Revealing Nonlinear Relationships and Thresholds of Human Activities and Climate Change on Ecosystem Services in Anhui Province Based on the XGBoost–SHAP Model

Abstract

Under the combined influence of global climate change and intensified human activities, ecosystem services (ESs) are undergoing substantial transformations. Identifying their nonlinear driving mechanisms is crucial for promoting regional sustainable development. Taking Anhui Province as a case study, this research evaluates the spatial patterns and temporal dynamics of six key ecosystem services from 2000 to 2020—namely, biodiversity maintenance (BM), carbon fixation (CF), crop production (CP), net primary productivity (NPP), soil retention (SR), and water yield (WY). The InVEST and CASA models were employed to quantify service values, and the XGBoost–SHAP framework was used to reveal the nonlinear response paths and threshold effects of dominant drivers. Results show a distinct “high in the south, low in the north” spatial gradient of ES across Anhui. Regulatory services such as BM, NPP, and WY are concentrated in the southern mountainous areas (high-value zones > 0.7), while CP is prominent in the northern and central agricultural zones (>0.8), indicating a clear spatial complementarity of service types. Over the two-decade period, areas with significant increases in NPP and CP accounted for 50% and 64%, respectively, suggesting notable achievements in ecological restoration and agricultural modernization. CF remained stable across 98.3% of the region, while SR and WY exhibited strong sensitivity to topography and precipitation. Temporal trend analysis indicated that NPP rose from 395.83 in 2000 to 537.59 in 2020; SR increased from 150.02 to 243.28; and CP rose from 203.18 to 283.78, reflecting an overall enhancement in ecosystem productivity and regulatory functions. Driver analysis identified precipitation (PRE) as the most influential factor for most services, while elevation (DEM) was particularly important for CF and NPP. Temperature (TEM) and potential evapotranspiration (PET) affected biomass formation and hydrothermal balance. SHAP analysis revealed key threshold effects, such as the peak positive contribution of PRE to NPP occurring near 1247 mm, and the optimal temperature for BM at approximately 15.5 °C. The human footprint index (HFI) exerted negative impacts on both BM and NPP, highlighting the suppressive effect of intensive anthropogenic disturbances on ecosystem functioning. Anhui’s ES exhibit a trend of multifunctional synergy, governed by the nonlinear coupling of climatic, hydrological, topographic, and anthropogenic drivers. This study provides both a modeling toolkit and quantitative evidence to support ecosystem restoration and service optimization in similar transitional regions.

1. Introduction

Ecosystem services (ESs) refer to the various direct and indirect benefits that natural ecosystems provide to human society through biodiversity and ecological processes, including food provision, water regulation, climate regulation, soil retention, habitat support, and cultural values [1,2]. With the intensification of global climate change, land-use transformation, and rapid population growth, ecosystem services are facing unprecedented pressures and challenges [3]. International frameworks such as the United Nations 2030 Agenda for Sustainable Development and the Convention on Biological Diversity have highlighted the importance of ESs by integrating them into global goals for ecological conservation and sustainable development [4,5,6]. Recent assessments by the Intergovernmental Science-Policy Platform on Biodiversity and Ecosystem Services (IPBES), including the 2023 scoping for the second global assessment and the 2025 transformative change assessment, emphasize the need for integrated approaches to address biodiversity loss and enhance ecosystems. In China, the world’s largest developing country, rapid economic and social development has led to substantial ecological changes [7,8]. As a key component of the middle and lower reaches of the Yangtze River, Anhui Province is located in a transition zone characterized by fast-paced economic expansion and ecological vulnerability in eastern China [9]. The province’s development trajectory, land-use changes, and increasing human activity have exerted profound impacts on the spatial patterns and temporal dynamics of its ecosystem services [10]. Understanding the spatiotemporal evolution and driving mechanisms of ES in Anhui is thus of great significance for safeguarding regional ecological security, promoting coordinated urban–rural development, and advancing the national strategy of ecological civilization [11].

Significant progress has been made in the study of ecosystem services (ESs) across theoretical frameworks, monitoring techniques, and policy applications. Foundational initiatives such as the Millennium Ecosystem Assessment have established a classification and evaluation basis for ES [12], followed by the development of integrated models such as InVEST, CASA, and ARIES, which support multidimensional assessment of ecosystem functions [13]. Quantitative spatial analysis based on remote sensing and geographic information systems (GISs) has greatly improved the accuracy and timeliness of dynamic ES monitoring [14]. Recent literature on ecosystem services and resilience has further emphasized the role of biodiversity in enhancing ecosystem system stability and adaptive capacity, as seen in research on forest resilience principles and multifunctional synergies. Moreover, the integration of machine learning and big data techniques has overcome the limitations of traditional linear or single-variable models, enabling deeper insights into the complex nonlinear mechanisms driving changes in ES [15]. Among numerous machine learning methods, the XGBoost–SHAP framework stands out for its exceptional handling of high-dimensional data, nonlinear dependencies, and variable interactions while achieving superior prediction accuracy and efficiency compared to models such as random forests (which can be less scalable) or neural networks (which often lack interpretability). SHapley Additive exPlanations (SHAP) enhances XGBoost by providing model-independent interpretations of feature contributions, enabling the quantification of marginal effects, threshold responses, and interaction dynamics—features that are particularly superior to traditional methods such as linear regression, and even to other interpretable models such as LIME, which may not scale well to ensemble predictions. This makes XGBoost–SHAP an ideal choice for this study, as it effectively reveals the multifaceted, nonlinear drivers of ecosystem services (ESs) in complex transitional regions, where traditional models (such as linear regression) struggle to capture threshold effects and multifactorial synergies. Existing studies often focus on limited ES types and static evaluations, overlooking cross-service interactions and dynamic analyses, while driving mechanism research insufficiently addresses nonlinear influences and thresholds amid Anhui’s complex socio-economic heterogeneity [11,16]. There is an urgent need to integrate multiple models and advanced machine learning algorithms into a unified research framework to accurately assess the dynamic changes and key drivers of ES. This is essential for providing scientific support for regional ecological governance and sustainable development strategies [17].

Building on the above context, this study integrates the InVEST and CASA models to comprehensively assess key ecosystem service (ES) indicators in Anhui Province, including biodiversity maintenance, soil retention, water yield, carbon fixation, and net primary productivity (NPP). The objective is to characterize the spatiotemporal dynamics of these services over the past two decades. To identify the dominant climatic, land-use, and socio-economic drivers and their nonlinear response relationships, we employ the XGBoost machine learning algorithm in combination with SHAP (SHapley Additive exPlanations) values. This modeling approach addresses the limitations of traditional statistical methods in capturing complex causal relationships [18] and enables a more precise quantification of the contribution paths and interactions of multiple driving forces. By systematically integrating multi-model frameworks, diverse datasets, and a multi-variable driver analysis, this study aims to reveal the underlying mechanisms of spatiotemporal ES evolution in Anhui. Furthermore, it proposes targeted strategies for optimizing and managing different ecosystem service functional zones, thereby providing theoretical and technical support for regional ecological civilization and green development.

2. Materials and Methods

2.1. Study Area Overview

Anhui Province (114°54′–119°37′ E, 29°41′–34°38′ N) is located in the inland core of eastern China, a region characterized by high population and economic density (Figure 1). Covering a total area of approximately 140,100 km², the province governs 16 prefecture-level cities. As the only province in the Yangtze River Delta urban agglomeration that spans both the Yangtze and Huai River basins, Anhui exhibits pronounced north–south transitional geographic and ecological characteristics. From north to south, the topography transitions through the Huaibei Plain (20–50 m elevation), the Jianghuai Hills (50–200 m), the Dabie Mountains (up to 1774 m), the Yangtze River Plain (10–30 m), and the southern Anhui Mountains, where Lotus Peak reaches 1864 m. This gradient forms a climatic and vegetative transition zone from the warm temperate to the mid-subtropical zone, blending the ecological traits of both northern deciduous broadleaf forests and southern evergreen broadleaf forests. Studies have shown that this ecotone harbors high biodiversity and ecological sensitivity, representing a vital component of China’s biogeographic framework [19].

Over the past two decades, however, widespread and unregulated human activities—especially urban expansion, infrastructure development, and intensified land use—have led to severe ecological degradation across many parts of Anhui. Problems such as habitat fragmentation, forest degradation, and water pollution have become increasingly prominent, resulting in a marked decline in ecosystem service functionality [20]. In ecologically sensitive areas such as the Huangshan and Dabie Mountains, tourism development has placed additional pressure on native forest cover and species habitat integrity [21]. Meanwhile, rapid industrial development along the Yangtze and Huai Rivers has outpaced the establishment of effective ecological compensation and pollution control mechanisms, exacerbating regional water-related environmental risks [22]. Therefore, ecological governance in Anhui must be tailored to its transitional geographic context, with a focus on integrated protection strategies across mountain–plain–river networks to enhance the resilience and stability of regional ecosystems.

2.2. Data Sources and Processing

This study integrates a variety of datasets encompassing topographic, climatic, soil, remote sensing, and socio-economic dimensions to support the spatiotemporal analysis of ecosystem services (ES) and their driving mechanisms in Anhui Province. Land use data from 2000 to 2020 are derived from the Wuhan University CLCD dataset (https://zenodo.org/record/5210928, accessed on 25 April 2024), with a spatial resolution of 30 m. NDVI (Normalized Difference Vegetation Index), data from the National Tibetan Plateau Data Center, and DEM data (90 m) from the Geospatial Data Cloud supported slope calculations and hydrological modeling. Climate data, including monthly precipitation and temperature, were sourced from the China Meteorological Data Service and Anhui’s regional climate records. Meteorological observation data mainly consist of radiation, precipitation, and temperature, sourced from the “China Regional High Spatiotemporal Resolution Surface Meteorological Element Driving Dataset” (http://westdc.westgis.ac.cn/, accessed on 25 April 2024). Soil data comes from HWSD, whose dataset integrates soil survey data from more than 180 countries around the world (https://gaez.fao.org/pages/hwsd, accessed on 25 April 2024). Socio-economic data (e.g., grain output) came from official provincial statistics, and administrative boundaries were provided by the Chinese Academy of Sciences. The Human Footprint (HFI) is a comprehensive indicator that combines factors such as land cover, population density, infrastructure, and accessibility to quantify human pressure on the environment. It is widely used to identify habitat loss, deforestation and degradation risks, species extinction risks, and to assist in spatial planning, land use management, and the identification of critical intact habitats (https://doi.org/10.6084/m9.figshare.16571064, accessed on 25 April 2024).

All datasets were standardized through atmospheric and radiometric correction, resampled to 100 m resolution, and reprojected (Albers or WGS_1984_UTM_Zone_50N) to ensure spatial consistency. These procedures ensured high-quality integration of diverse data types for use in the InVEST and CASA models and the XGBoost–SHAP framework, providing a robust foundation for subsequent ES evaluation and driver analysis [11,23] (see

Table 1. Data sources and their meanings.

Data Type	Source/Platform	Spatial Resolution/Precision
Land use data	https://zenodo.org/record/5210928#.YcZ_nWBByUk (accessed on 25 April 2024)	30 m
NDVI index	The National Tibetan Plateau Data Center (https://data.tpdc.ac.cn, accessed on 25 April 2024)	250 m
DEM data	Geospatial Data Cloud (http://www.gscloud.cn/, accessed on 25 April 2024)	90 m
Meteorological data	China Meteorological Data Center (http://data.cma.cn/, accessed on 25 April 2024)	Interpolated to 100 m
Soil data	China Soil Database [http://data.casnw.net/portal/, accessed on 25 April 2024]	100 m
Human footprint index	https://doi.org/10.6084/m9.figshare.16571064, accessed on 25 April 2024	500 m
Administrative boundaries, roads, and other basic geographic data	RESDC, Chinese Academy of Sciences [https://www.resdc.cn/, accessed on 25 April 2024]	-

).

2.3. Ecosystem Service Assessment

To provide a clear overview of the study’s workflow, this research integrates multiple steps: (1) data collection and preprocessing from diverse sources (e.g., remote sensing, meteorological, and socio-economic datasets) to ensure spatial consistency; (2) quantification of six ecosystem services (ESs) using InVEST, CASA, and RUSLE models; (3) spatiotemporal analysis of ES patterns and trends via Sen’s slope estimator and Mann–Kendall test; and (4) identification of nonlinear drivers and thresholds through the XGBoost–SHAP framework, including variable selection, model training, validation, and interpretation. This sequential approach facilitates a comprehensive understanding of ES dynamics in Anhui Province.

Due to its complex and fragile geological conditions, Anhui Province requires careful evaluation of key ecosystem services (ESs) that contribute to ecological balance and stability. In this study, six critical ES indicators were selected for assessment: food production (FP), water yield (WY), soil conservation (SC), biodiversity maintenance (HQ), carbon sequestration (CF), and net primary productivity (NPP) [24]. These services were quantified using a combination of the InVEST model, the CASA model, and the Revised Universal Soil Loss Equation (RUSLE) [25].

2.3.1. Carbon Storage, CS

The carbon sequestration function was assessed using the InVEST model, which adopts a multi-pool carbon accounting approach. The ecosystem carbon stock is partitioned into four interacting carbon pools: aboveground biomass carbon (e.g., tree trunks and leaves), belowground biomass carbon (e.g., root systems), soil organic carbon (e.g., humus layer), and dead organic matter (e.g., litterfall and decomposing plant material) [26]. By quantifying the spatiotemporal dynamics of each carbon pool, the model captures vegetation succession processes and carbon flux balances. The total carbon storage is calculated as follows:

$${\mathit{C}}_{\mathit{total} }={\mathit{C}}_{\mathit{above} }+{\mathit{C}}_{\mathit{below} }+{\mathit{C}}_{\mathit{soil} }+{\mathit{C}}_{\mathit{dead}}$$

(1)

Among these, ${\mathrm{C}}_{\mathit{total} }$ represents the total carbon stock in the study area, ${\mathrm{C}}_{\mathit{above} }$ denotes the aboveground biomass carbon stock, ${\mathrm{C}}_{\mathit{below} }$ refers to the belowground biomass carbon stock, ${\mathrm{C}}_{\mathit{soil} }$ represents the soil carbon stock, and ${\mathrm{C}}_{\mathit{dead} }$ represents the dead organic carbon stock.The carbon density for different land use types and soil types is primarily derived from relevant literature on the study area [17].

2.3.2. Water Yield, WY

This study uses the water yield module of the InVEST model to simulate the water cycle processes at the land–atmosphere interface. The hydrological modeling framework captures the mechanism of precipitation redistribution by incorporating geospatial heterogeneity to quantify water supply potential at the grid-cell scale [27]. The core of this modeling system is a dynamic “input–storage–output” water balance, which integrates topographic factors (e.g., slope, flow direction), meteorological inputs (e.g., precipitation intensity, evapotranspiration demand), and vegetation characteristics (e.g., canopy interception, root uptake). Key biophysical parameters include atmospheric moisture input, soil water-holding capacity, and vegetation transpiration efficiency. These parameters were localized and calibrated to ensure model applicability in the loess plateau and mountainous regions of Anhui. The results support the assessment of how forest structure influences the spatiotemporal allocation of water resources, offering scientific guidance for watershed-scale water resource coordination and management [28]. The water yield for each grid cell is computed as follows:

$$\mathrm{Y}\left(\mathrm{x}\right)=\left(1-{\displaystyle \frac{\mathrm{A}\mathrm{E}\mathrm{T}\left(\mathrm{x}\right)}{\mathrm{P}\left(\mathrm{x}\right)}}\right)\times \mathrm{P}\left(\mathrm{x}\right)$$

(2)

Among these, $\mathrm{Y}\left(\mathrm{x}\right)$ represents the annual water yield (mm) for each grid unit $\mathrm{x}$ within the watershed, $\mathrm{A}\mathrm{E}\mathrm{T}\left(\mathrm{x}\right)$ denotes the annual actual evapotranspiration (mm) for grid unit $\mathrm{x}$, and $\mathrm{P}\left(\mathrm{x}\right)$ represents the annual precipitation (mm) for grid unit $\mathrm{x}$ [29].

2.3.3. Soil Retention, SR

To assess soil retention services, this study employs a process-driven erosion control modeling framework grounded in the theory of surface material transport. The model incorporates spatial heterogeneity by integrating multiple parameters, including terrain variability (slope and aspect), the spatial distribution of rainfall kinetic energy (erosivity), vegetation coverage heterogeneity (leaf area index), and soil resistance to erosion (e.g., organic matter content). This coupled system is designed to analyze how improvements in ecosystem structure reduce soil erosion rates across diverse landscapes. The soil retention module of the InVEST model is based on the Universal Soil Loss Equation (USLE) and is used here to evaluate the soil conservation capacity of forest ecosystems in the loess regions of western Anhui. The model requires input data on elevation, vegetation, precipitation, and soil properties. The soil retention (SR) is calculated as the difference between potential and actual soil loss:

$$\mathrm{S}\mathrm{R}=\mathrm{A}\mathrm{p}-\mathrm{A}\mathrm{r}=\mathrm{R}\times \mathrm{K}\times \mathrm{L}\times \mathrm{S}\times \left(1-\mathrm{C}\times \mathrm{P}\right)$$

(3)

Among these, $\mathrm{S}\mathrm{R}$ is the soil retention amount (t·ha⁻¹·yr⁻¹), $\mathrm{A}\mathrm{p}$ is potential soil loss without vegetation or conservation measures, $\mathrm{A}\mathrm{r}$ is actual soil loss under current land use and vegetation cover, $\mathrm{R}$ is the rainfall erosivity factor (MJ·mm·ha⁻¹·hr⁻¹·yr⁻¹), $\mathrm{K}$ is the soil erodibility factor (t·ha·hr·ha⁻¹·MJ⁻¹·mm⁻¹), $\mathrm{L}$ and $\mathrm{S}$ are the slope length and slope steepness factors (often combined as LS), $\mathrm{C}$ is the cover-management factor (ranging from 0 to 1), and $\mathrm{P}$ is the support practice factor (ranging from 0 to 1).

The values for $\mathrm{C}$ and P were obtained from previous studies specific to the region (

Table 2. p value and C value of different land use types.

Type	Agriculture	Forests	Grass/Shrub	Water	Constructed	Bared
P	0.29	0.7	0.5	0.2	0.16	0.27
C	0.27	0.01	0.06	0	0.2	0.35

). This modeling approach provides spatially explicit estimates of the ecosystem’s capacity to reduce soil erosion, serving as a basis for land management and conservation planning [30].

2.3.4. Biodiversity Maintenance, BM

Referring to the guidelines for delineating ecological conservation redlines and relevant studies, we quantitatively assess the biodiversity maintenance service capacity index in the study area using the Net Primary Productivity (NPP) method. The calculation formula is as follows:

$${\mathrm{S}}_{\mathrm{bio} }=\mathrm{N}\mathrm{P}\mathrm{P}\times {\mathrm{F}}_{\mathrm{pre} }\times {\mathrm{F}}_{\mathrm{tem} }\times \left(1-{\mathrm{F}}_{\mathrm{alt} }\right)$$

(4)

Among these, ${\mathrm{S}}_{\mathrm{bio}}$ represents the index of biodiversity maintenance service capacity, NPP stands for annual net primary productivity of vegetation, ${\mathrm{F}}_{\mathrm{pre}}$ denotes annual precipitation, ${\mathrm{F}}_{\mathrm{tem}}$ corresponds to annual temperature, and ${\mathrm{F}}_{\mathrm{alt}}$ represents the altitude factor [31].

2.3.5. Crop Production, CP

According to relevant studies, a significant linear relationship exists between crops and NDVI (Normalized Difference Vegetation Index). By utilizing NDVI data, this research approximately allocates the grain yield in the study area to croplands. The calculation formula is as follows:

$${\mathrm{G}}_{\mathrm{i}}={\displaystyle \frac{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{i}}}{{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}\mathrm{u}\mathrm{m}}}}\times {\mathrm{G}}_{\mathrm{s}\mathrm{u}\mathrm{m}}$$

(5)

Among these, ${\mathrm{G}}_{\mathrm{i}}$ represents the grain yield allocated to the ith cropland grid; ${\mathrm{G}}_{\mathrm{s}\mathrm{u}\mathrm{m}}$ denotes the total grain production across different cities in the study area; ${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{i}}$ refers to the NDVI value of the ith cropland grid; and ${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{s}\mathrm{u}\mathrm{m}}$ represents the total NDVI value of croplands in various cities within the study area [32].

2.3.6. Net Primary Productivity, NPP

Net primary productivity (NPP) reflects the interaction between vegetation and external environmental factors. As a key indicator of vegetation vitality, NPP also serves as a proxy for terrestrial ecosystem quality and represents a fundamental component of the terrestrial carbon cycle [31]. Given the close link between vegetation carbon budgets and ecosystem productivity, quantifying NPP is critical for assessing an ecosystem’s carbon sequestration capacity and overall ecological health. In this study, NPP was estimated using the Carnegie–Ames–Stanford Approach (CASA) model and subsequently normalized to a scale of 0–1 to facilitate comparative analysis. The core NPP computation is based on light-use efficiency theory and is defined as follows:

$$\mathrm{N}\mathrm{P}\mathrm{P}\left(\mathrm{x},\mathrm{y},\mathrm{t}\right)=\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}\left(\mathrm{x},\mathrm{y},\mathrm{t}\right)\mathrm{\ast }\mathsf{\epsilon }(\mathrm{x},\mathrm{y},\mathrm{t})$$

(6)

Among these, APAR(x,y,t) represents the photosynthetically active radiation (MJ·m²·month⁻¹) absorbed by the pixel at spatial position (x,y) in month t, and ε(x,y,t) refers to actual light energy utilization (MJ·m²·month⁻¹) of the pixel at position (x, y) in month t (g C·MJ⁻¹).

$$\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}\left(\mathrm{x},\mathrm{y},\mathrm{t}\right)=\mathrm{S}\mathrm{O}\mathrm{L}\left(\mathrm{x},\mathrm{y},\mathrm{t}\right)\mathrm{\ast }\mathrm{F}\mathrm{A}\mathrm{P}\mathrm{A}\mathrm{R}\left(\mathrm{x},\mathrm{y},\mathrm{t}\right)\mathrm{\ast }0.5$$

(7)

Among these, SOL(x,y,t) represents the total solar radiation energy (MJ·m⁻²) of the pixel at position (x,y) in month t, and FAPAR(x,y,t) refers to the fraction of the incident photosynthetic effective radiation assimilated by vegetation.

The factor 0.5 assumes that 50% of the incoming solar radiation is photosynthetically active radiation (PAR), consistent with standard CASA model assumptions. The CASA model thus enables dynamic, spatially explicit estimation of NPP, serving as a foundation for carbon cycle modeling and vegetation productivity assessment.

2.4. Nonlinear Relationship Assessment

Understanding the effects of multiple environmental and anthropogenic factors on ecosystem services (ESs) is critical due to the inherently nonlinear nature and strong spatiotemporal heterogeneity of ecological processes. The provision and maintenance of ES are influenced by a complex interplay of natural drivers—such as climate change, land-use dynamics, and biodiversity loss—and human interventions. These factors often interact in nonlinear ways, with multivariate feedback loops, which pose substantial challenges to accurate modeling and prediction. Traditional statistical methods, such as linear or logistic regression, are limited in their ability to capture such nonlinear interactions, as they typically assume fixed relationships between variables and cannot reveal the threshold or synergistic effects present in real-world ecological systems. To overcome these limitations, this study employs the eXtreme Gradient Boosting (XGBoost) algorithm—a tree-based ensemble machine learning method based on gradient boosting—which combines multiple weak learners (i.e., decision trees) to optimize predictive performance. Compared to traditional regression methods, XGBoost demonstrates superior capability in handling high-dimensional data, capturing nonlinear dependencies, and modeling variable interactions effectively [33,34]. The objective function in XGBoost is defined as:

$$\mathrm{Obj}=\mathrm{L}+\mathsf{\Omega }$$

(8)

Among these, $\mathrm{L}$ is the training loss function (e.g., mean squared error), and $\mathsf{\Omega }$ is a regularization term that penalizes model complexity.

Specifically, when using squared error loss:

$$\mathrm{L}=\sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{y}}_{\mathrm{i}}-\widehat{{\mathrm{y}}_{\mathrm{i}}}{)}^2$$

(9)

Among these, ${\mathrm{y}}_{\mathrm{i}}$ is the observed value and $\widehat{{\mathrm{y}}_{\mathrm{i}}}$ is the predicted value.

The complexity term is defined as:

$$\mathsf{\Omega }=\mathsf{\gamma }\mathrm{T}+{\displaystyle \frac{1}{2}}\mathsf{\lambda }\sum _{\mathrm{j}=1}^{\mathrm{T}}{\mathrm{w}}_{\mathrm{j}}^2$$

(10)

Among these, T is the number of trees, ${\mathrm{w}}_j$ is the weight of the $j$th leaf node, and $\mathsf{\gamma }$ and $\mathsf{\lambda }$ are regularization parameters that control overfitting.

To improve model interpretability, this study utilizes SHAP (SHapley Additive exPlanations), which explains the contribution of each input feature to a model’s prediction based on cooperative game theory. The SHAP value ${\mathsf{\varphi }}_{\mathrm{i}}\left(\mathrm{v}\right)$ for feature $\mathrm{i}$ is defined as:

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

(11)

Among these, N is the set of all features, S is a subset of features excluding i, and $\mathrm{v}\left(\mathrm{S}\right)$ is the prediction based on subset S.

SHAP values quantify the marginal contribution of each feature by averaging over all possible feature subsets [35], thereby revealing not only the importance but also the directionality and threshold behaviors of influencing factors. In this study, the XGBoost machine learning model was constructed using the Python3.8 language, and each key parameter was set as follows: learning rate = 0.02, number of estimators = 88, maximum tree depth = 7, and minimum child weight = 2.

To evaluate the model’s predictive performance, we compared predicted values to observed values for six key ES indicators: biodiversity maintenance (BM), carbon fixation (CF), crop production (CP), net primary productivity (NPP), soil retention (SR), and water yield (WY). Model accuracy was quantified using the coefficient of determination (R2) and mean squared error (MSE). As shown in Figure 2, all six scatter plots exhibit a strong agreement between predicted and actual values, with data points closely aligning along the 1:1 reference line. Quantitative results indicate high model performance, with R2 values exceeding 0.90 for all target variables—most surpassing 0.95—and MSE values consistently below 0.0045, demonstrating the model’s robustness and ability to capture the dominant variance in ecosystem service outcomes.

To clarify the validation procedures, we implemented a rigorous process to ensure model reliability. The dataset was split into 80% training and 20% testing sets, with performance metrics (R² and MSE) evaluated on the unseen test data to assess generalization. Additionally, we employed 5-fold cross-validation during hyperparameter tuning, where the data were divided into five subsets, with each fold used once as validation, while the remaining four served as training. This yielded average cross-validated R² values of >0.88 across folds, confirming consistent performance and minimal variance between training and validation sets.

For input features, six explanatory variables were selected based on theoretical frameworks, literature evidence [1,36], and the biophysical and socio-economic characteristics of Anhui Province. These include: DEM (digital elevation model), HFI (human footprint index), TEM (temperature), PRE (precipitation), PET (potential evapotranspiration), and LS (slope length–steepness factor). The selection of these variables is grounded in their established roles as primary drivers of ecosystem services in transitional regions like Anhui, as supported by multiple studies [8,37,38,39]. Specifically, topographic factors such as DEM and LS are crucial for capturing spatial heterogeneity in ES distribution, as they influence hydrological processes, soil erosion, and vegetation patterns—key elements in models like InVEST and RUSLE [27,40,41,42]. Climatic variables (TEM, PRE, PET) are selected due to their direct impact on biomass production, water balance, and regulatory services, with precipitation and temperature often identified as dominant drivers in climate-sensitive ecosystems [15,43]. HFI represents anthropogenic disturbances, integrating human activity intensity (e.g., urbanization and infrastructure) that suppresses ES functionality, as evidenced in global footprint analyses [40]. This combination balances natural (climatic and topographic) and human-induced dimensions, enabling the model to reveal nonlinear interactions and thresholds. Variables such as GDP, population, and NDVI from Table 1 were not included to minimize multicollinearity and redundancy. For instance, GDP and population density are highly correlated with HFI (which incorporates nighttime lights as a proxy for socio-economic activity), potentially inflating variance and reducing model interpretability [40]. NDVI, while valuable for vegetation assessment, is often collinear with climatic drivers (e.g., PRE and TEM) and ES outcomes like NPP, making it more suitable as an intermediate or output variable rather than an independent input [32,41]. Excluding these ensures focus on non-redundant, high-importance predictors that effectively capture core dynamics, as validated by variable importance rankings and cross-validation tests. This approach aligns with comprehensive research on ES drivers, prioritizing parsimony for robust predictions in similar studies [12,15].

3. Results

3.1. Spatial Patterns of Ecosystem Services

Based on the ecosystem service (ES) evaluation results for Anhui Province from 2000 to 2020, significant spatial heterogeneity was observed across all six ES indicators (Figure 3). These spatial patterns were closely associated with regional natural geographic conditions and the intensity of human activities [36]. Specifically, biodiversity maintenance (BM), water yield (WY), carbon fixation (CF), net primary productivity (NPP), and soil retention (SR) exhibited a distinct south-high, north-low gradient. This distribution corresponds to the dense forest cover and favorable ecological baseline found in the Dabie Mountains and the southern Anhui mountain regions. In contrast, crop production (CP) displayed the opposite spatial pattern, with high values concentrated in the Huaibei Plain and the Jianghuai Hills—traditional agricultural zones—highlighting the spatial trade-off between agricultural production and ecological conservation.

Overall, the BM, CF, NPP, and WY indicators consistently displayed higher values in southern mountainous regions. The Dabie Mountains and southern Anhui (e.g., Huizhou area), characterized by high elevations and complex topography, were dominated by deep blue or green zones, with ES index values mostly ranging from 0.7 to 1.0. This indicates strong vegetation coverage, high ecological connectivity, robust primary productivity, and effective water retention capacity. The transitional Jianghuai Hills in central Anhui showed moderate values (0.4–0.7), while the low-elevation northern plains (20–50 m) such as the Huaibei Plain and parts of the Jianghuai Plain were predominantly pale in color, with BM, NPP, and WY values < 0.4 and CF < 0.3, reflecting flat terrain, low vegetation heterogeneity, limited infiltration, and weak soil organic matter accumulation. Conversely, crop production (CP) reached its highest levels in the northern and eastern lowland agricultural regions, including the Huaibei and Yangtze River plains. CP index values in these zones ranged from 0.8 to 1.0, indicating a well-structured cropping system and high agricultural productivity potential. In contrast, CP values in the mountainous south were lower (0.2–0.5) due to steep slopes, shallow soils, and limited suitability for large-scale cultivation. The soil retention (SR) pattern showed a “band–patch” spatial structure. High values (SR > 0.6) were concentrated along the steep forested slopes of the Dabie Mountains and southern Anhui, with localized hotspots reaching 0.8–1.0 in areas undergoing vegetation restoration and terracing. In contrast, the flat plains experienced lower SR values (0.1–0.3) due to concentrated surface runoff and low vegetation coverage. In summary, the spatial distribution of ecosystem services in Anhui Province is primarily driven by the interplay of topography, vegetation structure, and precipitation, which together create distinct ecological gradients across the region.

3.2. Trends in Ecosystem Service Changes

3.2.1. Spatial Heterogeneity Analysis

Based on a combined analysis using Sen’s slope estimator and the Mann–Kendall trend test, notable spatial heterogeneity in the dynamics of six ecosystem services (ESs) in Anhui Province from 2000 to 2020 was revealed (Figure 4). For biodiversity maintenance (BM), degradation followed a clear urban–rural gradient. In peri-urban areas, such as the Hefei metropolitan region, 0.0758% of the land showed significant degradation, primarily due to the expansion of construction land, resulting in habitat fragmentation and corridor disruption. In the Huaibei Plain, 41.112% of the area experienced slight degradation, likely associated with landscape homogenization driven by intensive agricultural expansion. Meanwhile, 35.216% of the central to southern region showed slight improvement and 15.900% significant improvement, mainly attributed to ecological restoration projects such as reforestation and the establishment of nature reserves.

Carbon fixation (CF) exhibited the highest level of spatial stability, with 98.307% of the region maintaining stable conditions, thanks to the persistent carbon stock in the mature subtropical evergreen broadleaf forests of southern Anhui. However, 1.302% of land near urban areas showed significant degradation, reflecting carbon sink losses due to urban expansion—e.g., northwest Hefei where forest land was converted.

Crop production (CP) showed a dualistic spatial pattern: in the Huaibei Plain, 64.083% of land experienced significant improvement due to agricultural modernization and high-standard farmland development. In contrast, around urban centers such as Hefei and Huangshan, 15.799% of land (10.661% slightly degraded and 5.138% significantly degraded) experienced productivity losses—reflecting the conversion of high-quality farmland to non-agricultural uses during rapid urbanization.

Soil retention (SR) and water yield (WY) were mainly driven by natural geographic factors, showing north–south differentiation. SR significantly improved in 12.666% of steep forested areas in southern Anhui, while 24.505% of the northern plains showed slight degradation due to seasonal rainfall concentration and tillage-induced topsoil erosion. WY exhibited a “split zone” pattern along the topographic watershed: 29.766% of central Anhui saw slight decline, while 66.121% of southern Anhui improved, indicating terrain control on hydrological processes.

Net primary productivity (NPP) improved across 88.09% of the region, with 50.125% showing significant improvement, particularly in Dabie and southern mountainous areas—likely due to forest structural optimization and CO₂ fertilization effects. In contrast, 4.339% of the northern plains showed slight degradation, possibly due to soil organic matter depletion under intensive farming. The spatial correlation between NPP improvement and gains in BM and SR further supports the central role of vegetation recovery in enhancing multiple ES functions simultaneously.

3.2.2. Proportional Change Statistics

To further illustrate the differences in functional evolution among ES, the proportion of land pixels falling into five change classes—significantly degraded, slightly degraded, stable, slightly improved, and significantly improved—was calculated for all six ES types (

Table 3. Proportion of pixel-based ecosystem service changes (2000–2020, %).

Type	BM	NPP	CP	CF	WY	SR
Significantly degraded	1%	1%	5%	1%	0%	0%
Slightly degraded	41%	4%	11%	0%	30%	25%
Stable	7%	6%	0%	98%	0%	12%
Slightly improvement	35%	38%	20%	0%	66%	51%
Significantly improvement	16%	50%	64%	0%	4%	13%

). Results show that improvement dominates across services, suggesting a co-evolutionary trend toward multifunctional ecosystem enhancement.

NPP and CP had the highest shares in the “significantly improved” class, accounting for 50.12% and 64.08%, respectively—demonstrating substantial gains in ecosystem productivity over the past two decades. BM showed a clear upward trend as well, with 51% of pixels falling under either slightly or significantly improved, indicating the effectiveness of ecological restoration programs such as reforestation and ecological corridor construction. In contrast, CF was almost entirely stable (98.31%), confirming its structural persistence and low vulnerability to short-term anthropogenic disturbance. However, carbon stock loss in urban fringe areas requires policy attention. Regulatory services—WY and SR—showed improvement mainly in the “slightly improved” class (66.0% and 51.0%, respectively), indicating moderate but recoverable responses to changes in terrain, vegetation, and precipitation. Although overall degradation was limited, BM, WY, and SR showed widespread but low-level retreat, implying latent risks to ecosystem functionality that may require targeted monitoring and adaptive intervention.

In summary, both the statistics and spatial trends reveal strong spatial coupling and structural regularity in ES enhancement. Production and regulation services are most sensitive to ecological restoration efforts, while services such as CF remain stable over time—providing a scientific basis for prioritizing ecosystem management strategies.

3.2.3. Mean Trend Analysis

Time series statistics (2000–2020) of the six ES indicators were compiled to explore temporal trends, variability, and response characteristics (

Table 4. Annual average values of ecosystem services (2000–2020).

Time	BM	NPP	CP	CF	WY	SR
2000	69.02	66.87	203.18	395.83	150.02	794.94
2001	78.46	67.08	402.70	451.42	122.10	429.74
2002	111.67	67.15	181.23	504.85	202.03	762.43
2003	135.13	67.25	196.93	467.67	203.89	940.59
2004	96.27	67.40	136.76	500.96	137.77	510.93
2005	102.24	67.44	225.65	454.63	159.67	708.18
2006	90.20	67.51	213.59	469.05	139.89	598.70
2007	124.92	67.54	246.13	516.30	145.07	657.14
2008	116.80	67.50	235.87	513.07	149.22	609.42
2009	110.26	67.47	263.02	487.29	161.37	662.03
2010	87.85	67.47	272.25	499.57	234.76	813.46
2011	85.97	67.50	276.93	499.21	129.07	496.23
2012	79.26	67.39	282.88	501.33	150.60	472.58
2013	111.66	66.98	268.10	494.92	142.56	510.94
2014	132.16	66.85	267.69	568.39	193.54	736.67
2015	121.28	66.73	279.27	557.17	263.73	948.60
2016	103.29	66.66	291.24	506.19	238.83	917.21
2017	112.76	66.67	281.84	505.60	173.60	654.46
2018	118.48	66.67	286.42	537.23	170.16	694.28
2019	82.24	66.58	287.68	484.99	157.50	539.11
2020	107.48	66.408	283.788	537.598	243.288	935.18

).

NPP and CP consistently remained at high levels, showing phased upward trends. NPP increased from 395.83 in 2000 to 537.59 in 2020, with a pattern of early fluctuation, mid-term stability, and late acceleration—likely driven by improved climatic conditions, elevated CO₂ levels (fertilization effect), and optimized forest structure. CP rose from 203.18 to 283.78, reflecting the long-term impact of agricultural modernization and farmland infrastructure enhancement. BM experienced an oscillating upward trend, especially between 2000–2003, rising from 69.02 to 135.13, and remained high in 2014–2015, signaling the positive impact of ecological projects like reforestation and corridor development on biodiversity maintenance. In contrast, CF showed minimal interannual variation, consistently between 66.40–67.53, confirming the long-term structural stability of carbon storage—also reflected in its dominance in the “stable” category discussed in Section 3.2.2. SR displayed a U-shaped trend—early fluctuation, a drop to a low of 137.77 in 2004, then a steady increase to 243.28 by 2020—likely due to climate extremes and changing slope management policies (e.g., forest/grassland restoration). WY exhibited a two-stage pattern: drastic early fluctuation and later recovery. After dropping from 794.94 in 2000 to 429.74 in 2001, WY rebounded to 940.59 by 2003 and peaked again in 2015 and 2020, reflecting strong sensitivity to precipitation variability, runoff regulation, and vegetation recovery.

In summary, from 2000 to 2020, most ESs in Anhui Province showed a pattern of predominant improvement, partial stability, and alternating fluctuation. Productive and regulatory services (e.g., NPP, CP, SR, WY) were more responsive to human management and climate variability, while CF reflected the “slow-variable” nature of ecosystem structure. The co-improvement of multiple services provides quantitative evidence of overall ecosystem enhancement and forms a foundation for developing targeted service-based regulation strategies.

3.3. Influencing Factors of Ecosystem Services

To quantitatively reveal the dominant drivers of spatiotemporal variation in six key ecosystem services (ESs) in Anhui Province—crop production (CP), water yield (WY), soil retention (SR), biodiversity maintenance (BM), carbon fixation (CF), and net primary productivity (NPP)—we employed SHAP (SHapley Additive exPlanations) analysis on top of an XGBoost machine learning model. Specifically, we used SHAP summary plots in the form of beeswarm diagrams to visualize the relative importance and effect patterns of six key environmental variables (Figure 5): digital elevation model (DEM), human footprint index (HFI), temperature (TEM), precipitation (PRE), potential evapotranspiration (PET), and slope length–steepness factor (LS). The beeswarm plot integrates three key layers of information into a single Figure 6:

Mode 1: Precipitation-Dominated Hydrological Regulation Services. This mode includes water yield (WY), crop production (CP), and soil retention (SR). For both WY and CP, annual precipitation (PRE) emerged as the most influential variable. The beeswarm plots clearly show that high PRE values (red dots) align with high positive SHAP values, confirming that abundant rainfall is a decisive factor in enhancing water provisioning and agricultural productivity. In contrast, for SR, although PRE remains the most important driver, its effect is negative—increased precipitation intensifies erosion potential, thereby reducing soil retention. This finding reflects the dual role of rainfall in ecosystem regulation: it supports productivity but can also drive degradation via runoff and erosion.

Mode 2: Topography-Dominated Carbon Sequestration and Productivity Services. This mode encompasses NPP and CF, both of which are dominantly driven by DEM. Elevation acts not only as a direct terrain measure but also as a proxy for land cover and ecosystem type. The beeswarm plot shows that high-elevation zones (red dots) correspond to positive SHAP values, while low-elevation zones (blue dots) align with negative impacts. This pattern mirrors the geographic reality of Anhui: high-elevation southern and western regions are forest-rich, with strong photosynthetic activity and large carbon stocks, in contrast to low-elevation plains dominated by farmland. Therefore, DEM effectively differentiates between forest–mountain systems and farmland–plain systems, making it the strongest predictor of spatial NPP and CF variation.

Mode 3: Multi-Factor Synergistic Habitat Support Services. Biodiversity maintenance (BM) exhibited the most complex driver mechanism. Although precipitation (PRE) ranked highest, temperature (TEM) and LS (slope factor) also showed strong influence, highlighting the cooperative interaction of water, heat, and topographic heterogeneity in shaping species distributions.

Anhui lies in a climatic transition zone, with diverse temperature gradients that support species richness. Complex terrain generates microhabitats and elevational niches, enabling coexistence and biodiversity persistence. This mode reflects the non-additive, synergistic nature of habitat provision across landscapes.

Feature importance is ranked vertically based on the mean absolute SHAP value for each variable—the higher a variable appears, the more influential it is to the model prediction. Influence direction and magnitude are represented by SHAP values on the horizontal axis: positive values indicate a promoting effect (e.g., increasing service value), while negative values indicate suppression. Each colored dot represents an individual sample (e.g., a grid cell), and the color reflects the original feature value (red = high, blue = low). By analyzing both position and color, we can clearly infer the nature of nonlinear relationships. For instance, if red dots (high values) concentrate on the positive SHAP side, the feature has a strong positive influence on the service. This visualization framework revealed three dominant driving modes in Anhui Province, primarily shaped by climate and topography, with anthropogenic impacts (HFI) playing a secondary role.

To further investigate the complex nonlinear relationships and identify critical thresholds in how environmental factors influence ecosystem services, this study employed SHAP dependence plots for detailed analysis of model outputs. Unlike SHAP summary plots, dependence plots explicitly show how variations in a single feature affect its marginal contribution to the model’s predictions. By pinpointing changes in SHAP value direction or slope, we were able to determine inflection points—thresholds where the effect of a given driver shifts significantly. These thresholds are crucial for understanding potential ecological transitions and for informing precise, responsive management strategies (Figure 7).

Annual precipitation (PRE) exhibited particularly strong threshold effects in hydrologically driven services. For water yield (WY), a clear inflection was observed at 1220 mm. Below this level, precipitation is largely consumed by evapotranspiration and soil moisture retention, leading to negative SHAP values. Once precipitation exceeds 1220 mm, however, it begins to generate substantial surface and subsurface runoff, and its contribution to water yield rapidly becomes positive. This indicates a key tipping point in the shift from water limitation to water surplus. For crop production (CP), 1150 mm of precipitation marked a decisive threshold. Below this, crop growth is constrained by water stress; above it, rainfall sufficiently meets crop water demand, and SHAP values remain strongly positive. Similarly, in the case of soil retention (SR), 1150 mm also emerged as a critical point—beyond which increased rainfall intensifies erosive forces, sharply increasing soil loss risk and causing SHAP values to turn negative, indicating a loss of service function.

Vegetation growth and carbon sequestration services, driven jointly by topography and climate, also exhibited multi-threshold responses. For net primary productivity (NPP), the elevation threshold was found to be 70 m. Below this point, lowland agricultural areas dominate, with generally lower NPP values; above it, the transition to hilly and mountainous forested terrain results in significantly greater productivity. Temperature responses for NPP followed a typical unimodal pattern, with optimal SHAP values concentrated between 14.8 °C and 16.8 °C, the most favorable range for photosynthetic accumulation. Values below or above this range led to diminished marginal contributions. Carbon fixation (CF) showed similar behavior, with a corresponding elevation threshold around 80 m, reflecting the superior carbon sink capacity of higher-altitude forest systems. Additionally, a slope-length factor (LS) threshold of approximately 5 m was observed, separating flat lowland zones with weaker sequestration potential from steeper regions where vegetation density and litter accumulation enhance carbon storage.

As an integrated service, biodiversity maintenance (BM) exhibited a more composite threshold response shaped by multiple factors. Precipitation showed a threshold effect around 1200 mm, beyond which sufficient water availability becomes a foundational condition for supporting high biodiversity. Temperature exerted a pronounced positive influence above 15.5 °C, likely corresponding to a transition from warm-temperate to subtropical bioclimatic zones, which favor richer species assemblages. Meanwhile, potential evapotranspiration (PET) exhibited a negative threshold near 1100 mm: beyond this point, excessive atmospheric water demand may trigger drought stress, thus reducing biodiversity maintenance and species persistence.

4. Discussion

This study integrated the InVEST and CASA models to systematically evaluate the spatiotemporal evolution of six key ecosystem services (BM, CF, CP, NPP, SR, and WY) in Anhui Province from 2000 to 2020. The use of the XGBoost–SHAP framework allowed for the identification of dominant drivers and their nonlinear responses. The results revealed a pronounced “high in the south, low in the north” spatial pattern of ecosystem services, closely tied to topographic gradients, vegetation cover, and human activity intensity. Temporally, most services showed an overall trend of improvement, though urban–rural disparities still pose degradation risks. These findings not only deepen our understanding of ecological dynamics in China’s transitional eastern interior but also provide quantitative evidence to support regional ecosystem management.

Rather than merely confirming spatial and temporal patterns, these results engage with broader debates on ecosystem service trade-offs and resilience in rapidly urbanizing regions. For instance, the inverse distribution of CP (high in northern plains) versus regulatory services like BM and WY (high in southern mountains) aligns with the “production–protection” conflict highlighted in Costanza et al.’s (2014) global valuation frameworks [38], but our nonlinear threshold analysis adds nuance by quantifying precipitation tipping points (e.g., 1200 mm for WY), which could inform adaptive management in climate-vulnerable areas. Comparatively, similar studies in China’s Yangtze River Economic Belt, such as Chen et al. (2020) on land-use impacts in the middle reaches [20], report analogous north–south gradients driven by urbanization; however, our incorporation of HFI reveals stronger anthropogenic suppression in Anhui’s urban fringes than in less densely populated provinces like Jiangxi, suggesting a need for tailored ecological compensation mechanisms. Internationally, our findings resonate with European ES assessments [14], where precipitation thresholds similarly influence water services but differ in scale: European models often emphasize policy-driven restoration (e.g., EU Green Deal), whereas Anhui’s gains stem from national initiatives like Grain for Green, highlighting the role of centralized governance in transitional economies. Critically, while these comparisons affirm the universality of climatic drivers, they underscore gaps in integrating socio-economic feedbacks—our HFI analysis addresses this, but future work could explore dynamic interactions using agent-based models to better simulate urban expansion scenarios [41].

Driver analysis further underscores the leading roles of climate and topographic variables. XGBoost results identified precipitation (PRE) and elevation (DEM) as top-ranked predictors for multiple services. SHAP dependence plots confirmed clear nonlinear effects, with PRE showing peak contributions within the 1100–1250 mm range. This supports prior studies in European ES assessments that found precipitation thresholds beyond which flood risks rise rather than benefits [8]. The unimodal temperature response curve for NPP, peaking around 15 °C, reflects an optimal thermal range for vegetation growth, consistent with temperature sensitivities observed in mid-latitude ecosystems [39]. The negative influence of the human footprint index (HFI) on NPP and BM suggests a significant urbanization pressure effect, matching global footprint maps that show 10–20% service loss in high-HFI zones [40]. Furthermore, the interaction between drivers (e.g., PRE and PET) intensified variability, highlighting the importance of accounting for multi-factor coupling, as also emphasized in studies employing geographic detectors [41]. Nevertheless, this study acknowledges uncertainties in model assumptions that could influence results. For instance, the CASA model’s NPP estimates rely heavily on NDVI inputs, which may introduce biases in cloudy or seasonally variable regions like Anhui, potentially overestimating productivity in agricultural zones [31]. Similarly, InVEST’s soil retention module simplifies erosion processes (e.g., assuming uniform rainfall erosivity), overlooking micro-scale factors like soil microbial activity or extreme events, as noted in RUSLE critiques [30]. These limitations could amplify uncertainties in threshold identifications, particularly for WY and SR, where hydrological data resolution (e.g., interpolated AET) may not fully capture local variability. Future research should incorporate sensitivity analyses or ensemble modeling to quantify these uncertainties and enhance predictive confidence [43].

These findings have clear policy implications for operationalizing threshold insights in ecosystem management and land-use planning. As a core province of the Yangtze River Economic Belt, Anhui should prioritize ecological conservation in the southern mountains, serving as a regional green buffer, while also promoting sustainable agricultural practices in the northern plains to reduce trade-offs. For example, zones with PRE > 1200 mm could be designated as priority water conservation areas through zoning regulations, integrating SHAP-derived thresholds into spatial planning tools like GIS-based decision support systems—an approach aligned with the United Nations’ Sustainable Development Goal 15 on ecosystem restoration [42]. To operationalize this, policymakers could establish monitoring networks for real-time PRE and TEM tracking, triggering adaptive measures (e.g., afforestation in erosion-prone LS > 5 areas) when thresholds are approached. Furthermore, incorporating HFI thresholds (e.g., >0.5 signaling high risk) into urban planning could guide green infrastructure investments, such as ecological corridors in Hefei’s fringes, fostering resilience against urbanization pressures. By linking nonlinear drivers to actionable strategies, this study provides a blueprint for evidence-based governance in similar transitional regions.

Nevertheless, this study acknowledges uncertainties in model assumptions that could influence results. For instance, the CASA model’s NPP estimates rely heavily on NDVI inputs, which may introduce biases in cloudy or seasonally variable regions like Anhui, potentially overestimating productivity in agricultural zones [31]. Similarly, InVEST’s soil retention module simplifies erosion processes (e.g., assuming uniform rainfall erosivity), overlooking micro-scale factors like soil microbial activity or extreme events, as noted in RUSLE critiques [30]. These limitations are particularly relevant under extreme climate conditions, such as floods or droughts, where abrupt nonlinear responses (e.g., intensified erosion or altered water yields) may exceed the models’ static assumptions, leading to inaccuracies in threshold estimations. For example, the identified PRE threshold of ~1200 mm for WY could be less reliable during intense rainfall events, as InVEST and CASA do not fully account for dynamic hydrological feedbacks or vegetation stress in high-stress scenarios, potentially underestimating tipping points in vulnerable transitional zones like Anhui. Future research should incorporate sensitivity analyses or ensemble modeling to quantify these uncertainties and enhance predictive confidence [43]. Moreover, including more dynamic socio-economic indicators (e.g., GDP trajectories) could enrich the understanding of human drivers.

5. Conclusions

From 2000 to 2020, ecosystem services (ESs) in Anhui Province exhibited an overall trend of improvement. A distinct spatial gradient emerged, characterized by enhanced regulating services in the south and stronger provisioning services in the north, reflecting the complex co-influence of geographical conditions, climatic variables, and human activities on ecosystem dynamics. The main conclusions are as follows:

• From 2000 to 2020, ecosystem services (ESs) in Anhui Province exhibited an overall trend of improvement, with a distinct “high in the south, low in the north” spatial gradient. Regulating services like biodiversity maintenance (BM), net primary productivity (NPP), soil retention (SR), and water yield (WY) were enhanced in southern mountainous areas, while provisioning services such as crop production (CP) improved in northern plains. NPP and CP showed the most significant gains (50.12% and 64.08% of areas, respectively), with mean values rising by 36% and 39%; carbon fixation (CF) remained stable across 98.31% of the region, indicating associations with vegetation and land-use dynamics.
• Precipitation (PRE) emerged as the dominant driver for most services, with optimal effects in the 1100–1250 mm range, while elevation (DEM) was key for CF and NPP. Temperature (TEM) showed nonlinear responses around 15 °C, and the human footprint index (HFI) exerted negative influences, particularly where HFI > 0.5, highlighting potential suppressive effects of human activity on ES functionality.
• Regulating services like SR and WY proved sensitive to natural variables, with improvement in 51.36% and 70.21% of areas, though degradation persisted in marginal zones. These findings underscore the need for spatially differentiated management strategies based on identified thresholds, such as prioritizing conservation in high-PRE southern areas and sustainable practices in northern plains to foster ecosystem resilience.