGBoost algorithm and SHAP interpretability, the study identified the primary drivers of desertification and their interconnected feedback mechanisms. A regression model was subsequently built, using precipitation, temperature, and soil sand content to forecast future vulnerability under four SSP scenarios for 2030. Key findings include:

(1)

Spatiotemporal Variation in Vulnerability: From 2000 to 2020, desertification was more pronounced in the southern and western regions, with lower vulnerability in the north and east. While extreme vulnerability areas showed a slight reduction, high vulnerability zones expanded, underscoring the need for targeted control efforts. Positive trends were observed in Inner Mongolia, while Mongolia continued to face significant vulnerability.

(2)

Dominant Drivers: Climate factors—particularly temperature, wind speed, and precipitation—emerged as the most significant drivers, followed by soil characteristics and vegetation (NDVI). Human activities also increasingly influence vulnerability, emphasizing the importance of enhanced climate adaptation, soil management, and sustainable land-use practices.

(3)

Projected Risks: Future projections indicate that desertification vulnerability in the Mongolian Plateau will increase significantly under high-emission scenarios, such as SSP3-7.0 and SSP5-8.5, with high vulnerability areas expanding notably. In contrast, lower-emission pathways like SSP1-2.6 and SSP2-4.5 can mitigate some desertification impacts, though high vulnerability regions will persist. The results emphasize the critical need for effective climate mitigation and adaptation strategies to address the escalating desertification risks, particularly in the southern and western regions, where the threat of land degradation remains highest.

In conclusion, this study shows that integrating machine learning with remote sensing enhances the accuracy of desertification vulnerability assessments and provides valuable insights for planning and mitigation. By identifying key drivers such as climate, soil degradation, and human activities, the findings can guide targeted interventions and inform sustainable land-use practices. Incorporating socioeconomic data further improves future projections, supporting more effective desertification management and adaptation strategies.