Nonlinear Water–Heat Thresholds, Human Amplification, and Adaptive Governance of Grassland Degradation Under Climate Change

Highlights

What are the main findings?

• 30 m structure–function assessment links LUCC transitions with NPP, NEP, soil conservation (SC), and grassland supply (GS), revealing a polarized 2003–2023 trajectory in the ELB.
• Explainable XGBoost–SHAP identifies climate dominance with accessibility amplification and quantifies operational thresholds (e.g., ~200 mm rainfall gate; road density ~0.06 km km⁻²; grazing windows 2.2–4.2 and 4.65–5.61 SU km⁻²).

What are the implications of the main findings?

• Thresholds translate directly into adaptive zoning (“two belts–four zones–one axis”), enabling season-/year-specific controls on grazing, roads, and water allocation.
• The workflow is transferable to mountain–oasis–desert basins, supporting climate-risk reduction and stabilization of carbon/soil functions in drylands.

Abstract

Dryland grasslands face elevated risks of rapid threshold crossing under a regime of warming, precipitation redistribution, and intensified interannual hydrothermal variability. Using the Ebinur Lake Basin (ELB) as a case, we developed an integrated structure × function assessment—linking land-use/cover change (LUCC) transitions with functional indicators of net primary productivity (NPP), net ecosystem production (NEP), soil conservation (SC), and grass supply (GS)—and coupled it with Bayesian-optimized XGBoost, SHAP, and partial dependence plots (PDPs) at a 30 m pixel scale to identify dominant drivers and ecological thresholds, subsequently translating them into governance zones. From 2003 to 2023, overall grassland status was dominated by degradation (20,160.62 km²; 69.42%), with restoration at 8878.85 km² (30.57%) and stability at 2.79 km² (0.01%). NPP/NEP followed a rise–decline–recovery trajectory, while SC exhibited marked bipolarity. Precipitation and temperature emerged as primary drivers (interaction X3 × X4 = 0.0621), whose effects, together with topography and accessibility, shaped a spatial paradigm of piedmont sensitive–oasis sluggish–lakeshore vulnerable. Key thresholds included an annual precipitation recovery threshold of ~200 mm and an optimal window of 272–429 mm; a road-density divide near ~0.06 km km⁻²; and sustainable grazing windows of ~2.2–4.2 and ~4.65–5.61 livestock units (LU) km⁻². These thresholds underpinned four management units—Priority Control (52.53%), Monitoring and Alert (21.53%), Natural Recovery (20.40%), and Optimized Maintenance (5.55%)—organized within a “two belts–four zones–one axis” spatial framework, closing the loop from threshold detection to adaptive governance. The approach provides a replicable paradigm for climate-adaptive management and ecological risk mitigation of dryland grasslands under warming.

1. Introduction

Grasslands are a major component of the terrestrial biosphere, underpinning biodiversity, carbon sequestration, hydrological regulation, and pastoral livelihoods. In China, they cover nearly 40% of the land surface, making their trajectories pivotal to socio-ecological sustainability in arid and semi-arid regions [1,2]. Over the past two decades, the climate signal of “warming + precipitation redistribution → intensified interannual hydrothermal variability”, together with cropland expansion, road densification, and intensified grazing, has driven more frequent crossings of key hydrothermal thresholds in drylands. These crossings have produced concurrent structural fragmentation, functional volatility, and trade-offs among ecosystem services [3,4,5]. Compared with humid regions, dryland grasslands are more sensitive to precipitation pulses, soil moisture, and atmospheric vapor pressure deficits (VPDs), and their responses are often strongly nonlinear with platform-tipping behavior. The relative dominance of soil moisture versus VPD varies regionally, and their synergy or antagonism jointly governs both the magnitude and the direction of interannual variability [6,7,8,9]. In parallel, accessibility gradients (e.g., road density, distance to settlements) and human activity intensity markedly amplify degradation risk near thresholds, underscoring the need for governance strategies centered on disturbance control, pressure limits, and optimized allocation [10,11,12,13].

At the regional climate scale, the Tianshan–Central Asia “water tower” has exhibited a prominent pattern of “warming + precipitation redistribution → intensified interannual hydrothermal variability” in the 21st century: large interannual swings and alternating wet–dry years are reshaping hydrothermal gradients and seasonal rhythms across the mountain–fan–oasis–playa sequence [14,15]. Shifts in rain–snow phase and melt timing further modify runoff generation and surface water–salt regimes, redefining the “sensitive windows” for soil retention and carbon sequestration [16,17,18]. Globally and in China, semi-arid ecosystems contribute disproportionately to the interannual variability of the terrestrial carbon cycle and are the most climate-sensitive, reinforcing the urgency of building indicator-based, interpretable, and actionable monitor–assess–adapt frameworks for drylands [2,4,19].

Methodologically, despite advances in remote sensing and composite indicators, three bottlenecks persist. First, many studies rely on single cover or process metrics (e.g., NDVI), which cannot simultaneously capture coupled structure × function degradation and service trade-offs spanning net primary productivity (NPP), net ecosystem production (NEP), soil conservation (SC), and grass supply (GS). Second, linear or quasi-linear models miss nonlinearities and interactions among drivers, leading to conservative estimates of risk amplification near thresholds. Third, the translation chain from “mechanistic diagnosis → spatial governance” is incomplete: thresholds and platforms often remain graphical interpretations rather than operational, zoned, and tiered management lists linked to performance [2,3,20]. Recently, interpretable machine learning—e.g., XGBoost coupled with SHapley Additive exPlanations (SHAP) and partial dependence plots (PDPs) or accumulated local effects (ALE)—offers a pathway to quantify dominant factors, reveal pairwise interactions and nonlinear thresholds, and support governance zoning at pixel scale [21,22,23,24,25]. Concurrent multi-scale dryland studies emphasize that in spatial overlaps of “hydrothermal sensitivity × high accessibility”, road networks and grazing pressure greatly amplify degradation risk and thus must be managed in tandem with threshold identification [11,12,13].

The Ebinur Lake Basin (ELB) forms a canonical arid geo-utilization gradient along the north slope of the Tianshan Mountains—from alpine zones through piedmont alluvial fans to an oasis corridor and terminal lake–playa. Snowmelt-regulated runoff shapes fan irrigation and oasis expansion, while strong evaporation and salt–dust recirculation in the terminal basin feed back to surface processes [16]. Recent multi-source remote sensing and process studies report significant vegetation and ecological risks associated with land-use transitions (cropland expansion, road densification), salinization–aridification patterns, and lake-area changes, making ELB an ideal “natural experiment” to identify composite climate-dominant–human-amplified–topography-modulated drivers and to spatially translate thresholds/platforms [4,14,15]. Against this backdrop, building a pathway that couples structure–function assessment with interpretable ML to detect nonlinear thresholds and interactions, and then translates thresholds into zoned, tiered governance rules, is directly relevant to achieving land degradation neutrality (LDN) and climate-adaptive management in drylands [1,2,20,23].

Accordingly, we implement an integrated Assess–Diagnose–Govern pipeline: at a 30 m pixel scale, we jointly classify degradation/restoration using structural land-use/cover change (LUCC) maps and functional indicators (NPP, NEP, SC, GS); we apply Bayesian-optimized XGBoost with SHAP to identify dominant drivers, spatial heterogeneity, and pairwise interactions; and we use PDPs to extract single-factor thresholds and platforms, extrapolating them into zoned, tiered governance units to form a climate-informed spatial strategy. The study addresses three objectives: (i) characterize ELB’s structural and functional trajectories (2003–2023); (ii) reveal nonlinear thresholds and interaction chains among climate, hydrothermal conditions, topography, and human activities; and (iii) translate thresholds into executable management zones and graded intensities, providing a replicable pathway for climate-adaptive grassland management and indicator-based evaluation in arid regions.

2. Materials and Methods

2.1. Study Area

The ELB (43.61–45.65°N, 79.89–85.24°E) lies in northern Bortala Mongol Autonomous Prefecture, Xinjiang, China, and covers approximately 47,704 km² [26]. From south to north, it forms a typical arid inland transect consisting of the Tianshan alpine belt, piedmont alluvial fans, an oasis corridor, and a terminal lake–playa system draining into the endorheic Ebinur Lake (Figure 1). Elevation declines from ~4800 m in the Tianshan to ~130 m around the lake margin [27].

The basin is characterized by a mid-latitude continental arid climate, with a mean annual temperature of ~6.4 °C, annual precipitation of ~90–170 mm, and potential evapotranspiration near 1600 mm. This produces a persistent water deficit and a steep hydrothermal (water–energy) gradient from the mountain recharge areas to the oasis and terminal depression, which strongly constrains grassland productivity, stability, and degradation thresholds [28].

Land use follows this gradient. Alpine and montane grasslands transition downslope into irrigated cropland and mixed grazing systems along the Wenquan–Bole–Jinghe corridor. Over roughly the past two decades, cropland expansion, densification of transportation infrastructure, and intensified water extraction and salinization have increased grassland fragmentation and desertification risk, particularly along fan–oasis ecotones [29]. These coupled hydroclimatic and accessibility gradients make the ELB an effective natural laboratory for assessing grassland degradation processes and management zoning under arid-climate stress.

2.2. Data Sources

We assembled a multi-source database for 2003–2023: (i) 30 m land-cover maps for structural change detection; (ii) NPP, NEP, SC and GS for functional assessment; and (iii) 15 candidate drivers spanning hydro-climate, terrain, soils, vegetation, and human pressure (population density, human-activity intensity—HAI, grazing intensity, roads and distance to settlements). All rasters were harmonized to a 30 m Albers equal-area grid with co-registration and attribute checks to ensure temporal–spatial consistency.

Specifically, structural changes were mapped from the 30 m annual China Land-Cover Dataset (CLCD) for 1985–2023, from which we extracted five benchmark years (2003, 2008, 2013, 2018, 2023) and encoded pixel-wise grassland transitions (Appendix A,

Table A1. Datasets and preprocessing workflow used in the 2003–2023 grassland assessment of the Ebinur Lake Basin (ELB).

Data Layer	Primary Dataset/Citation	Native Spatial & Temporal Coverage	Key Preprocessing & Use in This Study (2003–2023)
Land–use/land–cover (LULC)	30 m annual land cover datasets in China, 1985–2023 [Data set] [51].	30 m; annual 1985–2023	Extract five benchmark years (2003, 2008, 2013, 2018, 2023); reproject to 30 m Albers equal–area; aggregate to six classes; co–register (<0.5 pixel); encode pixel–wise transition codes (A × 10 + B) for grassland transitions.
Net Primary Productivity (NPP)	China regional 250 m annual max–NDVI, 2000–2024 [Data set] + CMFD v2.0 (meteorology) [30,52].	NDVI: 250 m; annual 2000–2024. CMFD: 0.1°; daily 1951–2024	Quality control NDVI and derive annual maximum; resample NDVI and CMFD to 30 m grid; drive a CASA–type light–use–efficiency model (monthly APAR × ε) to estimate annual NPP for 2003–2023; extract benchmark–year values for mapping and use full annual series for trend tests.
Net Ecosystem Productivity (NEP)	Net ecosystem productivity (NEP) of China (2001–2023) [Data set] [31].	~1 km; annual 2001–2023	Extract 2003–2023; bilinear resample to 30 m; use as functional indicator for pixel–wise trend tests and benchmark–year mapping.
Soil conservation (SC)	Chinese soil conservation/water–erosion control dataset (1992–2019) + CHIRPS/CMFD precipitation + HWSD v2.0 soils [53,54].	SC product: ~1 km; annual 1992–2019. Precip.: 0.05°–0.1°; daily 2000–2024. Soils: ~1 km; quasi–static	Use annual SC layers for 2003–2019; extend to 2020–2023 by carrying forward the 2017–2019 mean (conservative extension); resample to 30 m; use as functional indicator.
Grassland supply (GS)	Annual 30 m global grassland extent (2000–2022) [Data set] + LHGI (1980–2022) [Data set] + NDVI–based biomass [36,52,55].	Grassland extent: 30 m; annual 2000–2022. Grazing intensity: 300 m; annual 1980–2022	Construct a forage–supply index (SU km−2 yr−1) by combining NDVI–based biomass, mapped grassland fraction, and standardized livestock load; resample inputs to 30 m; extend 2023 using 2022 grassland extent and 2020–2022 mean grazing intensity; interpret GS mainly in terms of spatial gradients and relative change.
Climate drivers (P, T, VPD, radiation, wind)	China Meteorological Forcing Dataset v2.0 (CMFD v2.0) [Data set] [30].	0.1°; daily 1951–2024	Aggregate daily fields to annual totals/means for 2003–2023; bilinear resample to 30 m; derive driver layers (e.g., annual P and mean T) for XGBoost.
Topography	GEBCO 2024 Grid [Data set] [56].	~15 arc–sec (~500 m); 2024 release	Extract elevation; derive slope; resample to 30 m; treat as quasi–static terrain constraints over 2003–2023.
Soil properties	Harmonized World Soil Database v2.0 (HWSD v2.0) [Data set] [54].	~30 arc–sec (~1 km); quasi–static	Extract SOC and pH (0–30 cm); bilinear resample to 30 m; treat as quasi–static drivers.
Vegetation greenness (NDVI)	China regional 250 m annual max–NDVI (2000–2024) [Data set] [52].	250 m; annual 2000–2024	Quality control and annual compositing; resample to 30 m; use both as a driver (X7) and as input to NPP/GS calculations.
Desertification degree	National desertification raster (10 m), snapshots 2010 & 2020 [Data set] [57].	~10 m; snapshots 2010, 2020	Re–index classes to an ordinal severity score; linearly interpolate 2010→2020 to obtain annual series; extend endpoints to 2003–2009 and 2021–2023 by nearest available snapshot; resample to 30 m.
Soil–erosion severity	Chinese soil conservation dataset preventing soil water erosion (1992–2019) [53].	~1 km; annual 1992–2019	Extract annual erosion/severity layer for 2003–2019; extend to 2020–2023 by carrying forward 2017–2019 mean; resample to 30 m; use as driver X9.
River density	China River Extent Maps (CRED), 2016–2023 [Data set] [58].	~10–30 m; annual 2016–2023	Compute river–length density (km km−2) on 30 m grid; use 2016–2023 multi–year mean as quasi–static drainage constraint for 2003–2023.
Population density	LandScan Silver Edition 2022: Global population distribution [Data set] [59].	~1 km; reference year 2022	Resample to 30 m; treat as quasi–static socio–economic pressure proxy in the driver set.
Human–activity intensity (HAI)	Annual Human Footprint (2000–2018) [Data set] + High–quality Daily Nighttime Light (HDNTL) (2012–2024) [Data set] [60,61].	Human Footprint: 1 km; annual 2000–2018. HDNTL: ~500 m; daily 2012–2024	Normalize both indices and fuse to a 0–10 HAI scale using overlap years for calibration; build annual HAI for 2003–2023 (use Human Footprint where available; use HDNTL–calibrated continuation for 2019–2023); resample to 30 m.
Grazing intensity	Long–term High–resolution Grazing Intensity (LHGI) for China (1980–2022) [Data set] [36].	0.1° (~10 km) for 1980–2000; 0.0025° (~250 m) for 2001–2022; annual	Extract 2003–2022; resample to 30 m; extend to 2023 using 2020–2022 mean; use as driver X13 and for GS construction.
Road density & distance to settlements	Road network + National Geoinformation Public Service Platform (Tianditu), GS(2024)0650 settlements [62,63].	Vector infrastructure layers; ~2024 snapshot	Rasterize roads to 30 m; compute road density (km km−2) within each pixel (or moving window, as specified in Methods); compute Euclidean distance to nearest settlement (km); treat as quasi–static accessibility proxies (X14–X15).

). Grassland functions were characterized by four indicators at 30 m resolution: net primary productivity (NPP), net ecosystem productivity (NEP), soil conservation (SC), and grassland supply (GS). NPP was simulated using a light-use-efficiency (CASA-type) model (Table 1), driven by MODIS MOD13 NDVI (250 m, 16-day composites) and daily meteorological forcing from CMFD v2.0 (0.1°) [30], and was aggregated to annual NPP for 2003–2023 on the 30 m analysis grid. NEP was taken from a national gridded NEP product for China (2001–2023) [31], which combines satellite-based NPP with respiration estimates based on CMFD meteorology and soil organic carbon. SC was obtained from a Chinese soil conservation dataset that applies a RUSLE-type water erosion scheme using CHIRPS precipitation, HWSD v2.0 soils, and a 30 m DEM. GS was represented by a forage-supply index (stock units km⁻² yr⁻¹) combining NDVI-based biomass, mapped grassland extent, and standardized livestock units at the county level. The software versions we use are ArcGIS 10.8 and R 4.4.1.

Table 1. Indicators used to characterize grassland ecosystem functions and their computation.

Indicator	Formula	Notes
NPP	$${\mathit{NPP}}_{x,t}={\mathit{APAR}}_{x,t}\times {\epsilon }_{x,t}$$ (1)	Pixel-scale monthly NPP is computed using a light-use efficiency framework. APARx,t is absorbed photosynthetically active radiation, estimated from total shortwave radiation (SOL) and fraction of absorbed PAR (FPAR, derived from NDVI). εx,t is realized light-use efficiency, constrained by εmax (maximum biome-specific efficiency), temperature scalars Tε1, Tε2, and a moisture scalar Wε. Monthly NPP (2003–2023) is aggregated to annual values and mosaicked to a 30 m grid.
NEP	$${\mathit{NEP}}_{x,t}={\mathit{NPP}}_{x,t}-R_{\mathrm{h}}\left(x,t\right)$$ (2)	NEP is represented as the balance between production and heterotrophic respiration. Rh(x,t) is soil heterotrophic respiration, parameterized as a function of mean annual temperature T (°C) and annual precipitation P (mm), with empirical coefficients α,β,γ. NEP>0 indicates a net carbon sink; NEP<0 indicates a net carbon source.
SC	$${\mathit{SC}}_x=A_R\left(x\right)-A_C\left(x\right)$$ (3)	Soil conservation is quantified as the avoided erosion, i.e., the difference between potential erosion AR(x) and actual erosion AC(x). The Revised Universal Soil Loss Equation (RUSLE)-type factors include rainfall erosivity R, soil erodibility K, slope length–steepness factor LS (derived from the DEM), cover factor C (derived from NDVI), and support practice factor P (set to 1 for natural grassland or assigned according to conservation measures).
GS	$${\mathit{GI}}_{x,t}=\kappa \times {\mathit{NDVI}}_{x,t}\times \left({\displaystyle \frac{{\mathit{GA}}_z}{A_z}}\right)\times {\mathit{SU}}_{u,t}$$ (4)	GIx,tis the annual forage supply index for pixel x (in standardized sheep units, SU·km−2·yr−1). κ is a biomass conversion coefficient; GAx/Ax is the grassland area fraction within the pixel; SUu,t is the standardized livestock load (sheep units, SU) for the corresponding administrative unit u at time t.

Table 1. Indicators used to characterize grassland ecosystem functions and their computation.

Indicator	Formula	Notes
NPP	$${\mathit{NPP}}_{x,t}={\mathit{APAR}}_{x,t}\times {\epsilon }_{x,t}$$ (1)	Pixel-scale monthly NPP is computed using a light-use efficiency framework. APARx,t is absorbed photosynthetically active radiation, estimated from total shortwave radiation (SOL) and fraction of absorbed PAR (FPAR, derived from NDVI). εx,t is realized light-use efficiency, constrained by εmax (maximum biome-specific efficiency), temperature scalars Tε1, Tε2, and a moisture scalar Wε. Monthly NPP (2003–2023) is aggregated to annual values and mosaicked to a 30 m grid.
NEP	$${\mathit{NEP}}_{x,t}={\mathit{NPP}}_{x,t}-R_{\mathrm{h}}\left(x,t\right)$$ (2)	NEP is represented as the balance between production and heterotrophic respiration. Rh(x,t) is soil heterotrophic respiration, parameterized as a function of mean annual temperature T (°C) and annual precipitation P (mm), with empirical coefficients α,β,γ. NEP>0 indicates a net carbon sink; NEP<0 indicates a net carbon source.
SC	$${\mathit{SC}}_x=A_R\left(x\right)-A_C\left(x\right)$$ (3)	Soil conservation is quantified as the avoided erosion, i.e., the difference between potential erosion AR(x) and actual erosion AC(x). The Revised Universal Soil Loss Equation (RUSLE)-type factors include rainfall erosivity R, soil erodibility K, slope length–steepness factor LS (derived from the DEM), cover factor C (derived from NDVI), and support practice factor P (set to 1 for natural grassland or assigned according to conservation measures).
GS	$${\mathit{GI}}_{x,t}=\kappa \times {\mathit{NDVI}}_{x,t}\times \left({\displaystyle \frac{{\mathit{GA}}_z}{A_z}}\right)\times {\mathit{SU}}_{u,t}$$ (4)	GIx,tis the annual forage supply index for pixel x (in standardized sheep units, SU·km−2·yr−1). κ is a biomass conversion coefficient; GAx/Ax is the grassland area fraction within the pixel; SUu,t is the standardized livestock load (sheep units, SU) for the corresponding administrative unit u at time t.

The 15 candidate drivers used in the machine learning analysis comprise hydroclimate (annual precipitation and annual mean temperature), topography (elevation and slope), soils (soil organic carbon and pH), ecosystem state indicators (NDVI, desertification degree, water-erosion severity, and river density), and human-pressure indicators (population density, human activity intensity—HAI, grazing intensity, road density, and distance to settlements) (

Table 2. Driver system of grassland degradation.

System	Category	Indicator	Variable ID	Unit
Natural environment system	Topographic conditions	Elevation	X1	m
		Slope	X2	°
	Climatic factors	Precipitation	X3	mm yr−1
		Air temperature	X4	°C
	Soil properties	Soil organic carbon (SOC)	X5	g kg−1
		Soil pH	X6	—
	Ecosystem state	NDVI (Normalized Difference Vegetation Index)	X7	—
		Land desertification level	X8	% (dimensionless index)
		Soil erosion risk/severity	X9	% (dimensionless index)
		River (drainage) density	X10	km km−2
Human activity system	Socioeconomic pressure	Population density	X11	persons km−2
		Human Activity Intensity (HAI)	X12	—
	Direct resource extraction	Grazing intensity	X13	livestock units km−2
	Infrastructure disturbance	Road density	X14	km km−2
		Distance to settlements	X15	km

). Road density and distance to settlements were derived from national-scale infrastructure/settlement vector layers and used as quasi-static accessibility gradients over 2003–2023, whereas time-varying disturbance proxies (e.g., grazing intensity and HAI) follow their native annual sequences as detailed in Appendix A.

2.3. Research Framework

This study implements an integrated three-stage framework—Assessment, Diagnosis, and Governance—to link pixel-scale ecological processes to spatially explicit management in the ELB (Figure 2).

(1)

Assessment. We jointly evaluate structural change and functional change in grasslands at 30 m resolution. Structural change was characterized from the CLCD land-use/land-cover maps using five benchmark years (2003, 2008, 2013, 2018, and 2023), reporting both the net 2003–2023 transition and the stage-wise transitions across the four adjacent intervals. Functional change was derived from annual time series (2003–2023), and the integrated structure–function status was synthesized under an OOAO/LfL decision rule to support subsequent mechanism diagnosis and governance zoning.

(2)

Diagnosis. We then use an interpretable machine-learning approach. A gradient-boosted decision tree model (XGBoost) predicts the pixel-wise ordinal degradation–restoration response (Y) from terrain, hydroclimate, soil, vegetation, and human-use drivers. SHAP and partial dependence analysis are used to (i) quantify the relative importance of drivers, (ii) reveal nonlinear responses and pairwise interactions, and (iii) identify ecological thresholds and sensitive ranges.

(3)

Governance. Finally, we translate the identified thresholds—for example, water availability windows and disturbance limits—into management zoning. Pixels are assigned to differentiated management units (e.g., priority control, monitoring and alert, natural recovery, optimized maintenance), which together form a basin-scale spatial strategy. In this way, ecological thresholds become operational guidance for adaptive grassland management under a drying–warming climate.

This three-stage “assessment–diagnosis–governance” pathway builds on earlier calls to couple structural and functional indicators in dryland degradation assessments and to explicitly link ecological thresholds with management triggers [20,32,33]. Our contribution is to operationalize this idea at a 30 m pixel scale by jointly evaluating LUCC transitions and multiple grassland functions, diagnosing nonlinear driver–response relationships with interpretable machine learning, and then translating the derived thresholds into spatially explicit management units.

2.4. Integrated Assessment of Grassland Degradation

2.4.1. Structural Degradation

To characterize physical (structural) transitions in grasslands, we applied pixel-level land-use/land-cover (LULC) transfer mapping between 2003 and 2023 [34]. Both 30 m LULC rasters were reprojected to the Albers Equal-Area projection and strictly co-registered (co-location error < 0.5 pixel). Classes were encoded as: cropland = 1, forest = 2, grassland = 3, water = 4, built-up = 5, unused land = 6. For each pixel and each comparison interval, a two-date transition code was computed as C = A × 10 + B, where A and B are the land-cover class codes at the start and end dates, respectively, yielding combination values from 11 to 66 and a corresponding transfer matrix.

This coding was used to summarize the net 2003–2023 transition for the integrated assessment, and it was also applied to each adjacent benchmark interval (2003–2008, 2008–2013, 2013–2018, and 2018–2023) to support the stage-wise structural analysis. The use of land-cover transitions to characterize grassland structural degradation and restoration follows earlier work on dryland ecosystem change and rangeland condition assessment [12,35]. Structural status was then classified as follows: (i) degraded—grassland transferred out to another class (codes 31, 32, 34, 35, 36); (ii) restored—other classes transferred into grassland (codes 13, 23, 43, 53, 63); and (iii) stable—grassland remained grassland (code 33). For example, C = 31 denotes grassland (3) in 2003 converted to cropland (1) in 2023 and is therefore classified as structural degradation, whereas C = 13 denotes cropland converted to grassland and is classified as structural restoration.

2.4.2. Functional Degradation

To capture process–function change, we assessed four indicators: NPP, NEP, SC, and GS. The selection reflects core ecosystem service domains: NPP (supporting/biophysical input), NEP (climate regulation), SC (erosion control and ecosystem stability), and GS (forage provision and livelihoods, Table 1).

As summarized in Table 1, pixel-scale monthly NPP is computed with a CASA-type light-use-efficiency framework using absorbed photosynthetically active radiation and meteorological scalars. Monthly values for 2003–2023 are aggregated to annual NPP and mosaicked to the 30 m grid. NEP is obtained from an existing national NEP dataset for China and represents the balance between NPP and heterotrophic respiration. SC follows a Revised Universal Soil Loss Equation (RUSLE) formulation, quantifying avoided water erosion as the difference between potential and actual soil loss. GS is expressed as an annual forage-supply index that combines NDVI-based biomass, grassland extent, and standardized livestock units at the administrative-unit scale. The combination of productivity (NPP, NEP), soil conservation, and forage supply thus captures complementary aspects of grassland ecosystem functioning that are widely used in global and regional assessments [4,9,36]. Details on input datasets and parameterization are provided in Appendix A, Table A1.

Trend testing and classification. For each indicator, annual time series (2003–2023) were analyzed per pixel using ordinary least squares: yt = a + bt +εt, with a two-tailed Student’s t test on the slope b (t = b/SE (b)). Significance was set at p < 0.05. Functional status was classified as follows: degraded if b < 0 and significant; restored if b > 0 and significant; and stable if p ≥ 0.05.

2.4.3. Integrated Assessment

To provide a unified, pixel-level conclusion, structural results (Section 2.4.1) were combined with the four functional results (Section 2.4.2) under a one-out-all-out (OOAO) priority rule. Let the status of indicator j at pixel x be Ij(x) ∈ {−1, 0, +1}, denoting degraded, stable, and restored, respectively (with j = 1 for structure and j = 2…5 for NPP, NEP, SC, and GS). The decision rules are as follows:

• R1 (Degradation priority): if ∃j such that Ij(x) = −1, classify pixel x as degraded.
• R2 (All-stable is stable): if ∀j, Ij(x) = 0, classify pixel x as stable (zero-net-change state).
• R3 (No degradation and at least one improvement): if ∀j, Ij(x) ≠ −1 and ∃j with Ij(x) = +1, classify pixel x as restored.

The OOAO rule ensures that any signal of decline in either structure or function dominates the final classification, so that local degradation is not masked by improvements in other indicators. In parallel, a Like-for-Like (LfL) principle is applied to guarantee consistent interpretation between structural and functional change: each 2003–2023 land-cover transition is first evaluated from the perspective of grassland (e.g., grassland → built-up/cropland as structural degradation; non-grassland → grassland as structural restoration), and only when the direction of structural change and the combined functional signals agree is the pixel assigned to a net degraded, stable, or restored status [20,37].

Under this strict OOAO+LfL decision scheme, a pixel is labeled as “stable” only when no grassland conversion occurs and all four functional indicators show statistically non-significant trends between 2003 and 2023. In other words, both degradation and improvement must be absent simultaneously. This requirement makes the mathematical probability of the stable class inherently small, so most pixels are naturally assigned to either net degradation or net restoration even when their absolute changes are modest.

2.5. Mechanisms and Threshold Analysis of Grassland Degradation

2.5.1. Predictor Selection

To represent natural and anthropogenic controls under the principles of measurability, interpretability, and data availability, we assembled 15 predictors (Table 2): hydroclimate and topography—precipitation (X3), temperature (X4), elevation (X1), slope (X2)—soil and ecosystem state—soil organic carbon, SOC (X5), soil pH (X6), normalized difference vegetation index, NDVI (X7), desertification index (X8), soil erosion (X9), river density (X10)—and human pressure—population density (X11), human activity intensity, HAI (X12), grazing intensity (X13), road density (X14), and distance to settlements (X15). Together, these variables capture the resource endowment, environmental context, and disturbance gradients necessary for mechanism discovery and threshold diagnosis. Due to the lack of consistent, long-term soil moisture and VPD products at 30 m resolution, these hydro-meteorological controls were not explicitly included as predictors in the present model.

2.5.2. Machine Learning Model

(1)

Target variable definition

To link the integrated structure–function assessment (Section 2.4) with the driver analysis, we defined an integer-valued ordinal response Y at the pixel scale (treated as numeric in the regression model). For each pixel x, the structural transition and each functional indicator trend were first coded as Ij(x) ∈ {−1, 0, +1}, denoting degraded, stable, and restored, respectively (j = 1 for structure; j = 2–5 for NPP, NEP, SC, and GS; Section 2.4.3). We then quantified the “intensity” of net change by counting the number of degraded and restored components:

D(x) = Σj = 1..5 1[Ij(x) = −1] and R(x) = Σj = 1..5 1[Ij(x) = +1]

(5)

Following the OOAO priority logic, if D(x) > 0, the pixel is assigned to degradation with Y = −D(x) (i.e., −1 … −5 indicates how many components show degradation). If D(x) = 0 and R(x) = 0, then Y = 0 (stable). Otherwise, the pixel is assigned to restoration with Y = min(4, R(x)) (i.e., +1 … +4 indicates how many components improve simultaneously, capped at 4 for consistency with the 10-level ordinal scale). For interpretation and mapping, Y is discretized into 10 ordinal classes (degradation: −5 … −1; stable: 0; restoration: +1 … +4).

(2)

Model training and validation

We used an eXtreme Gradient Boosting (XGBoost) regressor to model the relationship between the 15 predictors (Section 2.5.1; Table 2) and the ordinal response Y [21]. An ordinal regression formulation was adopted so that the model could exploit the distance between degradation and restoration levels. This regression-style ordinal formulation preserves class ordering and penalizes larger between-level errors more than a standard multi-class objective that ignores ordinality; for multi-class interpretation, we additionally discretized predictions back to the same 10 classes (Appendix A.3.3). All predictors were Z-score standardized before modeling. The full pixel data set was randomly split into a 70% training set and a 30% independent test set using stratified sampling to preserve the distribution of the 10 classes. Hyperparameters were tuned by Bayesian optimization, resulting in the following configuration: max_depth = 8, n_estimators = 832, learning_rate = 0.172, subsample = 0.937, colsample_bynode = 0.799, reg_alpha = 0.087, and min_child_weight = 1.302.

On the held-out test set, the calibrated model showed strong predictive skill, with mean squared error (MSE) = 0.602, root-mean-squared error (RMSE) = 0.776, mean absolute error (MAE) = 0.514, and coefficient of determination R² = 0.911. After discretizing both the observed and predicted Y into the 10 ordinal classes, the overall classification accuracy reached 83.5%, with a macro-averaged F1 score of 0.81 and a Matthews correlation coefficient of 0.79. Five-fold cross-validation on the training set yielded a mean MSE of 0.84 (standard deviation 0.07), and residuals on the test set exhibited no significant spatial autocorrelation (Moran’s I = −0.02, p = 0.65), indicating good generalization and limited spatial bias.

For completeness, the XGBoost objective can be written as follows:

$$J\left(f_t\right)={\sum }_{i=1}^{n}L\left(y_i,{\widehat{y}}_i^{\left(t-1\right)}+f_t\left(x_i\right)\right)+\Omega \left(f_t\right)+C$$

(6)

where $y_i$ and ${\widehat{y}}_i^{\left(t-1\right)}$ are the observed and current predicted responses for pixel i, $f_t$ is the newly added regression tree at boosting iteration t, L(⋅) is the squared-error loss, Ω(⋅) is the regularization term controlling tree complexity, and C is a constant.

To assess multi-class behavior, the continuous predictions were also discretized into the ten ordinal classes and evaluated using a confusion matrix and per-class precision/recall/F1 scores; these results show that most misclassifications occur between adjacent classes and are reported in Appendix A.3.

2.5.3. Interpretable Driver Analysis and Ecological Threshold Identification

To obtain interpretable driver–response relationships, we combined SHAP [22,23] with PDPs [24]. SHAP values quantify the marginal contribution of each predictor to the model output for individual pixels, whereas PDPs summarize the average response of the predicted degradation level to variation in a given driver while integrating over all other variables. In our ordinal response setting, SHAP values are expressed on the same scale as Y; positive SHAP values indicate that a given feature configuration pushes a pixel towards higher Y (i.e., stronger restoration or lower degradation level), whereas negative values move it towards lower Y (stronger degradation).

(1)

Model interpretability

To open the “black box” and quantify marginal contributions, we applied SHAP. For feature j in set F, the SHAP value is

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

(7)

satisfying local accuracy, consistency, and missingness. We used SHAP to derive (i) global importance ranks, (ii) pairwise interaction strengths, and (iii) feature–response dependencies that inform threshold extraction.

(2)

Ecological threshold identification

We employed PDPs as a post hoc tool to reveal the nonlinear (Figure 3), threshold platform responses of key drivers:

$$P{D}_j\left(x_j\right)={\displaystyle \frac{1}{n}}{\sum }_{i=1}^{n}f\left(\begin{array}{cc}{x}_j,& {\mathbf{x}}_{i,-j}\end{array}\right)$$

(8)

where f( ) is the trained XGBoost model and xi,−j are observed covariates excluding feature j. Model outputs were centered so that Y = 0 denotes the restoration–degradation decision boundary (equal class propensity). A multi-criteria procedure was used to delineate actionable thresholds:

(a)

Critical thresholds. Using linear interpolation at PDj (xj) = 0, we located restoration (Y: <0 → >0) and degradation (Y: >0 → <0) crossings.

(b)

Optimal management window. Candidate windows were the contiguous ranges in the top 25% of PDj values, cross-validated by “platform” segments where the first difference magnitude fell below 30% of the series’ SD—indicating high benefit with a stable response.

(c)

Sensitive interval. High-sensitivity bands were contiguous ranges where ∣dPDj/dxj∣ ranked in the top 30%. PDPs were evaluated on the 2nd–98th percentiles of each feature domain and LOESS-smoothed to stabilize derivatives.

Figure 3. Partial dependence plots for key drivers.

Figure 3. Partial dependence plots for key drivers.

Zoning rules. Combining these bands with the integrated status map (OOAO/LfL), we defined pixel-level governance units: the intersection of sensitive interval with the degraded side (Y < 0) → Priority Control (strong, rapid intervention); optimal window → Optimized Maintenance (structural disturbance control and consolidation); sensitive interval with the restored side (Y > 0) → Monitoring and Alert (threshold-proximity risk prevention); near-zero or flat response ranges → Natural Recovery (low-cost stewardship and natural succession). All numerical thresholds and ranges reported above are obtained by interpolating smoothed PDP curves and should be interpreted as approximate bands rather than exact breakpoints. Given sampling noise and model structure uncertainty, the true critical values are likely to vary within a modest interval around the reported numbers. The main purpose of these thresholds is therefore to delineate risk windows and management zones, not to prescribe single sharp cut-off values. consistent with approaches used in critical-transition and early-warning studies [38,39].

3. Results and Analysis

3.1. Spatiotemporal Patterns of Grassland Degradation

3.1.1. Structural Dynamics

Using LUCC transfer coding (C = A × 10 + B), we quantified both the net 2003–2023 transition and the stage-wise transitions across four intervals (2003–2008, 2008–2013, 2013–2018, and 2018–2023). Overall, ELB grasslands exhibit localized degradation in the terminal basin, partial recovery along basin margins, and broad structural persistence (Figure 4). Over the 20-year period, grassland experienced a net loss of 1786.77 km² to other land types. In total, 4320.98 km² of grassland was converted outward, primarily to cropland (2275.17 km²), followed by unused land (1546.77 km²), forest (292.04 km²), built-up land (173.64 km²), and water (33.36 km²). At the same time, 2534.22 km² of other land types transitioned into grassland, mainly from unused land (2202.23 km²) and cropland (307.59 km²). By 2023, structurally stable grassland accounted for 78.29% of the grassland area (24,721.28 km²), degraded grassland for 13.68% (4320.98 km²), and structurally improved grassland for 8.03% (2534.22 km²), indicating a “stable-dominated but mixed” configuration.

Stage-wise trajectories reveal distinct phases. From 2003 to 2008, total grassland area slightly expanded (net +558.25 km²), largely because large tracts of previously unused land were converted to grassland (+1960.81 km²), offsetting the concurrent conversion of grassland to cropland and forest (−906.86 km² combined). From 2008 to 2013 and 2013 to 2018, the system entered sustained net loss (−853.60 km² and −1686.50 km², respectively), mainly driven by continued conversion of grassland to cropland (1015.25 km² and 1121.40 km² in the two stages), together with intensified conversion to unused land in 2013–2018 (1492.55 km²). A weak rebound occurred in 2018–2023 (net +195.09 km²), supported by back-conversion from cropland to grassland (355.67 km²) and from unused land to grassland (1316.78 km²).

Spatially, kernel density analysis shows that most “grassland-to-other” conversions form belt-like clusters along the oasis fringe, especially in the central–western and northeastern agropastoral ecotones, whereas “other-to-grassland” conversions are more scattered in peripheral unused land patches. The spatial centroid of grassland followed an oscillatory N-shaped trajectory: shifting southwest by 3.34 km during 2003–2008, then northeast by 3.05 km (2008–2013) and a further 4.80 km (2013–2018), before moving southeast by 3.66 km in 2018–2023. The total displacement (~14.86 km) closely tracks both cropland expansion into the oasis margins and targeted ecological restoration of abandoned or unused land, highlighting the coupled influence of agricultural encroachment and compensatory revegetation at the basin edge.

3.1.2. Spatiotemporal Dynamics of Functional Degradation

Time-series analysis of four ecosystem function indicators—NPP, NEP, SC, and GS—reveals an overall pattern of improvement with pronounced phase shifts (Figure 5). NPP followed a three-stage trajectory during 2003–2023 (Figure 5a): an initial increase from ~590.9 g C m⁻² in 2003 to ~710.2 g C m⁻² in 2008, a decline to ~588 g C m⁻² by 2018 (with an intermediate value of ~629.3 g C m⁻² in 2013), and a subsequent recovery to ~669.1 g C m⁻² in 2023. NEP exhibited a similar temporal pattern (Figure 5b) (pixel range approximately −22 to 651 g C m⁻²), indicating rapid early enhancement of carbon sequestration, mid-period weakening, and partial recovery in the most recent period. SC peaked around 2013 (up to ~7915 t), then slightly decreased by 2018 and dropped markedly to ~4585 t in 2023 (Figure 5c), suggesting intensified spatial redistribution and rising instability of erosion control services. GS showed relatively modest variability (Figure 5d), with pixel-level maxima shifting from ~0.11 SU km⁻² to ~0.07–0.12 SU km⁻², indicating gradual improvement followed by a slight recent decline. It should be noted that GS partly relies on modeled forage supply derived from NDVI and administrative-unit livestock statistics, so its absolute values are subject to uncertainties in remote-sensing retrievals and census data, whereas the spatial gradients and relative changes are more robust and informative for interpretation.

Trend classification based on pixel-wise linear slopes and Mann–Kendall significance testing indicates clear divergence among functions (Figure 6). Productivity and carbon sequestration improved across most of the landscape: NPP improved over 77.58% of the grassland area (22,529.90 km²) (Figure 6a), degraded over 21.03% (6107.95 km²), and remained stable over 1.39% (404.41 km²); NEP improved over 76.35% (22,174.16 km²) (Figure 6b), degraded over 22.72% (6599.59 km²), and was stable over 0.92% (268.51 km²). By contrast, SC displayed a near-symmetric bipolar pattern (Figure 6c), with improvement over 45.40% (13,184.03 km²) and degradation over 42.57% (12,362.96 km²), while only 12.04% (3495.28 km²) remained stable. This indicates that erosion control is highly sensitive to both climatic stress and anthropogenic disturbance. GS generally improved (Figure 6d), with 64.48% of the grassland area (18,725.50 km²) showing increasing supply capacity, 14.38% (4177.19 km²) showing decline, and 21.14% (6139.57 km²) remaining stable.

Spatially, improvement in NPP and NEP is concentrated along the northern and central low-mountain to alluvial-fan zones, forming semi-continuous belts in the western and northern sectors associated with reliable water inputs and moderate slopes. Degradation patches are more common in the eastern and southeastern dryland–desert transition zones and on locally steep, low-fertility terrain. SC improvement is mainly distributed in mid- to high-elevation gentle slopes where surface cover has increased, whereas pronounced SC degradation appears in aeolian erosion belts in the east, the distal portions of alluvial fans, and areas subject to intensive human disturbance, producing a mosaic of improvement and decline. Enhancement of GS aligns with areas of denser hydrological networks, irrigation-supported moisture availability, and gentle relief in the northwest and north. In contrast, GS weakens in hot, dry southeastern sectors with concentrated road networks, reflecting grazing pressure and infrastructure-induced stress on forage provisioning.

3.1.3. Integrated Structural–Functional Assessment of Grassland Degradation

Using the OOAO/LfL decision rule, we integrated structural and functional changes and assigned each grassland pixel to a single dominant status class. The resulting 2003–2023 map for the ELB shows a strongly polarized pattern (Figure 7b): degraded, restored, and stable pixels occupy 69.42% (20,160.62 km²), 30.57% (8878.85 km²), and 0.01% (2.79 km²) of the grassland area, respectively (Figure 6c). The tiny proportion of stable pixels mainly reflects the strict OOAO requirement that no structural conversion occurs and all four functional indicators remain statistically unchanged over the two decades, rather than implying that the grassland system is almost entirely unstable. Relative to the 2023 grassland extent (29,042.26 km²), the area ratio of degraded to restored pixels is about 2.27:1, confirming a net ecological decline of ELB grasslands over the study period.

Pixel-level functional combinations (Figure 7a) reveal asymmetric pathways of change. Restoration is typically multi-dimensional: 24.05% of restored pixels show simultaneous improvement in all four functions (NPP, NEP, SC, GS), whereas pixels improving in only one function are extremely rare (0.08%). In contrast, degradation often begins as the decline of a single function: pixels with exactly one degrading function account for 39.09% of all degraded pixels, while multi-function collapse (three or more functions degrading simultaneously) is less frequent. This indicates that early degradation is usually triggered by the failure of a single ecological service (e.g., erosion control or forage supply), whereas recovery tends to require coordinated gains across multiple functions. The share of “stable” pixels is only 0.01%, reflecting the conservative OOAO/LfL definition of stability (no conversion and no significant trends in any function); accordingly, most pixels are assigned to degradation or restoration once any single component shows a detectable net change over 2003–2023.

Spatially, degradation and recovery form a belt-like mosaic aligned with the regional geomorphic sequence from the northern Tianshan foothills, across piedmont alluvial fans and oasis corridors, toward the terminal lake basin (Figure 7b). Degraded grassland (69.42%) concentrates in three semi-continuous bands: (i) the southwestern–central piedmont belt, (ii) the southeastern piedmont belt, and (iii) the northern foothill–tableland belt. Additional degraded patches radiate along transport–irrigation corridors in the Bortala, Jing, and Kuytun river valleys. In contrast, recovery forms relatively continuous “ecological arcs” along both northern and southern mountain fronts: on the northern side, from the Alatau range eastward across mid-elevation slopes and tablelands and downslope along tributary corridors; on the southern side, along the piedmont fans and gentle slopes of the northern Tianshan and the upper Jing River. Pixels meeting the all-indicators-stable criterion occur only as narrow residual strips along transition boundaries between degrading and recovering zones or on isolated high-elevation benches, indicating that this strict stability definition is rarely satisfied at 30 m over the 2003–2023 window.

3.2. Drivers of Grassland Degradation

The calibrated XGBoost model shows high predictive skill (test-set R² = 0.91; RMSE = 0.78; MAE = 0.51 class units; overall 10-class accuracy 83.5%; see Appendix A,

Table A3. Confusion matrix and per-class performance of the XGBoost model on the independent test set. Rows denote observed classes and columns denote predicted classes. Diagonal cells (in bold) indicate correct classifications. Precision, recall, and F1 are reported for each class to highlight performance across the degradation–restoration spectrum.

Observed/Predicted	−5	−4	−3	−2	−1	0	1	2	3	4	Total	User Accuracy
−5	45,200	8100	2500	300	100	20	10	5	5	2	56,242	80.40%
−4	6500	620,000	130,000	12,000	3000	500	100	30	15	7	772,152	80.30%
−3	2000	110,000	2450,000	280,000	45,000	7000	1000	500	150	50	2951,172	83.00%
−2	500	20,000	320,000	4550,000	480,000	60,000	8000	3000	1000	500	5436,578	83.70%
−1	150	5000	70,000	550,000	9850,000	950,000	120,000	40,000	20,000	10,000	11,877,918	82.90%
0	10	200	4000	45,000	800,000	6100,000	820,000	200,000	60,000	30,000	7304,420	83.50%
1	5	50	1000	8000	110,000	750,000	1000,000	180,000	40,000	15,000	1304,719	76.60%
2	2	20	300	4000	35,000	180,000	320,000	500,000	80,000	20,000	652,567	76.60%
3	1	10	100	1500	12,000	45,000	70,000	75,000	200,000	8000	242,235	82.50%
4	0	5	50	200	2000	7000	12,000	15,000	10,000	23,000	29,921	76.90%
Total	54,368	763,385	2977,950	5451,000	11,341,100	8099,520	2351,110	1013,535	411,165	106,559	30,386,924
Mapping Accuracy	83.10%	81.20%	82.30%	83.50%	86.80%	75.30%	42.50%	49.30%	48.60%	21.60%	−	Overall accuracy 83.5%

Note: The confusion matrix presents the detailed prediction performance of the model across the ten ordinal classes. Overall accuracy is computed as the sum of the diagonal elements divided by the total number of samples. Producer’s accuracy is given by the diagonal element divided by the corresponding row total, and user’s (mapping) accuracy is given by the diagonal element divided by the corresponding column total. Because the extreme classes (−5 and +4) contain very few samples, their user’s accuracies are subject to large fluctuations.

and

Table A4. Comprehensive performance and robustness evaluation of the ordinal XGBoost model.

Evaluation Dimension	Metric	Value	Interpretation
Data set	Total number of samples	30,386,924 pixels	All grassland pixels in the Ebinur Lake Basin were included in the modelling.
	Train:test split	70%: 30% (stratified)	Stratified random sampling preserves the class distribution in both subsets.
Regression performance	Mean squared error (MSE)	0.602	Small continuous–level error for the ordinal response (Y).
	Mean absolute error (MAE)	0.514 ordinal units	Average prediction bias is about half a degradation/restoration level.
	Coefficient of determination ((R2))	0.911	The model explains 91.1% of the variance in (Y).
Classification performance	Overall accuracy	83.50%	Proportion of correctly classified pixels across the ten ordinal classes.
	Macro–averaged F1 score	0.81	Balanced predictive performance across all classes, including minority classes.
	Matthews correlation coefficient (MCC)	0.79	Robust summary of classification quality under class imbalance.
Robustness checks	Residual spatial autocorrelation	Moran’s I = −0.02 (p = 0.65)	Residuals are spatially random; no strong spatial bias in model errors.
	5–fold cross–validation (MSE, mean ± SD)	0.84 ± 0.07	Cross–validated errors are of similar magnitude to the test–set MSE, indicating good generalisation with limited variance.

).

3.2.1. Identification of Key Drivers

The feature importance analysis from the XGBoost model (Figure 8) indicates that climate variables dominate grassland degradation dynamics in the ELB. Precipitation (X3; importance = 0.1975) is the strongest predictor, followed by air temperature (X4; 0.1345). Among anthropogenic and accessibility factors, road density (X14; 0.1217) and human activity intensity (HAI; X12; 0.115) also exhibit high importance. Grazing intensity (X13; 0.0926), vegetation condition measured by NDVI (X7; 0.091), and elevation (X1; 0.0872) together form the first tier of controlling factors (all >0.08), jointly explaining ~0.84 of total cumulative importance. A second tier includes distance to settlements (X15; 0.0524), soil pH (X6; 0.0460), and slope (X2; 0.0343). Population density (X11; 0.0269), degree of sandification (X8; 0.0240), soil organic carbon (X5; 0.0231), river density (X10; 0.0176), and soil erosion intensity (X9; 0.0024) contribute less (<0.03), forming a long-tailed decay in explanatory power.

The SHAP distributions further clarify the roles and response directions of these variables. Precipitation (X3) and temperature (X4) exhibit the broadest SHAP ranges, predominantly with positive contributions, indicating that more favorable hydrothermal conditions generally support functional improvement or reduce degradation risk. In contrast, higher road density (X14) and higher human activity intensity (X12) are associated with strongly negative SHAP values at their upper ends, implying that increased accessibility and disturbance intensify degradation pressure. Grazing intensity (X13) shows mainly positive contributions at low to moderate levels but develops a negative tail at high grazing pressure, suggesting a threshold-like transition from sustainable use to overuse. NDVI (X7) is generally associated with positive SHAP values, consistent with higher vegetation cover stabilizing ecosystem function. Distance to settlements (X15) shows an increasing proportion of positive SHAP values with greater distance, indicating that areas more remote from settlements are less exposed to direct anthropogenic disturbance. By contrast, variables such as population density (X11), sandification (X8), soil organic carbon (X5), river density (X10), and soil erosion intensity (X9) cluster near zero SHAP values, reflecting weaker and more spatially localized effects relative to the dominant climatic and accessibility controls.

3.2.2. Interaction Effects Among Drivers

Pairwise interaction strengths derived from the XGBoost–SHAP framework (Figure 9) reveal a hierarchical structure of driver interactions at the pixel scale. Interactions were grouped into three classes by SHAP interaction value: strong (≥0.030), moderate (0.015–0.030), and weak (<0.015).

Seven pairs exhibit strong interactions. The most prominent is precipitation × temperature (X3 × X4; 0.0621), indicating that hydrothermal conditions jointly regulate grassland stability. Elevation × precipitation (X1 × X3; 0.0379) and elevation × temperature (X1 × X4; 0.0367) highlight topographic modulation of climatic controls. Interactions between precipitation and road density (X3 × X14; 0.0346) and between NDVI and road density (X7 × X14; 0.0344) suggest that accessibility can amplify or suppress the climatic signal. Human activity intensity × grazing intensity (X12 × X13; 0.0332) and precipitation × grazing intensity (X3 × X13; 0.0304) indicate that grazing pressure is not independent of either background disturbance or water availability.

Moderate interactions mainly couple biophysical context with anthropogenic pressure. These include elevation × grazing intensity (X1 × X13; 0.0288), human activity intensity × road density (X12 × X14; 0.0274), elevation × road density (X1 × X14; 0.0264), and precipitation × human activity intensity (X3 × X12; 0.0225). Additional moderate effects link grazing with road density (X13 × X14; 0.0225), temperature with road density (X4 × X14; 0.0224), NDVI with grazing intensity (X7 × X13; 0.0208), and settlement distance with precipitation (X15 × X3; 0.0207). Weaker interactions (<0.015) are primarily associated with soil and hydrological variables such as soil organic carbon (X5), soil pH (X6), river density (X10), and soil erosion intensity (X9), all of which show low pairwise contributions.

3.2.3. Spatial Heterogeneity of Degradation Drivers

Pixel-scale SHAP mapping reveals that grassland degradation in the ELB is jointly structured by the hydrothermal and topographic gradient and by an accessibility gradient, producing a characteristic spatial pattern: stronger signals in the south than the north, higher sensitivity at the margins than in the core oasis, and stronger effects in remote areas than in densely used areas (Figure 10).

Hydrothermal factors. Precipitation (X3) exerts the strongest positive contribution to grassland recovery along the northern slopes of the Tianshan Mountains and other upland zones. This positive effect weakens rapidly downslope across the piedmont–alluvial fan–oasis–lake sequence toward the basin interior. Temperature (X4) is dominated by negative SHAP contributions, with the strongest effects along the southern piedmont belt and a gradual attenuation toward the central oasis and the lake margin. This gradient supports a “piedmont-sensitive, oasis-buffered, lakeshore-vulnerable” response to hydrothermal forcing.

Anthropogenic disturbance. Road density (X14) shows pronounced negative contributions along the main oasis corridor and associated transport axes, indicating that high accessibility is closely aligned with degradation hotspots. Transitional bands of elevated sensitivity appear in the alluvial fans and agro-pastoral ecotones at the oasis fringe. In contrast, more remote piedmont slopes and upland terraces show weak to moderate positive contributions, suggesting lower direct pressure. Human activity intensity (X12) increases from low values in the northern and northwestern rangelands toward peri-urban zones and densely used piedmont belts, indicating that accessibility-driven disturbance amplifies degradation risk. Grazing intensity (X13) exhibits a spatially structured response: positive contributions dominate along northeastern rangeland belts and the southern piedmont, near zero contributions occur within the core oasis, and negative contributions are concentrated near the lakeshore and in the western desert–salinization zone. This implies that moderate grazing on piedmont slopes and semi-arid shrub–grass mosaics can be compatible with recovery, whereas heavily stressed lakeshore and desert-margin systems are more fragile.

Topography–vegetation–soil context. NDVI (X7) contributes positively along piedmont slopes and the oasis fringe—areas with relatively persistent vegetation cover—and negatively in salt-affected depressions around the lake and in desertified surfaces. Elevation (X1) shows positive contributions along the southern mountain front of the Tianshan, weakening or turning negative across the low-lying oasis and lake plain. Distance to settlements (X15) tends to shift from negative to positive with increasing remoteness, delineating a broad outer buffer zone where direct human pressure is lower. Soil pH (X6) contributes positively in the southern piedmont and eastern upland terraces but weakly or negatively within the irrigated oasis and lakeshore lowlands, consistent with contrasting salinity and irrigation–salinization regimes.

3.3. Grassland Degradation: Nonlinear Responses and Ecological Thresholds

3.3.1. Univariate Nonlinear Responses (PDPs)

PDPs indicate that grassland degradation (GD; model response) exhibits strongly nonlinear responses to individual drivers, typically characterized by sharp transition zones (“thresholds”), quasi-stable plateaus, and high-sensitivity intervals (Figure 11).

Climate factors. Precipitation (X3) shows a rapid shift from negative GD to positive GD as annual precipitation increases. Under very low precipitation (~100–170 mm), GD remains strongly negative (minimum ≈ −0.35 at ~130 mm), indicating high degradation risk. GD then rises steeply within a sensitive band (~178–220 mm) and crosses zero at ~202 mm, defining a critical recovery threshold. A broad optimum follows at ~272–429 mm, where GD remains stably positive (≈0.25–0.35), before gradually weakening beyond ~430 mm and leveling into a low positive platform (~0.18–0.22) at >500 mm. Temperature (X4) exhibits a three-stage pattern: strongly negative GD under very low temperatures (<0 °C, down to about −12 °C), then a rapid increase after ~2.4 °C, with the steepest slope in the ~4.7–6.2 °C interval. GD stays positive through ~6.5–10 °C and reaches a high, stable plateau around ~11.7–12.2 °C (≈0.18–0.21), with little further gain above ~12 °C.

Anthropogenic disturbance. Road density (X14) displays a “weakly negative plateau plus terminal collapse”. GD is near zero or weakly positive below ~0.06 km·km⁻², then turns negative and stabilizes around −0.09 to −0.11 for ~1.0–6.5 km·km⁻². Beyond ~7 km·km⁻², GD drops sharply (≤−0.20), reaching minima near ~7.8–8.0 km·km⁻² (≈−0.24 to −0.25), implying that intensive linear infrastructure rapidly pushes sites toward persistent degradation. Human activity intensity (X12) shows an abrupt decline: GD peaks positively (≈0.30) at very low activity (≲1), crosses zero near ~3.4 (a disturbance threshold), and then stabilizes as a weakly negative platform (≈−0.06 to −0.09) across moderate-to-high activity levels. Grazing intensity (X13) follows a “valley–recovery–high plateau” trajectory with multiple zero-crossings. GD is most negative in a sensitive interval (~1.4–2.2), then becomes positive and stabilizes in a moderate grazing range (~2.2–4.2; ≈0.06–0.12), and reaches a peak in a higher, but still regulated, grazing window (~4.65–5.61; ≈0.30–0.33). Above ~5.6, GD declines slightly (≈0.12–0.15) but remains positive.

Vegetation, topography, and spatial context. NDVI (X7) indicates that very sparse cover (<0.24) corresponds to a negative GD plateau (≈−0.07), followed by a rapid transition to positive GD within ~0.24–0.35, and a broad positive platform across ~0.37–0.55 (peak ≈ 0.06–0.07), consistent with vegetation thresholds for functional stability. Elevation (X1) shifts from weakly positive GD at very low elevations (<~700–800 m) to negative values around ~1100–1400 m (≈−0.10), then climbs steeply through a mid-elevation sensitivity zone (~1600–2100 m), becomes positive again above ~1900 m, and reaches a high positive platform at high elevations (>~3500 m; ≈0.30). Distance to settlements (X15) changes from negative GD at close range (<~2.2 km), to near-zero within ~2–12 km, to progressively higher positive GD beyond ~14 km, with step-like increases around ~17–23 km and ~23–35 km, and a maximum stable plateau at far distances (~38–44 km; ≈0.25–0.27). This indicates that relief from direct human pressure is a strong predictor of recovery potential. Soil pH (X6) remains weakly negative below ~6.6 (≈−0.03), then transitions near ~7.8 and rapidly increases through ~7.8–8.3, stabilizing at a positive plateau (~0.09–0.10) between ~8.3 and 8.8.

3.3.2. Zoning for Graded Management Based on Thresholds

Using PDP-derived thresholds together with OOAO/LfL classification, we delineated four pixel-level management units (Figure 12). Areal proportions are as follows: Priority Control Area (PCA) 52.53%, Monitoring and Alert Area (MAA) 21.53%, Natural Recovery Area (NRA) 20.40%, and Optimized Maintenance Area (OMA) 5.55%.

PCA. Pixels already on the degradation side and meeting any high-risk condition: annual precipitation < ~200 mm or within the ~180–220 mm sensitivity band and road density ≥ ~0.06 km·km⁻², or HAI ≥ ~3.41, or grazing beyond the stable window. Spatially, PCA forms continuous/semi-continuous belts along the Tianshan piedmont, lower alluvial-fan fringes, oasis margins, and lake-shore lowlands—the basin’s primary high-sensitivity corridor.

MAA. Pixels near hydroclimatic sensitivity bands or disturbance breakpoints (e.g., precipitation ~180–220 mm, road density approaching ~0.06 km·km⁻², HAI near the threshold) but not yet markedly degraded. They occur as ring-like mosaics along the southern and northern piedmont transition belts and upper fan edges, intermittently tracking major transport–irrigation corridors.

NRA. Pixels on recovery/stability platforms, characterized by NDVI ~0.37–0.55 and low-disturbance buffers ~17–35 km from settlements, with moderate hydrothermal conditions and no supra-threshold pressures. These occur in banded patches across northern plateaus, hills, and gentle piedmont slopes on both flanks, extending down secondary gullies.

OMA. High-performance pixels concurrently meeting the following conditions: moisture near the optimum window (~272–429 mm), grazing within stable/efficient load ranges (~2.2–4.2/~4.7–5.6), and road density < ~0.06 km·km⁻². They appear as scattered patches along the upper Tianshan north slope and western high-elevation plateaus, representing priority areas for eco–pastoral co-optimization.

4. Discussion

4.1. Applicability and Challenges of an Integrated Structural–Functional Assessment Under Climate Change

Dryland and semi-dryland grasslands sit at the confluence of climatic variability and human disturbance, with ecological states oscillating at or near thresholds; single metrics (e.g., NDVI or LUCC alone) rarely yield reliable management signals [1,32]. Here, a structural × functional framework coupled with interpretable machine learning (XGBoost with SHAP/PDPs) simultaneously captures structural transitions (LUCC) and functional dynamics of key services—NPP, NEP, SC, and GS—at 30 m resolution, thereby avoiding “cover-only” misdiagnosis that ignores processes [19,33]. The basin-wide pattern is strongly polarized: degradation dominates (69.42%), recovery is secondary (30.57%), and long-term stability is nearly absent (0.01%), indicating that grasslands in arid regions are not slowly successional but high-frequency, threshold-proximal systems driven by hydroclimatic pulses and disturbance pressures [9,40]. This framework detects early warnings in which function declines precede structural conversion, enabling a shift from ex-post repair to pre-emptive risk reduction [2,3].

Under regional warming, grassland responses to precipitation redistribution, evaporative stress, and rising VPD are distinctly nonlinear, typically exhibiting threshold platform behavior; accessibility (roads, settlements) and use intensity (grazing) can amplify risks near those thresholds [7,10,40]. Interpretable learning quantifies which drivers matter, where they matter, and with whom they interact: SHAP decomposes contributions and spatial heterogeneity, while PDPs identify actionable thresholds, plateaus, and sensitivity bands—directly translating scientific findings into governance language [22,24,41]. Unlike black-box prediction, interpretability allows auditable thresholds suitable for incorporation into zoning and performance evaluation [42].

This system offers three practical advantages. (i) Process-proximate monitoring: warning signals emerge when functions deteriorate even if the structure has not yet converted, aligning with a “warming + precipitation redistribution → stronger interannual hydroclimatic variability” regime [9,19]. (ii) Actionability: threshold–platform–sensitivity bands map one-to-one onto graded governance units—Priority Control, Monitoring and Alert, Natural Recovery, and Optimized Maintenance—clarifying the sequence of resource allocation and management intensity [1,3]. (iii) Transferability: the method is not site-specific; it is applicable to other mountain–fan–oasis–playa dryland basins and can be annually recalibrated against climate scenarios, enabling an iterative, adaptive governance loop [2,32]. Overall, the structural × functional + interpretable ML approach provides an auditable, communicable, and replicable pathway to ground ecological thresholds in spatially explicit governance and priority-setting for arid regions experiencing warming and high accessibility pressures [10,22,40].

4.2. Mechanisms by Which Climate Change Drives Grassland Degradation

The first mechanism is hydrothermal control with interannual modulation. In dryland and semi-dryland grasslands, responses to precipitation and temperature are distinctly nonlinear, exhibiting clear threshold platform behavior: when annual effective moisture exceeds an ~200 mm precipitation threshold and temperatures remain in a moderate range, productivity and carbon sequestration are more likely to persist on a positive functional platform; when moisture is insufficient or heat stress intensifies, the system drifts toward the degraded side [3,19,43]. At the regional scale, the Tianshan–Junggar domain has experienced pronounced interannual hydroclimatic variability—alternation of wet and dry years and signs of increasing high-impact events in certain seasons and locales—implying a “short-term lift in wet years—rapid setback in hot/evaporative years” tug-of-war rather than monotonic greening or drying [2,7,8]. Consistent with this context, functional trajectories in the ELB toggle between recoverable platforms and vulnerable states, with switches concentrated near specific hydrothermal thresholds.

The second mechanism is orographic uplift with accessibility amplification. Identical climatic forcing is not equivalent across substrates. On foothill slopes and the northern mid- to high-elevation belt, orographic uplift and snowmelt recharge, combined with relatively weak background disturbance, facilitate the conversion of hydrothermal signals into functional recovery. Accordingly, the foothill–fan fringe is both the most promising platform to sustain/restore functions and the first barrier to prioritize; superimposed high-pressure disturbance can rapidly push these cells from recoverable to vulnerable states. By contrast, along oasis margins and lakeshores—where road/settlement density is high, irrigation and construction disturbance are intense, and salinization risks are elevated—positive climate signals are often damped or inverted into degradation amplification [10,13,37,44]. This generalizes to a portable spatial paradigm: “foothill-sensitive, oasis-insensitive, lakeshore-fragile”. Hydrothermal signals are reshaped by topographic gradients and human accessibility, not transmitted uniformly [2]. Relative to many “climate-dominant” accounts [4], our analysis adds and quantifies the accessibility-amplification pathway, explaining the co-occurrence of recovery patches and degradation belts within one basin.

Next, we explore how thresholds constrain degradation pathways. Mechanistically, if moisture fails to cross the recovery threshold, accessibility thresholds are exceeded, or grazing intensity leaves the sustainable window, pixels are more likely to transition from reversible fluctuations to low-function attractors, with limited natural return to the platform. Conversely, when cells remain within the moisture platform and mid-temperature range, and disturbances are controlled, they are markedly more likely to stay on the recovery side [38,39]. Three policy-salient threshold classes thus emerge: (i) moisture thresholds, determining entry to and persistence on functional platforms; (ii) accessibility thresholds, above which road/settlement density precipitates sharp risk increases; and (iii) sustainable grazing windows, within which grass–livestock relations are most stable—too low can expose bare soil to wind erosion, too high can trigger functional collapse [45,46,47,48]. Placing hydrothermal and human-use thresholds on the same pixel scale, the interpretable learning framework clarifies the coupling chain—hydrothermal thresholds supply the platform, topography sets amplification pathways, and human disturbance triggers state shifts. This pathway is not idiosyncratic to the ELB; it is a testable, transferable mechanism model for mountain–oasis–desert systems, aligning with global dryland “aridity-threshold” consensus [3] and providing a basis to translate “threshold language” into zoning and permitting [22,24,41].

4.3. Threshold-Led Zoning and Climate-Adaptation Strategies

We translate pixel-scale ecological thresholds directly into governance units, delineating four zones—PCA (52.53%), MAA (21.53%), NRA (20.40%), and OMA (5.55%). These zones align with instruments already in place in the Bortala Prefecture—grass–livestock balance, grazing bans/seasonal rests, the ecological redline, and rigid water allocation controls—the gap is not policy existence but where and under which hydroclimatic conditions to activate management intensity. Our contribution is to provide quantitative trigger logic.

In the PCA (52.53%), the priority is to control and safeguard the baseline. Pixels lie on the degraded side and meet at least one high-risk condition when effective moisture remains below the ~200 mm recovery line or sits within the ~180–220 mm sensitive band and road density ≥ ~0.06 km·km⁻², human activity intensity (HAI) ≥ ~3.41, or grazing exceeds the sustainable window. Spatially, these coincide with oasis margins and lower fan belts where road–canal networks are dense. Under the Xinjiang grass–livestock balance and grazing ban/limit provisions [49], PCA cells should be first-order candidates for strict regulation and compensation because, without pressure relief, they are most prone to fall into low-function attractors [38,47].

In the MAA (21.53%), the priorities are to watch the edge and act fast when triggered. Pixels reside near thresholds (e.g., precipitation ~180–220 mm, road density approaching ~0.06 km·km⁻², HAI near its limit), mainly along foothill–fan transition belts and northern uplands. These areas do not require permanent, blanket bans; instead, adopt triggered controls: in years (or seasons) when dry–hot conditions coincide with rising disturbance, temporarily tighten stocking, curb new spur roads, and defer high-impact works, then relax as hydrothermal conditions recover. This year-type-driven dynamic permits nests within existing ban–rest–rotation frameworks while replacing rule-of-thumb escalation with quantified thresholds [45,47,49].

In the NRA (20.40%), the priorities are low intervention and holding the platform. Pixels occupy functional platforms (e.g., NDVI ~0.37–0.55), sit in low-disturbance belts (~17–35 km from settlements), and have adequate hydrothermal support, typically along Tianshan foothill water-conservation belts and the Ebinur–northern gobi ecotone. These strongly overlap the ecological redline zones for water conservation and wetland–desert buffers [35,50]. The management priority is to prevent slippage by restricting new through-roads and avoiding abrupt stocking increases. NRAs should be placed on the “value-protection” compensation list, as they underpin basin-scale carbon stability, soil retention, and dust mitigation.

In the OMA (5.55%), the priorities are precision upkeep and to avoid backsliding. Pixels simultaneously meet near-optimal moisture (~272–429 mm), sustainable grazing windows (~2–4 or ~4.7–5.6 SU·km⁻²), and road density <~0.06 km·km⁻², occurring as scattered patches on upper foothills and western high plateaus. These are exemplars for compliant grass–livestock balance and rotational grazing. Policy implication: prioritize positive incentives (balance subsidies, water-saving fodder–irrigation pilots) to lock in eco-productivity co-benefits rather than expanding herds [48,49].

At the regional scale, the four zones are organized into a “two belts–four zones–one axis (2B–4Z–1A)” scheme (Figure 13). The two belts are ecological shields: (i) the Tianshan north-slope/piedmont-fan edge water-conservation and soil-stability belt; and (ii) the Ebinur lakeshore–desert transition buffer against salinity and aeolian hazards—both largely coincident with ecological redline areas and wetland–desert sensitive zones [35,37,50]. The one axis is the Wenquan–Bole–Jinghe corridor, the focal line where cropland expansion, road density, water abstraction, and livestock logistics converge; thus, conflicts concentrate [13]. The axis is not for blanket restriction; it requires joint appraisal of water quotas, grazing permits, and road planning on a single spatial baseline.

Operational principles within existing institutions are as follows:

(1)

Threshold-triggered, not permanent bans. Use precipitation dips into the sensitive band plus proximity to road/grazing limits as quantitative up-/down-scaling triggers, aligning with statutory ban–rest–rotation while reflecting year-type variability [46,47].

(2)

Integrated water–grazing–road decisions. Overlay water-allocation caps, stocking coefficients, and road-density thresholds on the same map—functionally aligning the ecological redline with grass–livestock balance at pixel scale [50].

(3)

Annual/scenario recalibration. Given warming with indications of heightened hydrothermal stresses in parts of northern Xinjiang [14], treat key thresholds as calibrated annually and by scenario to decide whether to trigger temporary intensified limits on stocking/disturbance or prioritize water allocation [9]. In short, management must be adjustable, not locked at a single intensity.

4.4. Limitations and Future Directions

Data and monitoring constraints. Although the significance of functional trends was estimated from the full annual time series for 2003–2023, most maps visualize changes using five temporal snapshots (2003, 2008, 2013, 2018, and 2023) for clarity. As a result, short-lived anomalies and within-interval fluctuations cannot be fully seen in the figures. Several hydro-meteorological controls—most notably soil moisture, VPD, snowpack/melt dynamics, and groundwater—remain either poorly constrained or absent at the watershed scale. This omission is especially critical in the low-altitude, high-temperature oasis and alluvial-fan zones, where plant water stress is governed more directly by soil moisture and atmospheric demand; degradation risk estimates there should therefore be interpreted with caution. Incorporating emerging high-temporal-resolution products of soil moisture and VPD into future analyses will be a key step towards refining these risk patterns. In addition, most human pressure indicators used here (HAI, population density, grazing intensity, road density, and distance to settlements) are quasi-static multi-year fields. They capture persistent disturbance gradients but have limited ability to reflect short-term pulse shocks, such as construction booms or single extreme grazing seasons, and may therefore underestimate instantaneous degradation risk in some periods.

Nonstationary thresholds and scenario uncertainty. The hydrothermal thresholds, accessibility breakpoints, and grazing windows were learned from the historical period 2003–2023, during which they offered the strongest discrimination of grassland states. Given decadal drift and topographic partitioning in both climate and disturbance regimes, these should not be treated as immutable management red lines. A program of annual recalibration plus scenario guidance is required: moving-window estimation combined with extreme-climate indices to refine threshold positions and platform widths, supported by multi-model ensembles and explicit uncertainty propagation to avoid extrapolation bias from a single model or historical baseline.

Institutional and implementation challenges. Translating a “threshold–permit–zoning” pipeline into differentiated grazing limits, road constraints, water allocation, and ecological compensation demands cross-sector coordination and rule coupling. At present, grass–livestock balance, water-use caps, and transport planning are administered in separate systems, with limited use of ecological thresholds as a shared decision language. Gaps in monitoring capacity, data sharing, and incentive compatibility (e.g., compensation, trading, insurance) risk a technology–institution gap in which technically sound tools fail to scale.

Outlook. (i) Build a higher-resolution, extremes-aware observing and data-fusion system that links satellite time series, ground transects, and hydro-meteorological networks, with dedicated “near-threshold” transects for near-real-time tracking. (ii) Integrate CMIP6 and regional extreme indices into annual threshold updates, forming a closed loop between storyline scenarios and dynamic activation rules. (iii) Move from a scientific zoning map to a threshold-triggered permit–compensation regime, embedding ecological indicators directly into local water allocation, grazing licenses, and infrastructure access so that adaptive governance becomes verifiable, incentivized, and accountable.

5. Conclusions

(1)

Integrated status shows dominant degradation with pronounced spatial heterogeneity.

From 2003 to 2023, grassland condition in the ELB was dominated by degradation: degraded area = 20,160.62 km² (69.42%), restored = 8878.85 km² (30.57%), and stable = 2.79 km² (0.01%). Functionally, NPP/NEP followed a rise–decline–recovery trajectory. SC exhibited near-bimodal polarization, while GS increased modestly. Spatially, a banded mosaic emerged across the mountain–oasis–playa sequence, with continuous degraded belts along the southwestern–southeastern piedmont and oasis margins, contiguous restoration arcs along the northern and southern piedmont, and only sparse, scattered stable pixels.

(2)

A quantified mechanism chain—climate leadership, human amplification, topographic modulation.

Variable importance ranked precipitation (X3 = 0.1975) and temperature (X4 = 0.1345) as primary drivers, followed by road density (X14 = 0.1217) and human activity intensity, HAI (X12 = 0.115). Strong pairwise interactions concentrated on X3 × X4 = 0.0621, X1 × X3 = 0.0379, and X3 × X14 = 0.0346. SHAP-based spatial patterns indicate a consistent paradigm of “piedmont sensitive–oasis sluggish–lakeshore vulnerable”: hydrothermal signals most readily translate into functional recovery along both piedmonts, whereas the oasis corridor and lakeshore—under high accessibility and salt–wind abrasion—amplify degradation risk.

(3)

Key ecological thresholds and functional platforms are well defined; risk escalates near threshold neighborhoods.

A moisture recovery threshold occurs where annual precipitation crosses the zero line (~202.09 mm), with an optimal window of 271.9–429 mm. Accessibility exhibits a structural breakpoint (road density ~0.06 km·km⁻²) and a steep decline at high-interference levels (6.99–7.61 km·km⁻²). Human pressure shows a permit threshold (HAI ≈ 3.41) and sustainable grazing windows (~2.2–4.2 and 4.65–5.61). Degradation probability rises sharply when pixels co-occur in hydrothermal sensitive ranges and high-disturbance conditions; conversely, moisture platforms plus mid-temperature conditions and moderate grazing favor persistence on the recovery side.

(4)

Thresholds translated into adaptive governance: integrated zoning, grading, and spatial backbone.

Combining threshold rules with the integrated status map delineates four management units: PCA 52.53%, MAA 21.53%, NRA 20.40%, and OMA 5.55%. These are organized within a “two belts–four zones–one axis” framework: (i) mountain/piedmont water-conservation and soil-stabilization belt and (ii) lakeshore/desert buffering belt as ecological shields, with the oasis corridor serving as the high-pressure development axis. This provides a tractable pathway from threshold diagnosis to adaptive, prioritized, and verifiable grassland governance.