Threshold Extraction and Early Warning of Key Ecological Factors for Grassland Degradation Risk

Abstract

Grassland degradation poses a serious threat to ecosystem stability and the sustainable development of human societies. In this study, we propose a framework for grassland degradation risk assessments and early warning based on key ecological factors (KEFs) in Xilingol. The NDVI, NPP, and grass yield were selected as KEFs to represent vegetation coverage, ecosystem productivity, and actual biomass, respectively. By constructing a grassland degradation index (GDI) and integrating K-means clustering, the average curvature, and a gravity center shift analysis, we quantified the degradation risk levels and identified the threshold values for different grassland types. The results showed the following: (1) the grass yield was the most sensitive indicator of grassland degradation in Xilingol, with high-risk thresholds decreasing from 115.67 g·m⁻² in the temperate meadow steppes (TMSs) to 73.27 g·m⁻² in the temperate typical steppes (TTSs), and further to 32.30 g·m⁻² in the temperate desert steppes (TDSs); (2) the TDSs exhibited the highest curvature value (2.81 × 10⁻⁴) in the initial stage, indicating a higher likelihood of rapid early-stage degradation, whereas the TMSs and TTSs reached peak curvature in the latest stages; and (3) the TTSs had the largest proportion of high-risk areas (33.02%), with a northeast–southwest distribution and a probable westward expansion trend. This study provides a practical framework for grassland degradation risk assessments and early warning, offering valuable guidance for ecosystem management and sustainable land use.

1. Introduction

Grasslands serve a critical ecological and economic role worldwide, particularly in arid and semi-arid regions [1]. They provide essential ecosystem services, including wind erosion control, soil and water conservation, and biodiversity maintenance, while also supporting economic activities such as livestock farming [2,3]. However, grassland degradation is a pervasive issue, particularly in the arid regions of northern China [4,5], where declining biomass and worsening soil erosion are becoming increasingly severe. The main drivers include overgrazing, climate change, and unsustainable land use practices [6,7,8,9,10], which collectively contribute to the gradual decline in ecosystem productivity and the loss of vital ecological services.

Remote sensing has emerged as a powerful tool for grassland degradation monitoring due to its broad spatial coverage, timeliness, and cost-effectiveness [11]. Indicators such as the normalized difference vegetation index (NDVI) and the net primary productivity (NPP) have been widely applied to assess vegetation dynamics and ecological productivity [12,13,14,15,16,17,18,19,20,21]. Nevertheless, remote sensing data alone may be affected by environmental variability and sensor limitations [22,23,24], leading to uncertainties in representing ground-level ecological conditions. Integrating field data for validation and constructing a multidimensional indicator system are therefore essential for improving the accuracy of grassland degradation risk assessments.

Grasslands differ substantially in their species composition, ecological function, and resilience to disturbances [25,26,27,28], leading to heterogeneous responses to identical degradation pressures [29,30]. Applying a single degradation threshold across diverse grassland types risks oversimplifying these dynamics and may result in misclassification or a biased risk estimation [31,32]. For instance, Zhao et al. (2014) demonstrated significant differences in NPP-based biomass patterns between moist meadow steppes and arid desert steppes in Xilingol [33]. Despite such evidence, many regional assessments still rely on uniform degradation criteria or fixed NDVI/NPP thresholds [34,35,36], neglecting type-specific ecological baselines and explicit tipping points. As a result, few studies have systematically determined the thresholds for different grassland categories. To address this gap, the present study employed K-means clustering on grassland degradation index (GDI) curves, combined with a curvature analysis, to extract statistically supported high-risk thresholds for the NDVI, NPP, and grass yield, thereby enabling type-specific early warning.

However, an early warning of grassland degradation requires more than threshold detection. It also demands a comprehensive understanding of degradation trajectories and spatial evolution patterns [37,38]. Advanced spatial analytical methods—such as a curvature analysis, a standard deviation ellipse (SDE), or a gravity center shift (GCS)—can jointly characterize these processes [39]. A curvature analysis reveals ecosystem instability and acceleration towards critical points [40]; an SDE quantifies the spatial clustering and orientation of degraded areas [41,42]; and a GCS tracks spatiotemporal shifts in degradation hotspots [43,44,45]. Together, these techniques elucidate the spatial–temporal dynamics of degradation, offering a scientific basis for early warning and targeted intervention.

Xilingol, situated at the China–Mongolia border, is a representative grassland region featuring diverse grassland types and severe degradation challenges [46,47]. Although studies have examined the grassland degradation in this region [47,48,49], few have accounted for type-specific differences in the degradation thresholds or spatial dynamics. This study addressed these limitations by integrating the GDI, nonlinear curve fitting, and clustering methods to refine the degradation risk assessments across grassland types. Specifically, the NDVI, NPP, and grass yield were employed as key ecological factors (KEFs) to evaluate degradation from the perspectives of vegetation cover, ecological productivity, and actual yield. Through a statistical analysis and spatial trajectory modeling, this study aimed to (1) reveal degradation patterns across different grassland types by applying a curvature analysis; (2) extract statistically robust high-risk degradation thresholds based on a limited set of key ecological factors (KEFs); and (3) evaluate the grassland degradation risk and provide a comprehensive early-warning framework in Xilingol.

2. Materials and Methods

2.1. Study Area

Xilingol, located in central Inner Mongolia, China, is situated between 42°32′–46°41′N and 111°59′–120°00′E. It spans an area of 202,600 km², with 182,036 km² of grassland, accounting for 89.85% of the total area. This region has a typical temperate semi-arid monsoon climate, with annual temperatures ranging from 0 to 4 °C and precipitation between 200 and 400 mm, mainly concentrated from June to August. The precipitation is uneven, with higher amounts in the east and lower in the west. The terrain consists mainly of flat plateaus and hills, with elevations ranging from 1000 to 1500 m, providing a favorable environment for grassland growth [50].

The grasslands of Xilingol are primarily composed of temperate typical steppes (TTSs), temperate meadow steppes (TMSs), and temperate desert steppes (TDSs) [33], which, together, account for more than 80% of the total grassland area (Figure 1). These grasslands become progressively drier from northeast to southwest, leading to a gradual decline in vegetation cover and ecological functions. As a critical ecological barrier in northern China, the region is not only vital for maintaining ecological services, but also a focal point for grassland degradation [35,51]. Its diverse grassland types, pronounced degradation gradients, and ecological vulnerability make it an ideal site for studying degradation mechanisms and ecological restoration.

2.2. Data

This study primarily utilized two types of data: remote sensing data and field data, supplemented by spatial distribution data of the grassland types.

2.2.1. Remote Sensing Data

The remote sensing data used in this study included Landsat OLI images and MODIS NPP products. The Landsat OLI images with a 30 m resolution, sourced from Google Earth Engine, covered the period from June to August, 2022, with a median composition to align with the timing of field data, and were primarily used for calculating the vegetation indices. The MODIS NPP data, obtained from the Google Earth Engine MOD17A3HGF V6.1 product, with a 500 m resolution resampling to 30 m, were selected to match the timing of the field data and were primarily used for extracting the GDI and conducting a comprehensive assessment of the regional grassland degradation risks.

2.2.2. Field Data and Grassland Type Data

The field data used in this study were obtained from a 2022 grassland survey (Figure 2), which collected 743 sample points during the growing season (June to August). The sampling information included the central coordinates (latitude and longitude), vegetation cover type (herbaceous or shrub–herb), grassland origin (natural or artificial), grassland type, average grass height, grass yield per unit area (dry weight), and dominant grass species. For research purposes, data from natural grasslands with herbaceous vegetation cover and grassland types such as TTSs, TMSs, and TDSs were selected. After data cleaning, 574 valid data points were retained (Figure 1), including 100 from the TMSs (17.42%), 300 from the TTSs (52.26%), and 174 from the TDSs (30.31%). The proportion of sampling points across the steppe types was broadly consistent with the area proportions of these three major steppe types in Xilingol. Spatially, the TMS samples were mainly distributed in the eastern part of the study area, the TTS samples were concentrated in the central and southern parts, and the TDS samples were primarily located in the western part. The grassland type data were sourced from the Resource and Environmental Science Data Platform (https://www.resdc.cn/ (accessed on 17 March 2025)), with a lower spatial resolution suitable for large-area degradation risk assessments.

2.3. Methods

The workflow of the study is shown in Figure 3. First, data such as the SI, albedo, LSM, and NPP were collected or calculated, and a principal component analysis (PCA) was used to extract the weights. The GDI was then calculated to quantify the degree of degradation. After obtaining the GDI values at the sampling points, the data were grouped by grassland type for nonlinear fitting and K-means clustering to classify the degradation risk levels. A curvature calculation and the extraction of KEF ranges were used to identify the dynamic features of grassland degradation, and a statistical analysis was conducted to determine the degradation patterns and high-risk thresholds. Further, a regional degradation analysis, a GCS analysis, and an SDE analysis were used to assess the degradation trends, ultimately leading to the development of a grassland degradation risk early-warning system.

2.3.1. Grassland Degradation Index (GDI)

The GDI is a commonly used index for assessing the degree of grassland degradation on a large scale [52,53,54,55]. The calculation method varies by region, with different areas emphasizing different factors. Salinization is one of the significant factors limiting grassland productivity in the study area [56]. Based on previous studies and the local conditions, the GDI was constructed by integrating the NPP, salinity–alkalinity index (SI), albedo, and land surface moisture (LSM). The calculation formulas are shown in Equations (1)–(4):

$$GDI={\omega }_1\times \left(1-NPP\right)+{\omega }_2\times SI+{\omega }_3\times albedo+{\omega }_4\times (1-LSM),$$

(1)

$$SI=\sqrt{{\rho }_{Red}+{\rho }_{Green}},$$

(2)

$$albedo=0.356{\rho }_{Blue}+0.130{\rho }_{Red}+0.373{\rho }_{NIR}+0.085{\rho }_{SWIR1}+0.072{\rho }_{SWIR2}-0.01,$$

(3)

$$LSM=0.0315{\rho }_{Blue}+0.2021{\rho }_{Green}+0.3102{\rho }_{Red}+0.1594{\rho }_{NIR}-0.6806{\rho }_{SWIR1}-0.6109{\rho }_{SWIR2}$$

(4)

In the equations, ${\omega }_1,{\omega }_2,{\omega }_3,{\omega }_4$ represent the weights of each factor, while ${\rho }_{Red},{\rho }_{Green},{\rho }_{Blue},{\rho }_{NIR},{\rho }_{SWIR1},{\rho }_{SWIR2}$ refer to the red, green, blue, near-infrared, and two shortwave infrared bands of the Landsat OLI image, respectively. After calculating the NPP, SI, albedo, and LSM, data normalization was performed using the range method. The factor weights were then determined using the common factor variance, and the GDI was computed through a comprehensive calculation. The results of the weight calculation are shown in

Table 1. Factor weights for GDI calculation.

Weight	Value
${\omega }_1$	0.249
${\omega }_2$	0.274
${\omega }_3$	0.246
${\omega }_4$	0.231

.

2.3.2. Random Forest

Random forest is an ensemble learning method based on decision trees [57]. It constructs multiple decision trees and combines their results to improve the prediction accuracy and reduce overfitting. The random forest regression formula is expressed in Equation (5):

$$\widehat{y}={\displaystyle \frac{1}{B}}{\sum }_{b=1}^B{T}_b\left(x\right),$$

(5)

where $T_b\left(x\right)$ is the prediction of the b-th decision tree and B is the total number of trees.

Random forest has significant advantages in grass yield inversion, as its nonlinear modeling ability can effectively capture the complex relationship between remote sensing features and the grass yield. In this study, the Classifier.smileRandomForest function provided by Google Earth Engine was used for implementation. This study selected 12 vegetation indices, including the NDVI, EVI, DVI, etc., with the calculation formulas for these indices shown in

Table A2. Vegetation indices.

Vegetation Index	Formula
Difference Vegetation Index (DVI)	${\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{red}}$
Enhanced Vegetation Index (EVI)	${\displaystyle \frac{2.5\times \left({\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{red}}\right)}{1+{\rho }_{\mathrm{nir}}+6\times {\rho }_{\mathrm{red}}-7.5\times {\rho }_{\mathrm{blue}}}}$
Infrared Index (II)	$\left({\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{swri}1}\right)/\left({\rho }_{\mathrm{nir}}+{\rho }_{\mathrm{swri}1}\right)$
Modified Soil-Adjusted Vegetation Index (MSAVI)	${\displaystyle \frac{2{\rho }_{\mathrm{nir}}+1-\sqrt{{\left(2{\rho }_{\mathrm{nir}}+1\right)}^2-8\left({\rho }_{\mathrm{nir}}-{\rho }_r\right)}}{2}}$
Mid-Infrared Index (MidIR)	${\rho }_{\mathrm{swri}1}/{\rho }_{\mathrm{swri}2}$
Moisture Stress Index (MSI)	${\rho }_{\mathrm{swri}1}/{\rho }_{\mathrm{nir}}$
Normalized Difference Vegetation Index (NDVI)	$\left({\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{red}}\right)/\left({\rho }_{\mathrm{nir}}+{\rho }_{\mathrm{red}}\right)$
Ratio Vegetation Index (RVI)	${{\rho }_{\mathrm{nir}}/\rho }_{\mathrm{red}}$
Simple Ratio (SR)	${\rho }_{\mathrm{red}}/{\rho }_{\mathrm{nir}}$
Soil-Adjusted Vegetation Index (SAVI)	${\displaystyle \frac{1.5\times \left({\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{red}}\right)}{{\rho }_{\mathrm{nir}}+{\rho }_{\mathrm{red}}+0.5}}$
Transformed Normalized Difference Vegetation Index (TNDVI)	$\mathrm{T}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=\sqrt{NDVI+0.5}$
Visible Atmospherically Resistant Index (VARI)	${\displaystyle \frac{{\rho }_{\mathrm{green}}-{\rho }_{\mathrm{red}}}{{\rho }_{\mathrm{green}}+{\rho }_{\mathrm{red}}-{\rho }_{\mathrm{blue}}}}$

. Random forest regression was used to extract the grass yield, with the inversion process illustrated in Figure 4. After performing a correlation analysis, 10 vegetation indices (with a correlation value r greater than 0.75) were selected as features to input into the model (as shown in

Table 2. Results of correlation analysis.

Index	II	MSI	DVI	EVI	MSAVI	VARI	SAVI	MidIR	NDVI	RVI	SR	TNDVI
r	0.674	0.674	0.757	0.789	0.789	0.792	0.796	0.809	0.812	0.812	0.812	0.812

).

The field data were divided into three subsets: a training set, a testing set, and a validation set, with a ratio of 6:2:2. Using the RMSE as the evaluation metric, the optimal number of trees in the random forest model was determined based on the training and testing sets (Figure 5), indicating that, when the number of trees equals 115, the model has the lowest RMSE (24.63 g·m⁻²). Subsequently, the random forest model was constructed using both the training and validation sets under the identified optimal conditions. Finally, the model’s accuracy was assessed using the test set, with R² and the RMSE employed as performance indicators (Figure 6). The results showed an R² of 0.859 and an RMSE of 24.64 g·m⁻², which met the requirements of the study. The grass yield distribution map is presented in Figure 7c.

2.3.3. K-Means Clustering

K-means is a classic unsupervised learning algorithm widely used for data clustering, pattern recognition, and dimensionality reduction [58,59]. It divides the data into k mutually exclusive clusters such that the similarity within the same cluster is maximized while the similarity among different clusters is minimized. Its objective function is as follows:

$$J={\sum }_{i=1}^k{\sum }_{x\in C_i}{‖x_j-{\mu }_i‖}^2,$$

(6)

where

k: number of clusters;

$C_i$: the i-th cluster;

${\mu }_i$: the centroid of cluster $C_i$;

$x_j$: the j-th data point in cluster $C_i$.

2.3.4. Average Curvature

Curvature is an important geometric property that describes the degree of bending of a curve. It is commonly used in spatial analyses and ecological monitoring to analyze the characteristics or changing trends of curves [60]. In this study, a curvature analysis was used to explore the differences in degradation patterns among various grassland types. In grassland degradation, a higher curvature of the degradation curve indicates that the degradation process is accelerating and the system is approaching an unstable state, becoming more sensitive to external disturbances (such as climate change, overgrazing, soil degradation, etc.). During this phase, even small external pressures may trigger significant ecological changes. Therefore, paying particular attention to the interval with the maximum curvature is crucial for preventing the further degradation of the ecosystem and developing targeted intervention measures. The formula for calculating curvature is as follows:

$$k(x)={\displaystyle \frac{\left|y^{″}\right|}{{\left(1+{\left(y^{\prime }\right)}^2\right)}^{3/2}}},$$

(7)

where

k(x) represents the curvature at a point;

y′ is the first derivative of the curve;

y″ is the second derivative of the curve.

2.3.5. Separation Degree Index (SDI)

The SDI was constructed in this study to explore which key ecological factor has the strongest indication for grassland degradation risk. The basic idea is as follows: first, determine the total range of all intervals by calculating the difference between the maximum and minimum values of all intervals; next, calculate the proportion of each interval group within the total range; then, accumulate the proportions of each group to obtain the total overlap degree; and finally, take the reciprocal of the total overlap degree to obtain the separation degree (D). The calculation formula is shown below and the results are shown in

Table 3. Results of SDI calculation.

KEFs	Grassland Types	Average	Overall
Grass Yield	TMS	3.84	3.00
	TTS	2.83
	TDS	2.32
NDVI	TMS	2.91	2.76
	TTS	2.89
	TDS	2.47
NPP	TMS	2.72	2.78
	TTS	3.09
	TDS	2.52

.

Assume that there are n intervals [a_i, b_i], where i = 1, 2, 3, …, n.

$$D={\displaystyle \frac{1}{{\sum }_{i=1}^n\left(b_i-a_i\right)/(\mathit{max}\left(b_i\right)-min(a_i\left)\right)}}$$

(8)

3. Results

3.1. Curvature Analysis of Degradation Patterns for Different Grassland Types

To classify the degradation risk, the GDI values were first calculated, extracted at sampling points, and categorized by the grassland type. To reduce data dispersion and noise, nonlinear fitting was applied for smoothing and denoising. The fitted curves were then analyzed through clustering, enhancing the robustness and representativeness of the results while better capturing overall data trends, thereby aiding in the exploration of the degradation patterns. The nonlinear fitting results of the GDI are shown in Figure 8. The Y-axis represents the GDI values, while the X-axis corresponds to the sampling points, which are sequentially ordered from the lowest to the highest GDI value.

According to the Chinese national standard “Parameters for Degradation, Sandification and Salification of Rangelands” (GB 19377-2003) and related literature [5,61,62], the GDI values were classified into four clusters, representing four degradation stages, and the critical values were extracted (

Table A1. GDI values of the boundary points.

Grassland Type	Stage 1–Stage 2	Stage 2–Stage 3	Stage 3–Stage 4
TMS	0.093	0.139	0.305
TTS	0.295	0.426	0.581
TDS	0.771	0.874	0.930

). The curves were divided according to the critical values, and the curvature was calculated for each segment, yielding the results shown in

Table 4. Average curvature of the clustering interval.

Grassland Type	Stage 1	Stage 2	Stage 3	Stage 4
TMS	2.30 × 10−4	8.07 × 10−5	3.38 × 10−4	6.23 × 10−4
TTS	2.39 × 10−4	6.08 × 10−5	1.37 × 10−4	3.24 × 10−4
TDS	2.81 × 10−4	1.42 × 10−4	3.48 × 10−5	1.36 × 10−4

. For the TMSs, the average curvature of the clusters was as follows: stage 4 > stage 3 > stage 1 > stage 2. For the TTSs, the sequence was stage 4 > stage 1 > stage 3 > stage 2. For the TDSs, the sequence was stage 1 > stage 2 > stage 4 > stage 3. These results indicate significant differences in the degradation patterns among different grassland types. In the TMSs and TTSs, stage 4 had the highest curvature, whereas in the TMSs, stage 1 exhibited the highest curvature.

3.2. Extraction of High-Degradation-Risk Thresholds for KEFs

Based on the GDI values of the threshold points derived from Table A1, the grass yield, NDVI, and NPP of the sampling points were grouped. After grouping, boxplots were used to determine the distribution ranges of the KEFs under different degradation risk levels (Figure 9, Figure 10 and Figure 11). To reduce noise influence, the 95th and 5th percentiles were chosen as the maximum and minimum values for the interval range, respectively. Finally, the thresholds for each key ecological factor under different degradation risks were obtained (

Table 5. Thresholds of KEFs for early warning of grassland degradation risk.

Grassland Types	Factors	Thresholds
		No to Low	Low to Medium	Medium to High
TMSs	Grass Yield (g·m−2)	241.05	193.64	115.67
	NDVI	0.90	0.74	0.52
	NPP (kg·C/m2)	0.44	0.40	0.31
TTSs	Grass Yield (g·m−2)	143.00	99.83	73.27
	NDVI	0.59	0.51	0.37
	NPP (kg·C/m2)	0.34	0.27	0.21
TDSs	Grass Yield (g·m−2)	46.40	32.73	32.30
	NDVI	0.28	0.25	0.22
	NPP (kg·C/m2)	0.14	0.13	0.11

). The results show that, for the TMSs, the high-degradation-risk thresholds for KEFs were as follows: grass yield, 115.67 g·m⁻²; NDVI, 0.52; and NPP, 0.31 kg·C/m². For the TTSs, the thresholds were as follows: grass yield, 73.27 g·m⁻²; NDVI, 0.37; and NPP, 0.21 kg·C/m². For the TDSs, the thresholds were as follows: grass yield, 32.30 g·m⁻²; NDVI, 0.22; and NPP, 0.11 kg·C/m².

3.3. Assessment of Grassland Degradation Risk and Spatial Characteristics

After obtaining the thresholds of KEFs under different degradation risk levels, the grassland degradation risk in the Xilingol region was assessed based on the “law of the minimum”. The assessment results are shown in Figure 12. After evaluation, the area proportions under different degradation risks were statistically analyzed from both the overall and grassland type perspectives, as presented in

Table A3. The area proportion under different degradation risk levels.

	I	II	III	IV
TMS	33.16%	16.37%	40.14%	10.32%
TTS	30.63%	11.05%	25.30%	33.02%
TDS	48.21%	22.50%	15.56%	13.73%
SUM	36.22%	15.23%	24.60%	23.94%

. Overall, the area with no degradation risk accounted for the largest proportion, at 36.22%, followed by the high- and medium-degradation-risk areas, which had almost equal proportions of 23.94% and 24.60%, respectively. The area with a low degradation risk had the smallest proportion, at 15.23%. From the perspective of different grassland types, the high-degradation-risk area in the TTSs occupied the largest proportion, reaching 33.02%, followed by the TDSs, at 13.73%, while the high-degradation-risk area in the TMSs had the smallest proportion, at only 10.32%.

A GCS analysis was conducted to reveal the spatial dynamic changes of the same grassland type at different degradation stages. First, the centroids of the grassland distribution range under different risk levels were extracted. Then, arrows were drawn to connect the centroids sequentially, indicating the direction of risk shift. The results are shown in Figure 13. For the TMSs, the degradation risk shifted from low to high in the following direction: northeast–southwest–west (Figure 13d); for the TTSs, the shift was north–southwest–west (Figure 13c); and for the TDSs, the shift direction was northwest–southwest–northeast (Figure 13d). It can be seen that the shift directions for the TMSs and TTSs were similar, with high degradation risks generally expanding westward. In contrast, the shift direction for the TDSs showed a significant difference, with high degradation risks expanding northward. These patterns suggest that, for the medium- to high-degradation-risk zones of the TMSs and TTSs, grassland degradation should be prevented from spreading westward, while for the TDSs, the spread of degradation should be prevented from moving northward.

In addressing high-degradation-risk areas, it is essential not only to understand the direction of degradation shift, but also to clearly identify the distribution characteristics of the existing high-risk zones. Therefore, we used an SDE to visually represent the spatial distribution patterns and spatial clustering characteristics of the high-risk areas (Figure 14). It was observed that the high-degradation-risk areas of the TMSs and TTSs exhibited a northeast–southwest distribution, with the TMSs having a more scattered distribution and weaker clustering compared to the TTSs. In contrast, the high-degradation-risk areas of the TDSs showed a completely different feature, with a northwest–southeast orientation and the most pronounced clustering.

4. Discussion

4.1. Variations in Degradation Patterns Among Grassland Types

We hypothesized that different grassland types would exhibit distinct degradation trajectories, which has important implications for early warning and management. To test this, we constructed degradation curves by ordering the GDI values from low to high and quantified stage-specific change by computing the average curvature within defined GDI intervals. The curvature metric therefore serves as an indicator of the rate and acceleration of degradation across stages (Table 4).

The analysis revealed markedly different trajectories among the three grassland types. The TMSs displayed a relatively low curvature in the early stages, with a pronounced increase in the final stage, indicating initial resistance followed by late-stage acceleration. The TDSs showed the opposite pattern: a high curvature at early stages that attenuated later, suggesting rapid early decline and subsequent stabilization. The TTSs exhibited a moderate, more uniform curvature overall, with its peak degradation rate occurring in the later stage, but without the abrupt late surge observed in the TMSs.

Ecologically, these patterns reflect contrasts in stability and resistance. The late-stage acceleration in the TMSs implies greater short-term resistance to disturbances, but a vulnerability to crossing a tipping point once compensatory mechanisms are exhausted. In contrast, the early rapid decline in the TDSs indicates a lower resistance and a propensity for the quick loss of function, even under modest stress. The TTSs occupy an intermediate position, showing relative temporal stability, but still being susceptible to progressive deterioration.

These distinctions have direct implications for early-warning systems and adaptive management. For TDSs, monitoring and intervention should focus on the earliest detectable signals, since the system can deteriorate quickly. For TMSs, periodic monitoring that can detect gradual trends and indicators signaling an approaching inflection point is critical to prevent sudden collapse. For TTSs, a balanced approach combining routine surveillance and contingency measures may be appropriate. In practice, this implies tailoring the monitoring frequency, indicator thresholds, and management responses to the grassland type rather than applying uniform criteria.

4.2. Differences in KEF Thresholds and Indicator Strength

KEFs are decisive for assessing the grassland degradation risk. In this study, the NDVI, NPP and grass yield were selected as KEFs based on their relevance to the vegetation condition and on considerations of multidimensionality, complementarity, and data accessibility. High-risk thresholds were statistically derived (Table 5), and because areas classified as high-risk typically represent the ecosystem’s critical zones—where restoration demands substantially more time and resources—our discussion focuses on these thresholds. The results showed a clear decline in the high-risk thresholds from the TMSs to the TTSs to the TDSs. For example, the high-risk NDVI thresholds were 0.52 (TMSs), 0.37 (TTSs), and 0.22 (TDSs). Such differences indicate substantial variation in the ecological vulnerability and adaptive capacity among grassland types and imply that management and conservation strategies should be tailored to type-specific risk profiles.

To evaluate which KEF best indicates the degradation risk, we developed a separation degree index (SDI) to quantify the indicator strength (Table 3). Overall, the grass yield exhibited the highest separation degree (3.00), while the NDVI and NPP showed similar values (2.76 and 2.78, respectively). When disaggregated by grassland type, the grass yield had the highest SDI in the TMSs (3.84), whereas the NPP performed best in the TTSs and TDSs (3.09 and 2.52, respectively). Hence, if only a single indicator can be monitored, the grass yield is the preferred overall metric; however, when grassland type is considered, the grass yield is most informative for the TMSs and the NPP is preferable for the TTSs and TDSs. To improve the warning reliability in practice, we recommend adopting multi-factor or combined-criteria approaches rather than relying solely on a single threshold.

4.3. Comprehensive Grassland Degradation Risk Assessment and Early Warning

We combined sample-based KEF intervals, remotely retrieved the regional grass yield, and statistically derived KEF thresholds to produce a composite degradation-risk map for Xilingol. The results showed that 36.22% of the area was classified as having no risk, with the moderate- and high-risk areas accounting for 24.60% and 23.94%, respectively. Although the largest fraction remained in a comparatively good state, the combined area of moderate and high risk indicates substantial pressure on the regional grassland system and is consistent with previous assessments that identified Xilingol as a region with pronounced degradation dynamics.

The distribution of risk among the grassland types is non-uniform: the TTSs showed the largest proportion of high-risk area (33.02%), whereas the TMSs exhibited the smallest share. These inter-type differences mirror earlier findings that attributed a greater vulnerability in some typical-steppe areas to long-term anthropogenic pressure and land-use change, which can amplify soil erosion and reduce resilience [63]. Empirical analyses of grazing impacts and regional land-use studies have also reported that grazing pressure and conversion dynamics are primary drivers of the productivity decline in Xilingol and the adjacent northern grassland regions [49,64,65].

Climatic variability and soil attributes further modulate these patterns: consistent with national and regional syntheses, the interaction of precipitation anomalies and human disturbances determines whether a grassland is likely to cross a critical degradation threshold in a given year or decade [44,66,67,68]. This aligns with broader analyses showing that both climate variability and intensified anthropogenic use jointly explain much of China’s grassland degradation trends.

From an early-warning standpoint, our results reinforce two points established in recent literature. First, single-indicator approaches are often insufficient for operational early warnings because they may not capture the productivity, cover, and yield simultaneously [69]; multi-indicator frameworks improve the detection of emergent risks. Second, spatial–temporal metrics provide actionable information beyond threshold maps by indicating the direction and speed of degradation spread. Recent studies have demonstrated the value of combining productivity thresholds with spatial trajectory analyses to anticipate transitions to degraded states.

Applying these principles to Xilingol leads to differentiated early-warning recommendations: for TDSs and TTSs, where degradation is more concentrated and initiated earlier, management should prioritize the early detection and immediate control of grazing intensity and localized land-use pressures. For TMSs, where high-risk patches are more dispersed, but can accelerate toward critical points, we recommend monitoring systems capable of detecting gradual inflection signals (e.g., curvature-based early-warning indicators) and triggering preventive interventions. Across all types, we advocate for multi-factor thresholds, the periodic reassessment of thresholds under changing climatic trends, and targeted field validation to confirm remote sensing-based risk assignments.

4.4. Advantages, Limitations, and Future Directions

This study offers a compact and pragmatic framework by integrating complementary KEFs with a curvature-based trajectory analysis and spatial dynamic metrics. The combined approach enhances the sensitivity to both temporal tipping behavior and the spatial propagation of risk while remaining interpretable and suitable for operational monitoring.

However, this study also has certain limitations. First, the spatial resolution and temporal span of the remote sensing data may affect the precision of the high-risk degradation zone thresholds, especially in regions with different grassland types, where the representativeness of the data may vary. Second, the methods in this study were primarily focused on the Xilingol region and have not been widely validated in other regions or grassland types.

Therefore, future research needs to extend to additional areas to test the general applicability and adaptability of the models. We will focus on the following aspects: (1) implementing simple robustness checks and reporting uncertainty ranges for key thresholds; (2) conducting targeted field validations or using independent regional datasets to corroborate remote sensing assignments; and (3) exploring lightweight hybrid models (e.g., explainable ML combined with rule-based thresholds) to enhance the predictive performance without sacrificing interpretability. These refinements will increase confidence in the method’s applicability across regions and support its practical use in adaptive, type-specific early-warning systems.

5. Conclusions

This study established an integrated framework for assessing and predicting grassland degradation risks in Xilingol by combining key ecological factors (NPP, NDVI, and grass yield) with K-means clustering, a curvature analysis, and spatial dynamic methods. The approach enabled the identification of degradation patterns, the extraction of type-specific thresholds, and the development of a multi-dimensional early-warning system.

The findings revealed that the degradation dynamics differ substantially among grassland types. The TMSs demonstrated a high ecological stability, with the degradation accelerating only under high-risk conditions, whereas the TTSs showed a moderate stability with gradual risk accumulation and the TDSs exhibited a pronounced vulnerability, experiencing rapid degradation even at low risk levels. Although the degradation in Xilingol remains generally controllable, high-risk areas are concentrated in the TTSs and, to a lesser extent, in the TDSs, while the TMSs show the lowest risk proportion. A spatial analysis further highlights the westward and northward shifts in degradation centers, underscoring the need for tailored management strategies.

For early warnings, the TMSs require attention to dispersed degradation hotspots, the TTSs demand interventions to prevent westward expansion and transition toward TDSs, and the TDSs call for proactive measures to mitigate rapid-onset degradation. By integrating ecological thresholds with spatial dynamics, the proposed framework provides a data-efficient and transferable tool for supporting early-warning systems and guiding adaptive grassland management.