Leveraging Explainable Machine Learning to Decipher Ecosystem Health and Nonlinear Dynamics in the Henan Yellow River Basin

Abstract

Addressing national goals for ecological conservation in the Yellow River Basin, this study focuses on its Henan segment (HYRB). We developed a VOR-SQ assessment framework by augmenting the classic Vitality–Organization–Resilience model with ecosystem services and an enhanced ecological quality indicator. Using multi-source remote sensing and statistical data, we examine the spatiotemporal evolution of ecosystem health in the HYRB from 2000 to 2020. The XGBoost-SHAP algorithm was applied to identify nonlinear drivers and threshold effects. Key findings indicate (1) a persistent “high west, low east” health gradient with an overall declining trend; western mountains remain healthy, while eastern plains, urban, and intensive agricultural areas show degradation. (2) Natural factors—evapotranspiration (ET), elevation, NDVI, and slope—dominate health dynamics, with critical thresholds (~1153 mm, ~457 m, ~0.76, ~10.5°, respectively) beyond which their impacts shift markedly. (3) Anthropogenic factors (GDP, population/road density) contribute less globally but cause strong local negative disturbances in plains. For instance, road density > 434 km/km² or population density > 159 persons/km² reverses their effects from positive to negative. Accordingly, we propose tailored strategies: western conservation, central farmland optimization, and eastern development control. By coupling the VOR-SQ framework with XGBoost-SHAP, this study offers a robust diagnostic tool for ecosystem health and adaptive governance in fragile socio-ecological systems.

1. Introduction

Ecosystems underpin Earth’s life-support systems, driving the fundamental processes of hydrological cycles, energy flow, and biogeochemical dynamics. They provide the indispensable foundation for human survival by supplying food and water while performing irreplaceable regulatory functions such as climate modulation and biodiversity maintenance [1,2,3,4,5,6]. Since the early 2000s, however, ecosystems worldwide have faced unprecedented threats from anthropogenic pressures and climate change, contributing to declines in roughly 60% of global ecosystem services [7,8,9,10,11,12,13]. This degradation directly undermines ecosystems’ capacity to support human well-being, underscoring the urgent need to understand ecosystem structure and function [14,15,16].

This recognition has positioned ecosystem health (EH) as a pivotal goal for environmental assessment and management [17]. EH represents an ecosystem’s capacity to maintain structural integrity, function normally, recover from disturbances, and sustainably deliver services to society [18,19,20]. Assessment methodologies have evolved from single biological indicators to multi-dimensional integrated models [21,22]. Early approaches relying on indicator species were constrained by high costs and limited scalability [23,24,25]. Subsequent frameworks, such as the Pressure–State–Response model, incorporated anthropogenic drivers through causal linkages, a process that, while valuable, often depended on expert judgment, thereby introducing a degree of subjectivity [26,27,28,29,30,31].

The Vitality–Organization–Resilience (VOR) model provides a classic operational framework for quantitative EH assessment [32,33,34,35,36]. Recognizing that the VOR framework did not fully capture an ecosystem’s contributions to human well-being, researchers later integrated ecosystem services (ES) as a distinct dimension, leading to the VORS model [37,38]. More recently, efforts to enhance regional relevance have incorporated metrics for key pressures such as hydrological conditions and land use changes [39,40,41,42,43].

The spatiotemporal patterns of EH are co-driven by complex interactions between natural factors and human activities [44]. While traditional driver analysis methods effectively reveal statistical associations, they often fall short of capturing the intricate nonlinear dynamics characteristic of these systems [45,46,47,48]. Advances in machine learning algorithms offer robust nonlinear fitting capabilities, and when coupled with SHAP interpretation, these methods reveal localized effect pathways and thresholds in complex ecological systems [49,50,51,52,53,54,55,56,57].

A critical gap persists in existing frameworks: they often fail to capture the dynamic interplay between intrinsic ecosystem structure and external anthropogenic pressures. A critical gap, therefore, is the construction of a comprehensive framework that simultaneously captures intrinsic ecosystem structure (VOR), quantifies external service value (ES), and reflects dominant regional environmental conditions. This gap is particularly salient in vulnerable regions like the Yellow River Basin, characterized by intense human–environment conflicts [58]. To address this gap, this study proposes a novel “Vitality–Organization–Resilience–Services–Quality” (VOR-SQ) assessment framework, which augments the classic VOR model by integrating ecosystem services and an enhanced Ecological Quality index. Leveraging the XGBoost machine learning model and SHAP interpretation, this study quantitatively delineates key drivers of EH in the Yellow River Basin and unravels their nonlinear mechanisms. By innovating both assessment framework and analytical methodology, this study seeks to advance understanding of ecological health evolution in coupled human–environment systems and furnish an empirical foundation for adaptive management strategies within and beyond the Yellow River Basin.

2. Study Area and Materials

2.1. Study Area

The study area, the Henan section of the Yellow River Basin (HYRB), lies between 33°07′–36°36′ N and 110°13′–116°05′ E. Stretching roughly 711 km along the main channel, the region covers approximately 66,000 km² (Figure 1). Administratively, it includes ten prefecture-level cities: Zhengzhou, Kaifeng, Luoyang, Anyang, Hebi, Xinxiang, Jiaozuo, Puyang, Sanmenxia, and Jiyuan. The region has a warm-temperate, semi-humid continental monsoon climate, where most precipitation falls in summer. Mean annual temperature ranges from 12 to 15 °C, and annual precipitation from 500 to 900 mm. The terrain exhibits a pronounced west-to-east gradient, descending from the mountainous west to the low-lying east. The western part comprises the eastern foothills of the Qinling Mountains, with elevations exceeding 1000 m, whereas the central and eastern parts transition into hilly landscapes and alluvial plains. Rapid urbanization and intensive anthropogenic activities have imposed severe ecological pressures on the region, including biodiversity loss, water pollution, and landscape fragmentation [58,59,60]. Concurrently, the region faces the complex challenge of balancing ecological conservation, food security, and economic development [61,62,63,64]. Given the national strategy prioritizing environmental protection and sustainable growth in the Yellow River Basin [65], developing a scientifically robust EH assessment system for the HYRB is essential. Such a system is crucial for diagnosing the root causes of ecological issues, identifying degradation hotspots, and informing decision-making for basin-wide conservation, restoration, and sustainable management aimed at maintaining ecosystem services and achieving human–environment harmony [66,67].

2.2. Datasets

This study integrated multiple data sources: (1) LULC data; (2) MODIS remote sensing products; (3) solar radiation data; (4) agricultural statistics, primarily grain yield; (5) meteorological variables (temperature, precipitation, potential evapotranspiration); (6) a digital elevation model (DEM); (7) soil property data; (8) socioeconomic and nighttime light (NTL) data; and (9) ancillary parameters and inputs for the InVEST model (3.17.1). Detailed specifications regarding the spatial and temporal resolution, data sources, and preprocessing steps for each dataset are summarized in

Table 1. Detailed Information on Basic Data.

Dataset	Description	Spatiotemporal Resolution	Data Type	Source/Reference
Land Use/Land Cover (LULC)	Land cover classification maps	30 m (Annual, 2000–2020)	Raster	https://www.resdc.cn/
MODIS Products	Includes NDVI, ET, and other biophysical parameters	500 m	Raster	https://earthengine.google.com/
Climate Data	Solar radiation, temperature, precipitation, potential evapotranspiration	1 km (Monthly, 2000–2020)	Raster	https://www.geodata.cn/
Digital Elevation Model (DEM)	Elevation, slope, and derived topographic indices	30 m	Raster	http://www.gscloud.cn
Soil Properties	Soil type, organic carbon, depth to bedrock, etc.	1 km	Raster	https://www.fao.org/
Socioeconomic and Nighttime Light	Population density, GDP density, nighttime light (NTL) intensity	1 km (Annual, 2000–2020)	Raster	https://www.resdc.cn/
Agricultural Statistics	Henan Statistical Yearbooks	Provincial/Prefectural level (Annual)	Statistical Table	https://tjj.henan.gov.cn/

. All raster data were reprojected to the WGS 1984 UTM 49N coordinate system and resampled to a 30 m resolution using bilinear interpolation for continuous data and the nearest neighbor method for categorical data. Prior to aggregation, all indicators were normalized to a [0, 1] interval using min-max normalization.

3. Methods

Figure 2 presents the research framework. Multi-source data were first preprocessed. Based on the VOR-SQ model developed in this study, ecosystem health in the Henan section of the Yellow River Basin was assessed by calculating dimension-specific indicators to generate its spatiotemporal distribution. Trends and driving mechanisms were then analyzed using Theil–Sen/Mann–Kendall and XGBoost-SHAP methods, respectively. To support ecological restoration efforts.

3.1. Ecosystem Health Assessment Indicators

This study develops the VOR-SQ model to evaluate ecosystem health (EH) in the HYRB. This framework extends the classical Vitality–Organization–Resilience (VOR) model [34,35,68] by integrating Ecosystem Services (ES) and an enhanced Ecological Quality (EQ) index [69]. The composite Ecosystem Health Index (EHI) is calculated using a multiplicative geometric mean model, structured as a two-tiered aggregation:

$$EHI=\sqrt[3]{ES\times EQ\times NEHI}$$

(1)

$$NEHI=\sqrt[3]{EV\times EO\times ER}$$

(2)

where $EHI$ is the composite Ecosystem Health Index; $NEHI$ is the Natural Ecosystem Health Index; $ES$ is the comprehensive Ecosystem Services index; $EQ$ is the Ecological Quality index; $EV$, $EO$, $ER$ represent Ecosystem Vitality, Organization, and Resilience, respectively. All indicator values were normalized to a [0, 1] range using the min-max normalization method prior to aggregation. Based on the computationally derived EHI value, its health grade may be classified according to

Table 2. Classification of EHI Values.

EHI Value	Ecological System Health Range	Health Level
0~0.2	Low	I
0.2~0.4	Lower	II
0.4~0.6	Medium	III
0.6~0.8	Higher	IV
0.8~1	High	V

.

3.1.1. Ecosystem Vitality (EV)

Ecosystem Vitality (EV), representing metabolic activity and primary productivity [34], was quantified by net primary productivity (NPP) estimated using the Carnegie–Ames–Stanford Approach (CASA) model [35]:

$$NPP(x,t)=APAR(x,t)\times LUE(x,t)$$

(3)

where $NPP\left(x,t\right)$ is the net primary productivity (gC∙m⁻²) at pixel $x$ and month $t$; $APAR\left(x,t\right)$ is the absorbed photosynthetically active radiation (MJ∙m⁻²) at pixel $x$ and month $t$. $LUE(x,t)$ is the actual light use efficiency (gC∙MJ⁻¹) at pixel $x$ and month $t$.

3.1.2. Ecosystem Organization (EO)

Ecosystem Organization (EO) represents the complexity of interactions within an ecosystem, shaped by its components and their material–energy exchanges [70]. EO is quantified using landscape metrics that capture spatial heterogeneity and landscape connectivity, which together reflect ecosystem structural properties [71,72]:

$$\begin{array}{cc}\hfill EO=0.35\times LH& + 0.35\times LC+0.3\times IC\hfill \\ & =0.25\times SHDI+0.1\times AWMPFD+0.2\times FN+0.05\times CONT\hfill \\ & + 0.1\times COHE+0.05\times {FN}_1+0.025\times {CONT}_1\hfill \\ & + 0.025\times {COHE}_1+0.05\times {FN}_2+0.025\times {CONT}_2\hfill \\ & + 0.025\times COH{E}_2+0.05\times {FN}_3+0.025\times {CONT}_3\hfill \\ & + 0.025\times {COHE}_3\hfill \end{array}$$

(4)

the $EO$ index was calculated as a weighted sum (0.35, 0.35, 0.3) of landscape heterogeneity ($LH$, via $SHDI$ and $AWMPFD$), overall connectivity ($LC$), and connectivity of key ecological lands ($IC$: woodland, grassland, water) [73,74,75,76,77]. $LC$ and $IC$ were derived from fragmentation, connectivity, and cohesion indices, with subscripts 1–3 denoting respective land cover types. All indices were computed using FRAGSTATS 4.2.

3.1.3. Ecosystem Resilience (ER)

Ecosystem Resilience (ER) is the capacity to maintain structure and function under disturbance, comprising resistance (withstanding pressure) and resilience (recovery ability) [78,79]. ER is quantified using the Resistance Coefficient (RTC) and Resilience Coefficient (RLC) based on land use attributes (

Table 3. Resistance and Resilience of Different Land Use Types.

Land Use/Cover Type	Cropland	Forest	Grassland	Water	Construction	Unutilized Land
Resistance Coefficient (RTC)	0.4	0.5	0.8	0.7	0.2	1
Resilience Coefficient (RLC)	0.5	1	0.7	0.8	0.3	0.2

). For areas characterized by intensive industry, high population density, or severe degradation, a higher weighting is assigned to resilience, effectively prioritizing the RLC over the RTC in the calculation to reflect recovery needs [80]. ER is calculated as:

$$ER={\omega }_1\times \sum _{i=1}^n{RTC}_i\times A_i+{\omega }_2\times \sum _{i=1}^n{RLC}_i\times A_i$$

(5)

where ER denotes ecosystem resilience; ${RTC}_i$ and ${RLC}_i$ represent the resistance coefficient and resilience coefficient for the $i$-th land use type, respectively; $A_i$ is the area proportion of that land use type; where ${\omega }_1$ and ${\omega }_2$ are weighting coefficients, with ${\omega }_1$ = 0.4, and ${\omega }_2$ = 0.6.

3.1.4. Comprehensive Ecosystem Services Index (ES)

The Ecosystem Services (ES) index quantifies the outputs that functioning ecosystems provide to human societies. Based on multi-source data, this study assessed five key ES types in the Yellow River Basin (HYRB): Carbon Storage (CS), Water Yield (WY), Habitat Quality (HQ), Soil Conservation (SC), and Food Production (FP). The specific assessment approaches for these services are detailed in

Table 4. The Quantification and Indicators of Ecosystem Services.

Type	Quantification Methods	Equation	Explanation
CS	The InVEST model (3.17.1)’s carbon storage module was employed to estimate stored carbon, based on computations of mean carbon densities from four distinct pools for each land use category [81].	$C_{total}=C_{above}+C_{below}+C_{soil}+C_{dead}$	Among these, $C_{total}$ denotes total carbon stock (t· hm−2; $C_{above}$ denotes aboveground biogenic carbon stock in vegetation (t·hm−2); $C_{below}$ denotes belowground biogenic carbon stock in vegetation (t·hm−2); $C_{soil}$ denotes organic carbon stock in soil (t·hm−2); $C_{dead}$ denotes carbon stock in organic matter within litter (t·hm−2).
WY	In InVEST (3.17.1), water yield is estimated as the residual of mean annual precipitation minus actual evapotranspiration [82].	$Y_i=\left(1-{\displaystyle \frac{{AET}_i}{P_i}}\right)\times P_i$	Where $Y_i$ is the annual runoff (mm) for grid $i$; ${AET}_i$ is the annual actual evaporation (mm) for grid $i$; $P$ is the annual precipitation (mm) for grid $i$.
HQ	The InVEST (3.17.1) habitat quality module calculates habitat quality by considering existing land use patterns and associated threats to biodiversity [83].	$Q_{ij}=H_j\left(1-{\displaystyle \frac{D_{ij}^z}{D_{ij}^z+K^z}}\right)$	Here, $Q_{ij}$ represents the habitat quality of grid cell $i$ within land use type $j$; $D_{ij}^z$ denotes the total threat level of grid cells within land use type i; $K$ and $Z$ are scaling factors; $H_j$ indicates the habitat suitability of the land use type.
SC	Using the InVEST model (3.17.1)’s modules to analyze the discrepancy between potential erosion losses and observed erosion processes as a measure of soil conservation [84].	${SC}_i={RKLS}_i-{USLE}_i$	Among these, ${SC}_i$, ${RKLS}_i$, and ${USLE}_i$ represent the soil conservation amount (t·hm−2) in region $i$, the potential soil erosion amount (t·hm−2) without vegetation cover and soil conservation measures, and the actual soil erosion amount (t·hm−2), respectively.
FP	The evaluation of existing research findings on FP is based on two indicators: yield and NPP [85].	${FP}_{ij}={\displaystyle \frac{{NPP}_{ij}}{{NPP}_i}}{FP}_i$	Where ${FP}_{ij}$ denotes the crop yield (t·hm−2) of the $j$-th spot within the $i$-h region; ${FP}_i$ represents the total yield (t·hm−2) of the $i$-th region; ${NPP}_{ij}$ indicates the NPP value of the $j$-th grid cell within the $i$-th region, while ${NPP}_i$ signifies the total NPP value of the $i$-th region.

.

$$ES=\sum _{i=1}^n{ES}_i$$

(6)

$${ES}_a={\displaystyle \frac{ES-{ES}_{min}}{{ES}_{max}-{ES}_{min}}}$$

(7)

among these, ${ES}_a$ denotes the standardized result for each ecosystem service; $ES$ represents the initial value, while ${ES}_{max}$ and ${ES}_{min}$ denote the maximum and minimum values, respectively. ${ES}_i$ indicates the standardized result for the $i$ service.

3.1.5. Ecosystem Environmental Quality (EQ)

EQ represents the biophysical state of an ecosystem, the formation and evolution of which are jointly shaped by both natural conditions and anthropogenic pressures. To enable a more precise characterization of ecological quality (EQ) within the human–environment coupled system of the Henan Yellow River Basin (HYRB), this study introduces targeted enhancements to the classic Remote Sensing Ecological Index (RSEI). The original RSEI integrates NDVI (greenness), LST (heat), NDBSI (dryness), and wetness (WET) to reflect regional ecological conditions [86]. However, in the densely populated and rapidly urbanizing HYRB—a core grain-producing area—ecological quality is critically influenced by root-zone soil moisture and direct anthropogenic disturbance, both inadequately captured by the standard RSEI. The WET component derived from surface reflectance is a poor proxy for root-zone soil moisture, while human activity is only indirectly inferred from land cover-based indices such as NDBSI. Therefore, we optimized the index by introducing a Soil Moisture Index (SMI) to monitor root-zone water conditions [87] and incorporating Nighttime Light (NTL) as a direct proxy for anthropogenic intensity. The enhanced EQ integrates five indicators via principal component analysis: SMI, NDVI, LST, NDBSI, and NTL. All computations were performed on the Google Earth Engine (GEE) platform.

$${EQ}_i=1-PC1\left[f\left(SMI,NDVI,LST,NDBSI,NTL\right)\right]$$

(8)

$$EQ={\displaystyle \frac{{EQ}_i-{EQ}_{min}}{{EQ}_{max}-{EQ}_{min}}}$$

(9)

where $SMI$, $NDVI$, $LST$, $NDBSI$, and $NTL$ represent the Soil Moisture Index, Normalized Difference Vegetation Index, Land Surface Temperature Index, and Normalized Difference Building and Soil Index, respectively. These indices correspond to soil moisture, greenness, heat, dryness, and human activity intensity.

3.2. Theil–Sen Median Trend Analysis and Mann–Kendall Test

To characterize the temporal dynamics of ecosystem health, we employed a combination of Theil–Sen median trend analysis and the Mann–Kendall (MK) significance test. The Theil–Sen estimator, a non-parametric method, was used to calculate the magnitude of the trend (β) for the Ecosystem Health Index (EHI) over the study period. This approach is robust against outliers and does not assume any specific data distribution [88]. A positive β value indicates an improving trend in ecosystem health, while a negative β signifies degradation.

The statistical significance of these observed trends was subsequently evaluated using the Mann–Kendall test, a rank-based non-parametric method that does not require the data to be normally distributed [89,90].

3.3. Extreme Gradient Boosting (XGBoost)

XGBoost is a high-performance Gradient Boosting Decision Tree (GBDT) variant designed for modeling complex nonlinear relationships and threshold effects [91,92].

$$OBJ={\sum }_{i=1}^{n}loss\left(y_i,{\widehat{y_i}}^{\left(t\right)}\right)+{\sum }_{k=1}^t\epsilon \left(f_k\right)$$

(10)

in this formulation, $loss\left(y_i,{\widehat{y_i}}^{\left(t\right)}\right)$ denotes the loss function, which measures the discrepancy between the predicted value and the true value. It calculates the difference between the true value $y_i$ of the $i$-th sample and the model’s prediction ${\widehat{y_i}}^{\left(t\right)}$ after the $t$-th iteration. $\epsilon \left(f_k\right)$ is the regularization term for the $k$-th tree, preventing the model from overfitting.

3.4. Shapley Additive Explanation Model (SHAP)

To address the “black-box” nature of XGBoost, the SHAP framework was integrated to evaluate driver contributions [93]. Grounded in cooperative game theory, SHAP quantifies the average marginal impact of each feature across all possible feature combinations [94,95]. This approach facilitates both global feature ranking and local response analysis [54]. The Shapley value ${\phi }_i$ for feature $i$ is defined as:

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

(11)

where ${\phi }_i$ represents the Shapley value of the $i$-th feature variable; N is the set of all features; $S$ is the subset obtained by removing the $i$-th feature from $N$; $f_x(S\cup \{i\left\}\right)$ and $f_x\left(S\right)$ correspond to the model results when the $i$-th feature is included and excluded, respectively. $A$ Shapley value greater than 0 indicates a positive contribution of the feature to EHI, while a value less than 0 signifies a negative contribution.

By synthesizing these techniques, this study adopts an integrated XGBoost-SHAP analytical framework. This coupled approach delivers high predictive accuracy for EHI while crucially permitting the characterization of both the strength and direction of each driver’s effect, as well as its functional relationship, thus advancing the analysis from mere correlation towards causal mechanism inference.

4. Results

4.1. Spatiotemporal Variations of EV, EO, ER, ES, and EQ

From 2000 to 2020, the five EHI sub-indices in the HYRB followed distinct trajectories (Figure 3). Mean Ecosystem Vitality (EV) traced a “V-shaped” path, dipping from 0.618 in 2000 to a trough of 0.606 in 2005 before climbing to 0.651 by 2020—a net improvement. Spatially, EV generally decreased from the southwest to the northeast, though increases were widespread across the basin. The landscape heterogeneity indices (SHDI and AWMPFD) tell a story of disturbance and partial recovery. Both plummeted to their lowest points around 2005, then slowly increased through 2020, yet still ended the period lower than they started. Connectivity metrics followed a similar pattern: FN and CONT bottomed out in 2005 and saw a sluggish recovery after 2010. The cohesion index (COHE), however, told a different tale, declining persistently throughout the two decades. This contrast suggests that while some connectivity measures showed localized recovery, the overall landscape cohesion continued to unravel under high-intensity disturbance. Ecosystem Organization (EO) was highest in the southwestern foothills and Taihang Mountains, with moderately high values in the middle reaches. It rose until 2005, then entered a phase of fluctuating decline. Ecosystem Resilience (ER) showed a clear spatial divide: peak values in the forest- and grassland-dominated east, and much lower values in the intensively farmed and urbanized central and western reaches. Over the study period, ER steadily declined across 81.6% of the basin. Ecosystem Services (ES) broadly mirrored EV’s spatial pattern. Its mean value dropped to a low of roughly 0.352 around 2010, after which it began a steady recovery—a positive trend observed across 63.5% of the basin. Finally, Ecological Quality (EQ) displayed strong spatial heterogeneity, with the lowest values in urban areas and a general decrease from southwest to east. Temporally, EQ fluctuated but trended downward overall, with annual means between 0.365 and 0.467. The notable peak in 2010 (0.467) likely reflects the intersecting impacts of post-2005 economic growth and the implementation of national restoration projects, highlighting the complex interplay between human activity and ecological policy.

4.2. Spatiotemporal Evolution of EH in the HYRB Region

Over the 2000–2020 period, the mean Ecosystem Health Index (EHI) fluctuated within a narrow range (0.40–0.43) before settling at a lower level (Figure 4). Ecosystem health displayed marked spatial heterogeneity throughout the basin. The highest levels of ecosystem health were predominantly clustered in the basin’s southwestern and northwestern sectors. EHI values generally decreased eastward from the southern foothills of the Funiu, Xiong’er, and Taihang Mountains. This spatial structure creates a distinct west-high-east-low gradient across the HYRB. This gradient is fundamentally driven by topography: the western mountainous areas are dominated by natural and semi-natural ecosystems with stable structures and strong service functions, whereas the eastern plains are dominated by human-managed landscapes experiencing intensive anthropogenic disturbance. The observed temporal fluctuations in EHI indicate that the region is undergoing a critical phase reflecting the interplay between trade-offs and synergies linking ecological conservation with socioeconomic development.

Based on the Theil–Sen median trend (β) and Mann–Kendall significance test (Z) results (Figure 5), EHI trends were classified into four categories: significantly improved (β > 0, |Z| ≥ 1.96), non-significantly improved (β > 0, |Z| < 1.96), non-significantly deteriorated (β < 0, |Z| < 1.96), and significantly deteriorated (β < 0, |Z| ≥ 1.96). Over the study period, 47.36% of the HYRB area exhibited an improving EHI trend, comprising 46.17% with non-significant improvement and 1.19% with significant improvement. Areas with significant improvement were relatively more concentrated in the western Funiu–Xiong’er Mountains and the central hilly regions, appearing as isolated patches interspersed across broader forested areas. This spatial configuration is likely linked to focused restoration initiatives such as the Natural Forest Conservation and the Grain-for-Green Program. Non-significantly improved areas were widely distributed across the western and central regions, encompassing substantial portions of farmland and built-up land. This widespread but modest increase may reflect the broad yet possibly diffused effects of regional ecological management measures. EHI deterioration affected 52.64% of the area, though this decline was predominantly non-significant (51.87%). Areas of significant deterioration were limited (0.76%) and largely confined to urban fringes and intensive farmlands. Significantly deteriorated areas were predominantly located at the urban fringe or within intensive farmland, a pattern that suggests a link to ecosystem degradation associated with rapid urban expansion and agricultural intensification. The prevalence of non-significant deterioration points to a widespread but gradual decline in ecological health across the basin. This chronic stress may be attributed to diffuse pressures such as non-point source agricultural pollution, groundwater over-extraction, and climate change. Analysis of EHI class transitions revealed a path-dependent dynamic: ecosystems in lower health tiers showed a tendency toward further decline, while higher-tier systems demonstrated greater stability. Nonetheless, localized decline risks persisted. Notably, the proportion of areas transitioning from ‘Medium’ to ‘Lower’ EHI classes increased between 2010 and 2015, coinciding with agricultural expansion and drought events during that period.

4.3. Analysis of Driving Factors

We employed an integrated XGBoost-SHAP framework to decouple the nonlinear influences on the Ecosystem Health Index (EHI). Hyperparameter tuning via grid search ensured the XGBoost model’s robust performance across all 2000–2020 test datasets. Model performance was strong, with R² values ranging from 0.809 to 0.852 and RMSE between 0.046 and 0.062, confirming predictive stability (Figure 6). This model effectively captures the complex, nonlinear associations between EHI and its potential drivers, providing a reliable foundation for the subsequent SHAP-based attribution analysis.

SHAP decomposition was integrated with the XGBoost model to disentangle the directional and magnitude effects of various drivers on EHI across both specific time periods and the multi-year average (Figure 7). The multi-year average identified Evapotranspiration (ET, X6), Elevation (DEM, X1), Slope (X11), and NDVI (X4) as the primary determinants of EHI. ET as the most influential factor, with its SHAP values widely distributed but skewed negative. The prevalence of negative SHAP values for ET underscores that water availability remains the primary limiting factor for EH, persisting even in areas adjacent to the Yellow River. NDVI showed a complex, nonlinear relationship with EHI, with data points falling on both sides of the SHAP value axis, indicating strong context dependence. In contrast, Elevation and Slope demonstrated clear, positive monotonic associations; higher values (represented in red on the summary plot) consistently aligned with positive SHAP values. This trend likely reflects the designation of steeper, high-elevation areas as ecological conservation zones, where restricted human activity facilitates natural recovery. Socioeconomic factors, such as Population Density (X7) and GDP (X3), had weaker global influence, though their relative contribution rates increased modestly by 2020, indicating phase-specific and potentially unstable effects.

To investigate the nonlinear dynamics, SHAP dependence plots were constructed for the top six drivers: ET (X6), Elevation (X1), Slope (X11), NDVI (X4), Population Density (X7), and Land Surface Temperature (LST, X2) (Figure 8). The plots revealed clear nonlinearities and potential ecological thresholds. Topographic Factors: Elevation (X1) exhibited a threshold at approximately 457 m. Below this threshold, the association was slightly negative, while above it, the effect turned positive, suggesting medium-high elevations are more favorable to ecosystem health. Slope (X11) exhibited an inflection point ~10.5°, beyond which steeper slopes were correlated with improved EHI, likely due to reduced anthropogenic disturbance. Climatic and Biophysical Factors: Evapotranspiration (ET, X6) displayed a critical threshold near 1153 mm. Lower ET signals water stress, but values exceeding the threshold also yield negative SHAP values—suggesting that excessive atmospheric demand acts as an additional ecological stressor in the HYRB. NDVI (X4) transitioned to a consistently positive association beyond a threshold of 0.76, underscoring the foundational role of adequate vegetation cover. Socioeconomic Factors exhibited typical nonlinear patterns. Population Density (X7) exerted a strong negative pressure at lower densities, which plateaued to a stabilized, weaker inhibitory effect at higher concentrations. Road Density (X10) showed a three-stage pattern “promotion-inhibition-rebalancing”—across two critical thresholds, pointing to a complex, phased influence.

Overall, each driver’s influence shifted qualitatively around specific thresholds, revealing differential sensitivity and nonlinear response mechanisms in ecosystem health. These critical points provide a reference for implementing zoned ecological management.

Integrating time-series and multi-year mean analyses reveals a dynamic nonlinear evolutionary pattern defined by stable environmental constraints, coupled water–heat dynamics, and localized human disturbances (Figure 9). SHAP contribution and interaction matrices indicate that evapotranspiration (X6)—the globally dominant factor—forms a strong interactive pathway with vegetation indices (X4), defining how water–heat balance regulates ecosystem health nonlinearly. Elevation (X1) and slope (X11), as core regulators, exert their influence through interactions with meteorological variables. In 2010, for instance, terrain factors briefly outweighed climatic drivers in relative contribution, reflecting a spatial redistribution under environmental constraints. Socioeconomic factors (X3, X5, X7, X9, X10) show limited global influence but pronounced trade-offs at local scales, highlighting how human disturbance can modulate the marginal effects of natural factors. Overall, driver interactions have remained structurally stable over the past two decades, exhibiting a clear path dependence in their influence on ecosystem health.

5. Discussion

5.1. Spatial Heterogeneity and Drivers of Ecosystem Health

The HYRB exhibits a pronounced “West-High, East-Low” gradient in ecosystem health (EH), reflecting a fundamental coupling between environmental gradients and land use intensity. Viewed through the lens of social-ecological resilience, the forested western mountains—bolstered by superior VOR levels—occupy the “Conservation” phase of the adaptive cycle, where high biophysical memory and connectivity are maintained. The fragmented eastern plains, strained by rapid urbanization, appear to be transitioning toward a “Release” phase. This spatial divergence parallels findings in the Danube Basin, confirming that forested headwaters harbor greater functional redundancy than their high-intensity agricultural counterparts [96]. The fluctuating EH decline observed throughout the study period highlights the vulnerability of regional ecosystems under dual pressures, suggesting that the HYRB may be nearing a regime shift. As ecological buffers erode, the system’s capacity to absorb shocks without structural collapse inevitably diminishes. Within the framework of Ecosystem Health theory, this degradation marks a decoupling of “Vitality” from “Organization”—a process where human-dominated landscapes prioritize immediate provisioning services at the cost of long-term regulatory resilience [97]. Accordingly, management strategies should pivot from static preservation toward fostering adaptive capacity [98], thereby preventing the crossing of irreversible ecological thresholds.

5.2. Interaction with Ecological Theoretical Frameworks

The VOR-SQ framework developed in this research aligns with the “Structure-Process-Function” paradigm central to landscape ecology [99]. Using the XGBoost-SHAP model, we identified non-linear responses that offer empirical evidence of ecosystem regime shifts [97,100]. Specifically, the contribution thresholds identified—457 m for altitude, 10.5° for slope, and 0.76 for NDVI—serve as quantitative benchmarks for precision management. For instance, vegetation restoration yields the greatest ecological gains where NDVI falls below 0.76, whereas areas above 457 m warrant prioritization within ecological redline protections. Evapotranspiration (ET) exhibited threshold behavior around 1153 mm, underscoring how acutely these ecosystems—situated in a semi-arid to semi-humid transition zone—respond to water availability. This finding aligns with Liebig’s Law of the Minimum [101,102].

5.3. Limitations and Future Perspectives

Despite the comprehensive insights provided, the inherent constraints of this study must be acknowledged. The current evaluation model focuses on long-term evolutionary trends, which limits its capacity to capture the immediate impacts of extreme events, such as heatwaves or flash droughts [103,104]. Given that such disturbances often catalyze regime shifts, the coupling of high-frequency climatic variables is essential for enhancing predictive responsiveness. Regarding the methodology, the SHAP-based interpretability offers an internal logic of the model rather than a definitive physical truth; its efficacy is inevitably constrained by data quality and overlooked factors like anthropogenic water management [105]. Furthermore, the ER indicator is also limited by the subjectivity of its construction. Currently, land use types are simplified into homogeneous blocks, a process that fails to capture the nuanced physiological variations within those classes. Future work must bridge this gap by transitioning to dynamic monitoring. Leveraging vegetation-specific indicators and long-term data sequences will allow for a more objective evaluation of recovery rates and tipping points, ultimately refining our understanding of resilience at every scale.

6. Conclusions

This study employed an integrated VOR-SQ framework, coupled with the XGBoost-SHAP model, to evaluate ecosystem health (EH) across the Henan section of the Yellow River Basin (HYRB). Spatiotemporal analysis from 2000 to 2020 reveals a consistent “high-west, low-east” gradient, alongside a subtle but steady decline in the composite EH index—leaving over 52% of the basin area at risk of degradation. Threshold diagnostics pinpointed critical tipping points, specifically an NDVI of 0.76 and evapotranspiration of 1153 mm, revealing that while natural factors establish the EH baseline, localized health transitions are primarily driven by intensive anthropogenic disturbances at urban fringes and agricultural zones. These findings necessitate a differentiated management strategy, one that is explicitly tailored to the region’s nonlinear dynamics. Strict conservation measures should be prioritized within the western ecological barriers, while management in central agricultural zones should focus on strengthening landscape connectivity. For the eastern plains, enforcing stringent development “red lines” is essential to maintain anthropogenic pressures—particularly population and road density—below the identified tipping points.

The study’s insights, however, should be viewed in light of several limitations. The current reliance on multi-year static snapshots may fail to capture the cumulative or lagged effects of extreme climate events on ecosystem resilience. Furthermore, indicators for aquatic and groundwater health require refinement to better encapsulate the basin’s hydrological complexity. Future research should focus on synthesizing multi-source real-time sensing data with long-term field observations. Shifting from static assessments toward dynamic, real-time early-warning systems will be vital for supporting adaptive management and long-term sustainability in high-intensity socio-ecological systems like the Yellow River Basin.