Quantifying Grazing Intensity from Aboveground Biomass Differences Using Satellite Data and Machine Learning

Abstract

Accurately quantifying grazing intensity (GI) is crucial for assessing grassland utilization and supporting sustainable management. Traditional livestock-based approaches cannot capture the spatial heterogeneity of grazing or its dynamic response to climate variability. The objective of this study was to develop a remote sensing-based quantitative framework for estimating GI across the Inner Mongolian grasslands. The framework integrates MODIS vegetation indices, ERA5-Land climate variables, topographic factors, and field-measured data and GI was quantified as the proportional difference between potential and satellite-derived aboveground biomass (AGB), providing a spatially explicit measure of forage utilization. In this framework, potential AGB (AGBp) represents the climate-driven growth capacity under ungrazed conditions reconstructed using machine learning models, whereas satellite-derived AGB (AGBs) denotes the standing AGB remaining under current grazing pressure. Validation using 324 paired grazed–ungrazed plots demonstrated strong agreement between modeled and observed GI (R² = 0.65, RMSE = 0.18). This AGB-difference-based approach provides an effective and scalable tool for large-scale rangeland monitoring, offering quantitative insights into grass–livestock balance, ecological restoration, and adaptive management in arid and semi-arid regions.

1. Introduction

Globally, grasslands constitute one of the most extensive terrestrial biomes, covering more than one-third of the Earth’s land surface and supporting over 1.4 billion ruminants [1,2,3,4]. By sustaining forage production and regulating biogeochemical cycles, grasslands not only secure global food and fiber supply but also deliver essential ecosystem services [5]. These include provisioning services such as forage and livestock products and regulating and supporting services that enhance soil carbon storage, water conservation, and biodiversity maintenance, together underpinning ecological stability in arid and semi-arid regions [2,6,7]. In China’s Inner Mongolia, grasslands carry additional strategic significance: they underpin pastoral livelihoods while serving as a key ecological barrier that safeguards northern China from desertification and climate risks [8,9,10,11]. Yet these ecosystems are increasingly threatened by the combined impacts of intensive grazing and ongoing climate change, particularly in semi-arid steppe regions, where degradation processes unfold rapidly and ecosystem functions are severely compromised [6,12,13]. This context highlights the urgent need for precise, spatially explicit assessments of GI and its temporal evolution, which are critical for maintaining a balance between forage supply and livestock demand, promoting ecological restoration, and guiding sustainable rangeland governance in northern China [14,15].

Traditional approaches to GI assessment have typically relied on livestock census data, assuming that animal numbers directly translate into grazing pressure [16,17,18]. However, this assumption has multiple flaws. First, livestock statistics seldom differentiate between stall-fed animals and those grazing freely on natural pastures, which weakens the ecological relevance of census-based indicators and produces nonlinear or distorted relationships between animal counts and grassland utilization [19,20]. Second, such statistics are usually aggregated at county or prefectural scales and reported annually, resulting in poor spatial granularity and delayed availability [21,22]. Third, management practices such as seasonal rest grazing, rotational systems, and grazing bans vary widely across regions [21,23], further decoupling livestock inventories from the actual pressure exerted on the land. These limitations highlight the need for alternatives that directly capture the realized utilization of forage resources rather than relying solely on animal numbers. In addition, some studies have employed GPS-based livestock tracking to estimate GI from animal movement data [22,24]. This method can directly record grazing trajectories and durations, providing fine-scale information on pasture use. However, its application remains limited due to high equipment costs, battery endurance, device maintenance requirements, and potential impacts on animal health [18,24]. Overall, while GPS tracking improves the ground-level quantitative accuracy of GI estimation, it still faces major constraints in conducting long-term, large-scale, and operational grazing monitoring.

Remote sensing offers a powerful solution because it provides consistent, large-scale, and temporally frequent observations of vegetation status [25,26,27,28]. Yet, vegetation indices (VIs) derived from satellite imagery integrate both climatic influences and grazing effects [15,19,29,30], making it difficult to isolate the grazing signal. To address this, researchers have proposed reconstructing potential vegetation indices (VIp) that represent ungrazed, climate-driven productivity [31]. A persistent challenge, however, is that numerous vegetation indices exist, and it remains unclear which most reliably reflects baseline productivity under ungrazed conditions [32].

Machine learning (ML) techniques have increasingly been applied to grassland monitoring to disentangle climate-driven variability from grazing effects [33,34]. Algorithms such as random forests, boosting trees, and deep neural networks can capture nonlinear vegetation–climate relationships more effectively than traditional regression methods [25,29,35]. Nevertheless, these models differ in their interpretability, computational demands, and sensitivity to data structure [25,36]. No consensus yet exists on which ML approach is best suited for reconstructing VIp or for deriving GI consistently across large regions [25,37]. Understanding the strengths, limitations, and performance boundaries of different machine learning algorithms is therefore essential for reliably separating climatic influences from grazing impacts and for enabling high-resolution monitoring of GI. Another factor influencing model performance is the choice of spatial stratification. Subdividing training data by ecological zones or administrative units can reduce parameter heterogeneity, improve predictive stability, and enhance ecological interpretability [25,38,39]. If zoning is too coarse, critical differences among ecosystems may be lost, while overly fine partitions can lead to sample sparsity and overfitting. However, the relative merits of different subdivision strategies for reconstructing VIp and estimating GI remain underexplored, limiting progress toward spatially adaptive monitoring [40].

Inner Mongolia, one of China’s most important pastoral regions, faces accelerating degradation due to fragile ecosystems and long-term grazing pressure [9]. Yet a high-resolution, long-term dataset of GI is still lacking, constraining both ecological research and management decision-making. This study addresses these gaps by proposing a remote sensing-driven framework that defines GI as the proportional difference between AGBp and AGBs. Unlike census-based methods, this definition directly reflects the actual utilization of grassland production. Specifically, our objectives are to achieve the following:

• Identify the optimal combination of VIs, ML algorithms, and ecological zoning strategies for reconstructing VIp.
• Derive AGBp and AGBs from 2000 to 2024 and calculate GI as their proportional difference.
• Analyze the spatiotemporal dynamics of GI across Inner Mongolia and evaluate its implications for grass–livestock balance and sustainable rangeland management.

By combining machine learning, remote sensing, and field validation, this work delivers a high-resolution, operational, and cost-effective dataset for monitoring GI, offering new insights and practical tools for grassland conservation and governance in northern China

2. Materials and Methods

2.1. Study Area

The study was conducted across the grasslands of Inner Mongolia, northern China (Figure 1), a region shaped by a temperate continental monsoon climate [11]. Mean annual temperature (MAT) decreases from about 9 °C in the arid southwest to nearly −5 °C in the humid northeast, while mean annual precipitation (MAP) ranges from 40 to 580 mm, with more than 80% concentrated between May and October. The landscape is dominated by plateau terrain—including the Ordos, Xilingol, and Hulunbuir plateaus—generally rising to around 1000 m above sea level.

Vegetation composition reflects the strong precipitation gradient and can be broadly classified into three steppe types from west to east. The desert steppe, occurring under minimal rainfall, supports sparse vegetation and is highly vulnerable to degradation. The typical steppe in the central zone is dominated by Leymus chinensis and Stipa grandis, while the more humid meadow steppe in the east is characterized by species such as Stipa baicalensis and Filifolium sibiricum. As one of China’s largest pastoral regions, Inner Mongolia sustains intensive livestock grazing, mainly by sheep and cattle [22]. Inner Mongolia’s grassland husbandry is dominated by extensive free grazing, but the system has undergone substantial institutional change since the 1980s [41,42]. Following the household contract reform of grassland tenure, most pastures were divided and fenced at the household level, which has largely restricted the movement of wild herbivores except in border or protected areas. Since the enactment of the Grassland Law in 1985, reclamation and conversion of grasslands for cropland, mining, or other land uses have been strictly prohibited, ensuring that grazing and hay harvesting remain the primary modes of land utilization [42]. To avoid early mowing that could reduce seed availability and hinder vegetation regeneration, hay harvesting in most parts of Inner Mongolia is usually concentrated between mid- and late August each year [42]. Because our study focuses on vegetation conditions in mid-August, before the main mowing period, it can be reasonably assumed that grazing is the dominant factor driving vegetation changes in Inner Mongolia.

2.2. Data Usage

2.2.1. Satellite Data

The MODIS 8-day surface reflectance product (MOD09A1) was obtained via the Google Earth Engine (GEE) platform (Google LLC, Mountain View, CA, USA)., which has three reflectance bands, blue (459−479 nm), red (620−670 nm), and near-infrared (841−876 nm), at a spatial resolution of 500 m, from 2000 to 2024 were used for this study. These bands were used to calculate vegetation indices required for model input, as detailed in

Table 1. Summary of 17 driving variables from five categories used for VIp reconstruction and AGB conversion across Inner Mongolia.

Category (Resolution)	Variable (Unit)	Full Name/Definition	Description/ Ecological Meaning	Used in the Modelling Module
Satellite data (500 m)	NDVI [43]	Normalized difference vegetation index	Reflects vegetation greenness and photosynthetic activity	Reconstruction of VIp/AGB conversion
	EVI [44]	Enhanced vegetation index	Enhances vegetation signal by reducing atmospheric and canopy background noise	Reconstruction of VIp/AGB conversion
	EVI2 [45]	2-band enhanced vegetation index	Simplified two-band version of EVI, without blue band dependence	Reconstruction of VIp/AGB conversion
	SAVI [46]	Soil adjusted vegetation index	Minimizes soil brightness effects in sparsely vegetated areas	Reconstruction of VIp/AGB conversion
	NIRV [47]	Near-infrared reflectance of vegetation	Represents vegetation productivity using near-infrared reflectance	Reconstruction of VIp/AGB conversion
Regional subareas	Group [48]	8 sub-areas	Classified according to the Grassland Classification standard as the first-level grassland category, used for large-scale ecological zoning	Reconstruction of VIp
	Subgroup [48]	20 sub-areas	Second-level division under the Grassland Classification standard, representing finer ecological differentiation within major zones	Reconstruction of VIp
	Type [48]	224 sub-areas	Third-level grassland unit defined by vegetation composition according to the Grassland Classification standard	Reconstruction of VIp
	County-level	108 sub-areas	Administrative divisions used for statistical aggregation and spatial validation.	Reconstruction of VIp
Climatic factors (~10 km)	P (mm)	Annual precipitation from 2000 to 2024	Mean temperature	Reconstruction of VIp
	T (°C)	Annual temperature from 2000 to 2024	Cumulative precipitation	Reconstruction of VIp
	MAP (mm)	Mean annual precipitation from 1980 to 1999	Long-term average precipitation, representing regional water availability	AGB Conversion
	MAT (°C)	Mean annual temperature from 1980 to 1999	Long-term average temperature, representing regional thermal conditions	AGB Conversion
	AP (mm)	Annual precipitation anomaly from 2000 to 2024	Represents interannual variability of precipitation	AGB Conversion
	AT (°C)	Annual temperature anomaly from 2000 to 2024	Represents interannual variability of temperature	AGB Conversion
Topographic factors (30 m)	DEM (m)	Elevation derived from SRTM 30 m	Represents terrain and microclimate variation	AGB Conversion
Climate-driven metrics (500 m)	VIp	Potential vegetation index	Derived from climatic drivers to represent ungrazed vegetation conditions	AGB Conversion

Note: The VIs listed under satellite data served as predictors for both the reconstruction of the VIp and the AGB conversion. The “Used in modelling module” column corresponds to the two subsequent methodological components described in Section 2.3 (VIp reconstruction and AGB conversion), indicating where each variable was applied within the overall modelling framework. The VIp, reconstructed from climatic and topographic drivers under ungrazed conditions, represents an intermediate product linking these two steps.

. To ensure data quality and temporal consistency, invalid pixels were identified using the MOD09A1 quality control (QC) layer. Missing values were then gap-filled using valid observations within a ±48-day temporal window.

2.2.2. Regional Subareas Data

The data on grassland group, subgroup, and type reflect the spatial distribution patterns and structural composition of grassland resources in Inner Mongolia. Classification was based on a multi-factor approach integrating vegetation characteristics and habitat conditions, incorporating variables such as vegetation type, hydrothermal regimes, and topographic features (Figure 1, Table 1). The classification follows a minimum mapping unit of 4 m², and the dataset was provided by the Inner Mongolia Forestry and Grassland Monitoring and Planning Institute.

2.2.3. Climatic Data

Climate data were derived from the ERA5-Land daily data generated by the Copernicus Climate Change Service at the European Centre for Medium-Range Weather Forecasts (https://cds.climate.copernicus.eu/, accessed on 15 October 2024). This data product provides daily meteorological data on a 0.1° (~10 km) latitude/longitude climate modeling grid. The band temperature_2m and total_precipitation_sum from 1980 to 2024 were used for this study. These daily records were then used to calculate the precipitation, temperature, MAP, MAT, precipitation anomaly (AP) and temperature anomaly (AT) over the 20-year period (1980–1999).

2.2.4. Topographical Data

Topographical data with a spatial resolution of 30 m, obtained from the National Aeronautics and Space Administration (NASA) (https://earthdata.nasa.gov/), were resampled to 500 m resolution using ArcGIS 10.8 software for further analysis.

2.2.5. Ground Observation Data

Ground observation datasets were used to support model training and validation, including AGB measurements, GI quadrats, and ungrazed reference stations (Figure 1 and Figure 2,

Table 2. Summary of ground observation datasets used for VIp reconstruction, AGB conversion, and GI estimation in this study.

Type	Sampling Method	Number of Plots	Primary Purpose
AGB measurements	All herbaceous biomass was harvested in plots consisting of 5 quadrats of 0.1 m2 each, dried, weighted, and averaged.	2789	AGB Conversion
GI quadrats	Paired grazed and ungrazed plots (<1 km apart, each 10 × 10 m) were established to compare vegetation residuals. Grazing intensity levels were determined based on relative reductions in vegetation cover, grass height, edible forage, and residual biomass.	648 (324 paired grazed and ungrazed quadrats)	Estimation of GI
Ungrazed reference stations	Long-term fenced plots established by the Forestry and Grassland Monitoring and Planning Institute of Inner Mongolia, excluding livestock grazing to maintain natural vegetation conditions.	76	Reconstruction of VIp

Note: The three ground observation datasets correspond to the main methodological components described in Section 2.3, namely the reconstruction of VIp, AGB conversion, and GI estimation.

).

The measurements of AGB were derived from the National Inventory of Grassland Resources by the National Forestry and Grassland Administration of Inner Mongolia from 2020 to 2024 (http://www.forestry.gov.cn), and a total of 2789 field plots were collected during the peak growing season (July–August) (Figure 1). The sampling followed a stratified design, with plots evenly distributed across major grassland types to ensure representative coverage of vegetation and environmental gradients. Each plot was treated as an independent sample for analysis. The AGB was measured by harvesting the aboveground portion of the plants and oven-drying them to obtain the dry mass weight.

The GI quadrats were collected from grazed and ungrazed plots by comparing vegetation residuals within adjacent areas (<1 km apart), and 648 field plots were surveyed during the peak growing season (July–August) (Figure 2). The different GI levels of adjacent samples were primarily driven by grazing activities, as neighboring plots typically had similar climates and environments. The enclosed plots without livestock grazing were selected as the ungrazed samples. Then, using the ungrazed samples as benchmarks in paired sampling, the GI of other plots was determined based on the reduction rates of vegetation coverage, grass height, edible forage, and residual biomass [49]. Following the national standard Evaluating Criterion for Balance of Forage Supply and Livestock Requirement (LY/T 3320–2022) issued by the National Forestry and Grassland Administration, GI levels were classified as light, moderate, and heavy according to the ratio of grazed to ungrazed values for vegetation cover, height, and AGB, with reductions of <20%, 21–30%, and >30%, respectively. When the three indicators indicated different grades, the AGB-based level was adopted as the representative GI [49].

These evaluation criteria based on vegetation residuals are unaffected by grassland type and climate zone and represent a more practical approach than that based only on livestock number for grassland management. This approach assumes that differences in vegetation residuals between paired grazed and ungrazed plots are primarily caused by livestock grazing rather than by vegetation type or climatic gradients. Because each paired plot was located within a short distance (<1 km) and shared similar topography, soil, and climatic conditions, the effect of regional grassland types and climate zones on the residual-based evaluation criteria was minimized. Therefore, the derived GI mainly reflects grazing-induced vegetation reduction rather than environmental variability.

The ungrazed reference stations were established by the Forestry and Grassland Monitoring and Planning Institute of Inner Mongolia and are evenly distributed across the region’s major grassland types (Figure 2). Most of the plots were set up around 2020, with an area of 2−3.3 hectares each. Grazing has been strictly excluded within these sites, making them representative of ungrazed vegetation conditions.

2.3. Research Methods

To quantify GI from an ecological perspective, this study developed an integrated framework that combines multi-source remote sensing, climate reanalysis, topographic information, and field-based observations (Figure 3). The framework utilizes both VIs derived from MODIS imagery and reconstructed VIp representing ungrazed and climate-driven vegetation baselines. It comprises three main components: (1) reconstruction of VIp using climate and topographic predictors, (2) machine-learning-based estimation of AGB (AGBp and AGBs) under grazing-free and grazed conditions, and (3) calculation of GI as a dimensionless indicator of vegetation utilization, followed by spatiotemporal trend analysis. Unlike conventional livestock-based approaches, this framework explicitly links vegetation productivity with climatic and topographic constraints, enabling ecologically meaningful and spatially explicit quantification of GI across heterogeneous grassland environments.

2.3.1. Reconstruction of VIp

VIp representing ungrazed and climate-driven vegetation states, were reconstructed to describe the potential vegetation productivity unaffected by grazing disturbances. The reconstruction was based on long-term ungrazed reference plots (n = 76; Figure 2) that serve as ecological baselines. MODIS-derived vegetation indices (NDVI, EVI, EVI2, SAVI, and NIRv) were combined with concurrent temperature and accumulated precipitation over multiple lag periods (0−72 days) extracted from the ERA5-Land dataset to capture delayed vegetation responses to climatic drivers.

To model the climate–vegetation relationships, six machine-learning algorithms—Random Forest (RF), Gaussian Process Regression (GPR), Least Squares Boosting (LSBoost), Support Vector Machine (SVM), Multiple Linear Regression (MLR), and Elastic Net Regression (ENET)—were evaluated and compared (

Table 3. Summary of Machine Learning Algorithms.

Algorithms (Acronym)	Benefits	Limitations	References
Random Forest (RF)	Robust performance, automatically evaluates variable importance, less prone to overfitting	Less interpretable than linear models, high computational demands	[10]
Gaussian Process Regression (GPR)	Provides uncertainty estimates, suitable for small-sample, high-dimensional data	High computational complexity, not suitable for large datasets	[50]
Least-Squares Boosting (LSBOOST)	Integrates multiple weak learners, high accuracy, adjustable learning rate	High accuracy but computationally intensive, sensitive to parameter tuning	[51]
Multiple Linear Regression (MLR)	Simple model, interpretable coefficients, suitable for strongly linear data	Cannot model complex nonlinear relationships	[52]
Support Vector Machine (SVM)	Handles both linear and nonlinear data, effective for high-dimensional datasets	Sensitive to kernel and parameter selection, relatively high training cost	[53]
Elastic Net Regression (ENET)	Effective for high-dimensional data and correlated predictors	Less effective for capturing complex nonlinear patterns	[54]

). These models were trained under multiple ecological stratifications, including grassland class, subclass, type, administrative division, and pixel-level domains, to account for spatial heterogeneity in vegetation–climate interactions.

Model training adopted both hold-out (70/30 split) and five-fold cross-validation schemes, using R² and RMSE as primary evaluation metrics. For each subregion, the model with the best performance was selected to generate spatially continuous VIp predictions. This multi-model and stratified approach ensured that the reconstructed VIp accurately reflected potential vegetation productivity across diverse climatic and ecological gradients, providing a consistent ungrazed baseline for subsequent AGB modeling.

2.3.2. AGB Conversion

AGB was estimated through a unified RF regression framework designed to capture the nonlinear response of vegetation productivity to climatic and topographic factors. The model integrated vegetation indices, climatic anomalies (AT, AP), long-term climate means (MAT, MAP), and topographic variables derived from the digital terrain model (DTM). Both DEM and slope were initially tested as potential predictors. However, variable importance analysis (Figure S1) showed that DEM contributed substantially more than slope, and the two variables exhibited strong collinearity. Therefore, only DEM was retained in the final model to avoid redundancy and ensure model stability.

Two biomass states were simulated independently:

(1)

AGBp—potential biomass estimated from reconstructed VIp, representing vegetation growth capacity under ungrazed and climate-constrained conditions;

(2)

AGBs—remaining biomass derived from observed VIs, representing vegetation conditions after grazing disturbance.

The model formulation can be expressed as:

$$\mathrm{AGB} = \mathrm{ML}(\mathrm{AT}, \mathrm{AP}, \mathrm{MAT}, \mathrm{MAP}, \mathrm{VI}, \mathrm{DEM})$$

(1)

where ML denotes the machine learning-based Random Forest regression model; AT represents temperature anomaly; AP represents precipitation anomaly; MAT is mean annual temperature; MAP is mean annual precipitation; DEM represents elevation derived from topographic data.

AGB model training employed a combination of holdout validation and 5-fold cross-validation. Specifically, 10% of the samples were reserved for independent validation, while the remaining 90% were used for iterative training [25]. Within each iteration, four out of five data subsets were used for training and the fifth for validation, rotating through all subsets. This procedure was repeated four times, and average validation errors were computed across both holdout and cross-validation datasets. Finally, given the availability of ground-truth AGB data, the model was retrained on the entire dataset using 5-fold cross-validation to produce the final AGB upscaling models.

2.3.3. Grazing Intensity Estimation and Spatiotemporal Analysis

GI was defined as a dimensionless ratio between AGBp and AGBs. GI ranges from 0 to 1 and reflects the extent to which grassland productivity has been utilized. The GI value of 0 indicates no grazing, where the retention biomass equals the capacity biomass (AGBs = AGBp). The GI value of 1 represents complete vegetation removal, where the retention biomass is zero (AGBs = 0). Intermediate values indicate varying levels of utilization, with values closer to 0 reflecting low GI and values closer to 1 indicating high GI.

GI is calculated as follows:

$$\mathrm{GI}=\left({\displaystyle \frac{\mathrm{AGBp}-\mathrm{AGBs}}{\mathrm{AGBp}}}\right)$$

(2)

This formulation quantifies the proportion of vegetation consumed and serves as a standardized measure for assessing grazing pressure across spatial and temporal gradients.

To describe the overall temporal evolution of GI, an ordinary least squares (OLS) regression was applied to the regional mean GI time series (2000−2024), with year as the independent variable. The regression slope represented the rate of change over time. The Hurst exponent (H) was further calculated to evaluate the long-term persistence of the series. H values close to 0.5 indicated weak temporal dependence, suggesting that autocorrelation had little influence on the OLS-derived trend.

To characterize the spatially distributed temporal trends, the non-parametric Theil–Sen median slope estimator and Mann–Kendall test were applied at the pixel level to quantify the magnitude, direction, and significance of GI changes. This combined approach enables robust detection of both significant increases and decreases in GI while minimizing the influence of non-normality, outliers, and potential temporal dependence.

2.3.4. Validation and Accuracy Assessment

The reliability of remote sensing-derived GI was evaluated using 648 GI plots collected during the peak growing season. Field-based GI was calculated as the relative difference in AGB between paired plots. Model estimates were compared against these field values, with R², RMSE, and MSE used as evaluation metrics. Additionally, Pearson’s correlation coefficients were calculated to assess feature–response relationships. Higher R² and lower error values were interpreted as improved model performance. The formulas for these evaluation metrics are as follows:

$${\mathrm{R}}^2=1-{\displaystyle \frac{{\sum _{i=1}^n\left(\widehat{y_i}-y_i\right)}^2}{{\sum _{i=1}^n\left(y_i-\overline{y}\right)}^2}}$$

(3)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{{\sum _{i=1}^n\left(\widehat{y_i}-y_i\right)}^2}{n}}}$$

(4)

where $y_i$ is the observed value, $\widehat{y_i}$ is the predicted value, $\overline{y}$ is the mean of the observed values, and n is the sample size.

3. Results

3.1. Importance of Driving Variables

Five commonly used vegetation indices—NDVI, EVI, EVI2, NIRv, and SAVI—were reconstructed under grazing-free conditions to evaluate the influence of different climate driver combinations on model performance. The candidate predictor variables included multi-scale lagged precipitation and concurrent air temperature.

The five vegetation indices demonstrated varying sensitivities to these climatic drivers. NDVI achieved the highest reconstruction accuracy (R² = 0.545) when driven by a combination of 32-day lagged precipitation and concurrent temperature. EVI2 and SAVI performed slightly less accurately, while EVI showed the weakest performance.

Considering both model robustness and consistency in input–response relationships, potential NDVI (NDVIp) was selected as the core vegetation index for estimating GI. The optimal configuration was based on 32-day lagged precipitation and concurrent temperature (Figure 4).

3.2. Performance of the ML Models

Six machine learning algorithms—RF, GPR, LSBoost, SVM, MLR, and ENET—were evaluated using the top three combinations of climate drivers for NDVIp reconstruction. As shown in Figure 5, RF achieved the highest predictive accuracy (R² = 0.545) and the lowest RMSE, indicating strong model performance in capturing vegetation dynamics under ungrazed conditions. GPR (R² = 0.544) and SVM (R² = 0.541) performed comparably, albeit with slightly higher errors. In contrast, LSBoost, MLR, and ENet exhibited reduced accuracy (R² < 0.49) and greater variability across samples. Among the three algorithms with comparable predictive accuracy (RF: R² = 0.545, GPR: R² = 0.544, and SVM: R² = 0.541), RF demonstrated substantially higher computational efficiency, with training time reduced by more than half compared with GPR and SVM. Considering both accuracy and efficiency, RF was therefore selected as the foundational model for subsequent grazing-intensity estimation.

3.3. Effect of Spatial Stratification on Model Performance

To assess the influence of ecological and administrative stratification on model accuracy, five spatial zoning schemes were compared: pixel-level, vegetation group, subgroup, type, and county-level divisions. As shown in Figure 6a, stratification significantly improved the NDVIp model’s predictive accuracy across all schemes. Among them, type-level stratification yielded the highest coefficient of determination (R² = 0.864) and the lowest RMSE, followed closely by county-level (R² = 0.840) and subgroup-level (R² = 0.751) partitions. In contrast, pixel-based modeling performed the worst (R² = 0.545), likely due to high variability and sample sparsity. The performance ranking of zoning strategies was: type > county > subgroup > group > pixel. These results suggest that aligning model training with ecologically homogeneous units (e.g., vegetation types) enhances generalization by reducing within-group variability. As shown in Figure 6b, type-level stratification also minimized prediction error dispersion, further demonstrating its superiority for spatially adaptive modeling.

3.4. Quantitative Estimation of Grazing Intensity

Using the optimal driver combination—concurrent temperature and 32d lagged precipitation—identified by the Random Forest model stratified by type, NDVIp was reconstructed and integrated into a biomass model to simulate AGBp and AGBs. The model achieved R² values of 0.76 for AGBs and 0.69 for AGBp. The derived GI yielded a validation R² of 0.65 against observed values, with no systematic bias and uniformly distributed residuals, confirming its robustness for spatiotemporal analysis (Figure 7d).

In terms of spatial patterns, AGBs showed a more moderate gradient, with notable reductions in central and eastern regions due to grazing pressure. In contrast, AGBp exhibited a clear west-to-east increasing trend, with the highest values (>200 g·m⁻²) concentrated in the northeastern meadow steppes and forest-steppe transition zones. The spatial distribution of GI revealed relatively low intensity (<0.3) in the western desert steppes, moderate levels (0.3–0.5) in central typical steppes, and high GI (>0.6) primarily in the southeastern agro-pastoral ecotones. These patterns highlight the spatial heterogeneity of grazing impacts (Figure 7a–c).

3.5. Characterization of the Spatiotemporal Evolution of the Grassland Grazing Intensity

The overall declining trend of annual GI from 2000 to 2024 was well fitted by the OLS regression, and the Hurst exponent (H ≈ 0.52) suggested weak persistence, indicating that temporal autocorrelation had little influence on the observed trend. From 2000 to 2024, the overall GI in Inner Mongolia grasslands exhibited a declining trend, with an average annual decrease of approximately 1.2% (Figure 8a). A clear structural breakpoint was detected in 2011, after which GI values significantly dropped and stabilized at a relatively lower level, indicating a shift in grazing management or climate constraints.

Given that each pixel’s temporal trend was estimated independently using the Theil–Sen and Mann–Kendall tests, and that ecological stratification by grassland type was applied to reduce spatial heterogeneity, spatial autocorrelation is expected to have minimal influence on the robustness of the results. Spatially, the central and western regions showed the most substantial reductions in cumulative GI, particularly in previously overgrazed areas, reflecting the effect of targeted restoration programs and policy interventions. In contrast, slight increases in GI were observed in some northeastern and transitional zones, though these changes remained within a limited range (Figure 8b). Across the east–west gradient, GI exhibited distinct regional characteristics. The western region, dominated by desert steppes, showed relatively minor changes, while the central region—mainly typical steppe—experienced a pronounced decrease. In contrast, the eastern region, characterized by meadow steppes, maintained relatively low GI levels with limited variation.

4. Discussion

4.1. Spatial Heterogeneity of Grazing Intensities

In this study, we developed a dynamic framework for estimating GI by integrating remotely sensed NDVI data with field-measured AGB from both grazed plots and long-term ungrazed reference sites. In the study area, pastures have been fenced and contracted to households since the 1980s, grassland conversion has been legally prohibited since 1985, and hay mowing typically occurs in mid- to late August [42]. These management measures restrict the movement of wild herbivores, prevent large-scale land-use changes such as reclamation or mining, and thereby minimize non-grazing disturbances to vegetation dynamics. We therefore evaluated vegetation in mid-August, before the main mowing period, so that observed differences most likely reflect livestock grazing, while potential contributions from wild herbivores or other land-use disturbances are minimized. By incorporating a diverse set of ecological drivers, including vegetation type, climatic variables, and topographic features, our approach yields a more ecologically grounded and spatially explicit proxy for grazing pressure compared to traditional livestock-based assessments. In general, the multiyear average GI on Inner Mongolia increased from east to west during 2000 to 2024, with broad spatial heterogeneity. The average GI in Inner Mongolia from 2000 to 2024 was 0.15, suggesting that approximately 15% of AGB production was consumed by livestock annually during the growing season. Spatially, GI values were higher in the western desert steppe and eastern meadow steppe regions, and lower in the central typical steppe, revealing substantial regional heterogeneity. Temporal analysis revealed a weak but consistent declining trend (slope = −0.0025). This trend aligns with prior findings: [49] reported a decreasing pattern of GI based on grazing probability from 2015 to 2020, and [55] observed a similar long-term decline (2000−2022) using livestock statistics adjusted to reflect pressure on natural pastures (Figure 9). Although the methods differ in their definitions and magnitudes of GI, the convergence in trend underscores the robustness of our approach for characterizing grazing dynamics. Notably, a marked decline in GI was observed in 2011, which coincided with the initial implementation of Inner Mongolia’s ecological subsidy and grazing exclusion programs. This temporal correspondence suggests that these policies may have contributed to the reduction in grazing pressure, consistent with previous findings that reported similar policy effects in the region [56].

Overall, this study provides an accurate and spatially explicit assessment of GI by integrating biophysical variables and remote sensing indicators. The findings contribute to improved grassland monitoring, inform adaptive grazing management, and offer a scientific basis for policy-making in grassland conservation and restoration, thereby supporting the sustainable development of regional ecosystems.

4.2. Future Research Directions and Management Implications for Sustainable Grassland Development

This study generated a region-wide quantitative map of GI; however, the current spatial granularity is still insufficient to resolve micro-scale degradation patches or real-time livestock distribution [57]. Future work should therefore integrate high-resolution platforms—such as low-altitude UAVs and airborne LiDAR—with real-time animal-tracking technologies (e.g., GNSS collars and LoRa IoT tags) to create a “space–air–ground” multi-scale monitoring network. Coupling vegetation productivity and livestock activity through spatiotemporal data-assimilation algorithms would enable real-time identification of microscale overgrazed patches and grazing pathways, thereby providing the data foundation for precision management [16].

To promote sustainable utilization and fine-scale management of grassland resources, we propose a dynamic grass–livestock balancing strategy that leverages “grazing duration” as a regulatory lever. By combining the present grazing-intensity maps with industry or local overload thresholds, overstocked areas and their magnitudes can be spatially and quantitatively identified, facilitating targeted destocking. Recognizing that, under market conditions, destocking often takes the form of stall feeding rather than animal sale, we further convert the required destocking amount into an equivalent stall-feeding duration, enabling herders to estimate necessary fodder reserves and enhancing the operational feasibility and forward-looking nature of the management plan.

Overall, the proposed framework bridges the gap between ecological monitoring and operational management by enabling high-resolution, temporally explicit assessments of grazing pressure. Through the integration of remote sensing, machine learning, and ground-based validation, this approach offers a scalable and transferable tool to support sustainable grassland governance under changing climatic and socioeconomic conditions.

5. Conclusions

This study proposes an integrated, data-driven framework that combines remote sensing, climatic records, ground-based observations, and machine learning algorithms to generate high-resolution GI estimates across the Inner Mongolian grasslands. By reconstructing ungrazed vegetation baselines and quantifying the potential and observed AGB, the model enables monthly assessments of GI at a 500 m spatial resolution from 2000 to 2024—offering unprecedented insight into the spatiotemporal dynamics of grassland utilization. The results demonstrate strong agreement with field data and effectively capture the spatial heterogeneity and seasonal variability of grazing pressure. Compared to existing livestock census-based GI products, our model provides significantly improved accuracy, ecological relevance, and near-real-time monitoring capacity. These advances underscore the framework’s value as a scalable and transferable tool for rangeland monitoring, adaptive management, and early-warning systems. While the model exhibits strong predictive performance (R² = 0.65), uncertainties persist—particularly with respect to meteorological variability and biomass response modeling—highlighting the need for further refinement. Nonetheless, the proposed approach contributes a robust foundation for spatializing GI and aligning grassland governance with principles of ecological sustainability, climate resilience, and food security. As grazing remains the dominant land use in arid and semi-arid ecosystems, high-resolution GI monitoring will be vital for supporting sustainable land management, ecological compensation schemes, and policy-making at the interface of agriculture and environmental protection.