Spatio-Temporal Evolution and Driving Factors of Eco-Environmental Response to Land Use Transformation in China’s Southern Hilly Area During 2000–2020

Abstract

Hilly areas serve as critical ecological barriers yet face developmental challenges, drawing increasing attention to how land use transformation affects eco-environmental quality (EEQ). Systematic studies on EEQ drivers in complex terrains remain limited, particularly regarding nonlinear and interactive effects. This study examines Zhejiang’s hilly area—typical of southern China’s hills—using land use data from 2000, 2010, and 2020. Methods including land use transfer matrix, EEQI, hotspot analysis, and XGBoost-SHAP were applied to assess impacts and quantify drivers. Results show a slight but consistent decline in EEQ index (EEQI) (0.7635 to 0.7472), driven primarily by rapid built-up land (BL) expansion (276.41% increase). NDVI was the most influential factor (SHAP: 0.1226, >59%), followed by GDP per unit area and temperature. NDVI showed a threshold effect (>0.65 strengthens benefit), and strong interaction with per capita GDP. A slope-vegetation coupling mechanism was identified: on slopes > 30°, high NDVI significantly amplifies EEQ improvement, highlighting the importance of vegetation conservation on steep slopes. These findings provide a scientific basis for targeted land management in hilly regions of southern China and similar areas.

1. Introduction

Since the 21st century, China’s territorial spatial pattern has undergone profound restructuring under the dual objectives of rapid urbanization and ecological protection [1,2]. This process encompasses changes in land use types, structural adjustments, and multidimensional spatial reorganization [3,4]. Against the backdrop of ecological civilization construction being elevated to a national strategy in China, territorial spatial transformation has emerged as a core pathway for harmonizing human-land relationships and optimizing spatial structures [5]. As a critical indicator for assessing regional sustainable development levels, EEQ exhibits dynamic changes closely linked to land use patterns [6].

Significant scholarly achievements have accumulated regarding the eco-environmental effects of land use transformation, primarily concentrating on the development of EEQ evaluation methods, quantitative analyses of the impacts of land use transformation on EEQ, and investigations into spatial heterogeneity [7,8,9]. Commonly employed methodologies include the land use transfer matrix, ecological environment quality index (EEQI), ecosystem service value assessment models, and landscape pattern analysis [10,11,12]. Research indicates that early urban expansion often induced localized ecological degradation. However, following the advancement of China’s ecological civilization strategy, numerous regions have shifted towards quality-enhancing development models, resulting in a deceleration or even reversal of negative EEQ trends [13]. Spatially, pronounced ecological pressures are evident in the urbanized eastern coastal regions of China, whereas conditions in ecologically fragile central-western areas have improved due to policy interventions such as the Grain for Green Program [14,15,16]. Nevertheless, extensive research has focused on national, provincial, or plain/arid region scales, while systematic research targeting China’s southern hilly area remains relatively scarce. This region features fragmented terrain, diverse ecosystems, and a sensitive human-land relationship, resulting in a potential EEQ response mechanism that may significantly differ from other areas.

EEQ dynamics are co-driven by multiple factors, encompassing natural conditions and socioeconomic development [17]. Among these, natural elements such as topography and climate define the baseline conditions and sensitivity of regional ecosystems. Socioeconomic factors, including population density and GDP growth, influence ecological environment quality both directly and indirectly through land use transformation [18,19]. In recent years, scholars have widely adopted multiple regression models, geographical detector models, geographically weighted regression (GWR), structural equation modeling (SEM), and machine learning approaches to identify and quantify the influence and underlying mechanisms of various drivers on EEQ [20,21]. However, these approaches typically exhibit certain limitations: (1) They often struggle to effectively capture complex nonlinear relationships among variables, leading to an insufficiently profound understanding of driving mechanisms; (2) Analyses of interactions between driving factors remain relatively inadequate; (3) While capable of handling complex relationships, the “black-box” nature of machine learning methods poses challenges to model interpretability. A novel framework synergistically integrating XGBoost (for elucidating nonlinear relationships) and the SHAP explainer method (addressing “black-box” interpretability issues, quantifying importance, and revealing interactions) can circumvent these limitations, enabling the identification and quantification of key drivers influencing EEQ.

Although some studies have focused on the eco-environmental response mechanisms in the eastern plains of China, relatively less attention has been paid to the hilly regions in the south [22]. The eastern plains, characterized by gentle terrain and relatively homogeneous spatial factors, exhibit spatial agglomeration patterns dominated primarily by a single gradient of development intensity. In contrast, the hilly areas of southern China are marked by high topographic fragmentation, significant slope variability, rich biodiversity, high erosion sensitivity, and notable topographic constraints on construction activities. The superposition of vertical gradients along terrain features and horizontal gradients of urbanization has led to systematic differences in how EEQ responds to land use transformation compared to the eastern plains. However, there remains a lack of systematic research on the response of EEQ and its driving factors within complex hilly terrains, particularly regarding nonlinear influences and interaction effects.

The hilly area of Zhejiang serves as an ideal window for observing the complexity of eco-environmental responses. Since the implementation of the “Thousand Village Demonstration, Ten Thousand Village Improvement” project in 2003, Zhejiang has achieved remarkable success in integrating village restructuring, ecological restoration, and high-quality development, establishing itself as a national model [23]. Conducting a systematic study on the relationship between land use transformation and EEQ in the hilly area of Zhejiang as a typical case will not only help reveal the ecological effects of land use changes in complex geographical settings but also provide valuable insights and references for other hilly area in southern China.

Building upon this foundation, the hilly area of Zhejiang Province were selected as the study area. Land use data from 2000, 2010, and 2020 were utilized, and methods including the land use transfer matrix, EEQI, hotspot analysis, and other spatial analysis techniques were employed to reveal the response of EEQ to territorial spatial transformation. The XGBoost-SHAP method was innovatively introduced to analyze the nonlinear influences of driving factors. This study was designed to: (1) systematically examine the spatiotemporal response of EEQ to land use changes in hilly area; (2) quantitatively evaluate the relative importance of various driving factors, identify thresholds of their nonlinear effects, and quantify interaction effects between key factors by integrating the XGBoost machine learning model with the SHAP interpretability framework. The limitations of traditional linear models in capturing complex relationships were overcome through this methodology, which represents the core innovation of this research.

2. Materials and Methods

2.1. Overview of the Study Area

China’s southern hilly area represents a globally significant ecological region characterized by its vast coverage of subtropical evergreen broad-leaved forests, which play a crucial role in carbon sequestration, biodiversity conservation, and water regulation. This region is typified by complex topography and long-standing human-land interactions. Focusing on this broader context, this study selects the hilly area of Zhejiang Province as a representative case. Zhejiang, located on China’s southeastern coast, lies within the subtropical monsoon climate zone and is characterized predominantly by hilly terrain. The study area comprises 26 hilly counties, covering the entire prefectures of Lishui and Quzhou, as well as parts of Hangzhou, Wenzhou, Jinhua, and Taizhou. This region exemplifies the low-relief hilly landscapes typical of southern China (Figure 1) and represents a core implementation zone of Zhejiang’s “Thousand-Village Demonstration, Ten-Thousand-Village Renovation” project and rural revitalization strategy.

The area possesses a sound ecological foundation, with forest coverage generally exceeding 70% and abundant water resources. However, it also exhibits relatively high ecological vulnerability and is particularly sensitive to land use changes. Over the past two decades, driven by ecological civilization strategies and continuous rural policy implementation, the region has undergone significant land use transformation. This includes the consolidation and optimization of rural residential areas, spatial restructuring of built-up land, and shifts in agricultural land use types. These changes have profoundly influenced both socioeconomic development and EEQ.

The period from 2000 to 2020 was selected to cover the critical phase of ecological restoration policies and investigate the medium—to long-term impacts of land use transformation.

2.2. Data Sources

Land use data (2000, 2010, 2020; spatial resolution: 30 m × 30 m) [24] and GDP per unit area data (spatial resolution: 1 km × 1 km) [25] were obtained from the Data Center for Resources and Environmental Sciences, Chinese Academy of Sciences (https://www.resdc.cn/, accessed on 10 April 2023 and 21 April 2025). DEM data (spatial resolution: 30 m × 30 m) were acquired from the Geospatial Data Cloud (http://www.gscloud.cn/, accessed on 18 April 2025). Temperature data (spatial resolution: 1 km × 1 km) were sourced from the National Tibetan Plateau/Third Pole Environment Data Center (http://data.tpdc.ac.cn, accessed on 18 April 2025) [26]. NDVI data (spatial resolution: 1 km × 1 km) were downloaded from NASA’s EarthData portal (https://www.earthdata.nasa.gov/, accessed on 5 August 2024) [27]. Population density data (spatial resolution: 1 km × 1 km) were provided by WorldPop (https://www.worldpop.org/, accessed on 21 April 2025). Following established methodologies [28] and considering regional characteristics, land use types were classified into six categories: Cropland (CL), Forestland (FL), Grassland (GL), Water Bodies (WB), Built-up Land (BL), and Unused Land (UL). All datasets were uniformly reprojected to the “WGS_1984” coordinate system and resampled to a 1km grid resolution.

2.3. Methods

2.3.1. Land Use Transformation

(1)

Land Use Transfer Matrix

Land Use Transfer Matrix is a widely employed quantitative analytical tool for identifying mutual transformation relationships and intensities between different land use categories [29]. This method uses a two-dimensional matrix to capture land use gains and losses, revealing dynamic change processes. Specifically, the rows represent land use types in the initial year, the columns represent land use categories in the terminal year, and each element within the matrix quantifies the area transitioning from one land use category to another. This method enables the identification of dominant land use transformation directions, key categories experiencing net loss or gain, and provides foundational data for supporting subsequent eco-environmental impact assessments. Its mathematical expression is given by:

$$s_{ij}=\left[\begin{array}{cccc}{s}_{11}& s_{12}& \dots & s_{1n}\\ s_{21}& s_{22}& \dots & s_{2n}\\ \dots & \dots & \dots & \dots \\ s_{n1}& s_{n2}& \dots & s_{nn}\end{array}\right]$$

(1)

where n denotes the total number of land use categories in the study area, and s_ij represents the area converted from category i to category j during the study period.

(2)

Gravity Model

The gravity center model is a widely applied spatial analysis technique that integrates temporal and spatial information to visually represent the dynamic evolution of regional elements [30,31]. In this study, the gravity center model was employed to compare the coordinates of the gravity centers for each land use category at the three time points, thereby determining the spatial expansion direction and migration velocity of land use changes. Its calculation formula is expressed as follows:

$$\overline{x}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{M}_{ti}{x}_i}}{{\displaystyle {\sum }_{i=1}^n{M}_{ti}}}},\overline{y}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n{M}_{ti}{y}_i}}{{\displaystyle {\sum }_{i=1}^n{M}_{ti}}}}$$

(2)

where $\overline{x}$ and $\overline{y}$ denote the longitude and latitude coordinates, respectively, of the geographic gravity center for a specific land use category in year t; n represents the total number of spatial units; $x_i$ and $y_i$ are the geometric center coordinates of the i-th spatial unit; and M_ti signifies the attribute value (e.g., area) of the land use category within the i-th spatial unit in year t.

(3)

Standard Deviational Ellipse (SDE)

SDE method is a spatial econometric analysis technique capable of revealing the overall morphological characteristics of spatial element distributions [32,33]. This method generates an elliptical shape by calculating the projections of the standard deviations of a land use type’s spatial distribution onto its major and minor axes. It provides a visual representation of the directional trend and concentration level of the land use distribution [34]. The relevant formulas are:

$$\tan \theta ={\displaystyle \frac{\left({{\displaystyle {\sum }_{i=1}^n\overline{x_i}}}^2-{{\displaystyle {\sum }_{i=1}^n\overline{y_i}}}^2\right)+\sqrt{{\left({{\displaystyle {\sum }_{i=1}^n\overline{x_i}}}^2-{{\displaystyle {\sum }_{i=1}^n\overline{y_i}}}^2\right)}^2+4{\left({\displaystyle {\sum }_{i=1}^n\overline{x_i}}\overline{y_i}\right)}^2}}{2{\displaystyle {\sum }_{i=1}^n\overline{x_i}}\overline{y_i}}}$$

(3)

$${\sigma }_x=\sqrt{{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{\left(\overline{x_i}\cos \theta -\overline{y_i}\sin \theta \right)}^2}}{n}}},{\sigma }_y=\sqrt{{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{\left(\overline{x_i}\sin \theta -\overline{y_i}\cos \theta \right)}^2}}{n}}}$$

(4)

The center of the SDE corresponds to the geographic gravity center of a specific land use category. The rotation angle of the ellipse, denoted by θ, represents the clockwise angle between true north and the major axis and is derived from Equation (3). The semi-axis lengths of the SDE, ${\sigma }_x$ and ${\sigma }_y$, are calculated using Equation (4). The direction of the major axis indicates the orientation of greater dispersion, while the minor axis indicates the direction of higher aggregation. Consequently, an elongated ellipse signifies a discrete distribution of the land use category. Conversely, an ellipse approaching a circular shape indicates minimal difference between the lengths of the major and minor axes and reflects a more concentrated distribution.

2.3.2. Eco-Environmental Response

(1)

EEQI

EEQI is a land-use-classification-based metric widely applied for assessing regional EEQ [35]. It primarily reflects the EEQ during a specific period by integrating the EEQ value of each land use category and its proportional area [36]. This indicator not only considers the functional differences in various land use types within ecosystems but also captures the comprehensive impact of structural changes in land use on the eco-environment. The formula is expressed as:

$$E{V}_t={\displaystyle \sum _{i=1}^n{A}_{ki}\cdot \frac{R_i}{TA}}$$

(5)

where EV_t denotes the EEQI value in year t, R_i represents the EEQI weight assigned to the i-th land use category (

Table 1. EEQI weight of every land use category.

Land Use Category	Secondary Classification	EEQI Weight
CL	Paddy field	0.30
	Dry land	0.25
FL	Forest	0.95
	Scrubland	0.65
	Sparse forest	0.45
	Other forest	0.40
GL	High coverage grassland	0.75
	Medium coverage grassland	0.45
	Low coverage grassland	0.20
WB	Inland waterway	0.55
	Natural lake	0.75
	Artificial water	0.55
	Intertidal flat	0.45
	Floodplain	0.55
BL	Urban land	0.20
	Rural residential area	0.20
	Other construction land	0.15
UL	Marshland	0.65
	Bare soil	0.05
	Bare rock	0.01

, values referenced from [9]), n signifies the total number of land use categories, A_ki indicates the area of the i-th land use category, and TA is the total area. In this study, the EEQI classification was established using the natural breaks (Jenks) method in ArcGIS 10.8, supplemented by relevant literature [15]. The index values were categorized into five levels: lowest (0 ≤ EEQI < 0.2), low (0.2 ≤ EEQI < 0.5), moderate (0.5 ≤ EEQI < 0.7), good (0.7 ≤ EEQI < 0.85), and excellent (0.85 ≤ EEQI ≤ 1.00).

(2)

Hot spot analysis

Hot spot analysis, based on the Getis-Ord G_i* statistic, is a spatial statistical method used to identify statistically significant spatial clusters of high values (hot spots) or low values (cold spots) within a dataset [37]. Its fundamental principle involves calculating the correlation between the attribute value of each spatial unit and the values within its neighborhood to determine whether it is significantly higher or lower than the surrounding area [38]. In this study, hot spot areas typically represent regions with notable ecological environment improvement, while cold spot areas potentially face ecological degradation risks. This method facilitates the identification of spatial heterogeneity in eco-environmental responses. The formula is expressed as:

$${G_i}^{*}={\displaystyle {\sum }_{j=1}^n{W}_{ij}{Y}_j/}{\displaystyle {\sum }_{j=1}^n{Y}_j}$$

(6)

where G_i* denotes the hot spot statistic for spatial unit i, Y_j represents the attribute value at spatial unit j, W_ij is the spatial weight matrix defining the relationship between units i and j, and n signifies the total number of spatial units. A significantly positive G_i* value (G_i* > 0) indicates a hot spot, while a significantly negative value (G_i* < 0) identifies a cold spot. This study employed the Getis-Ord Gi* tool in ArcGIS 10.8 to perform hotspot analysis. The spatial relationship was determined using the fixed distance band method, with Euclidean distance selected as the distance method. Clusters with p-values < 0.1, <0.05, and <0.01 were identified as having 90%, 95%, and 99% confidence levels, respectively.

2.3.3. Driving Factors

(1)

Driving factors selection

Changes in EEQ are influenced by numerous factors. Based on relevant studies [39,40] and the specific context of the study area, natural factors and social factors were selected as the two primary categories of explanatory variables. To avoid multicollinearity among multiple explanatory variables [41,42], eight potential driving factors were chosen and tested for variance inflation factor (VIF) using SPSS 26. The results indicated that all VIF values were below 5, confirming the absence of multicollinearity, as presented in

Table 2. Driving factors of EEQ.

Type	Factors	Abbreviation	VIF
Natural factors	Elevation	ELV	1.627
	Slope	SLP	1.629
	Aspect	ASP	1.034
	Temperature	TMP	1.414
	NDVI	NDVI	1.449
Socioeconomic factors	Per capita GDP	GDP	1.340
	Population density	PD	1.323
	Road density	RD	1.021

.

(2)

XGBoost

XGBoost is an ensemble learning algorithm based on decision trees. This method iteratively trains multiple Classification and Regression Trees (CART), adjusting weights based on residuals in each iteration to minimize the loss function [43]. Compared to traditional machine learning algorithms, XGBoost significantly enhances training and prediction speeds through algorithmic optimization and parallel processing. It offers advantages in efficiency, accuracy, and the capability to model complex nonlinear relationships [44], and is widely applied in fields such as environmental science, remote sensing monitoring, and land use change studies [45]. The objective function is expressed as:

$$Obj\left(\theta \right)={\displaystyle \sum _{i=1}^{n}L\left(y_i,{\widehat{y}}_i\right)}+{\displaystyle \sum _{k=1}^K\mathrm{\Omega }\left(f_k\right)}$$

(7)

where $Obj\left(\theta \right)$ denotes the objective function, $L\left(y_i,{\widehat{y}}_i\right)$ represents the loss function between predicted values and true values, $\mathrm{\Omega }\left(f_k\right)$ is the regularization term that controls model complexity to prevent overfitting, and f_k signifies the output function of the k-th decision tree. The machine learning modeling and interpretation were conducted in Python 3.13. The XGBoost algorithm was implemented using the xgboost library.

(3)

SHAP

The SHAP interpretability model originates from the concept of Shapley values in cooperative game theory. SHAP assigns each feature a contribution score, known as the SHAP value, to quantify its impact on the model’s output [46]. This model can further decompose each feature’s contribution into its main effect and interaction effects, establishing SHAP as a universal framework for interpreting diverse algorithms [47]. This study employs the SHAP method to measure the relative importance of driving factors, their partial dependence relationships, and interaction effects. The SHAP values were calculated using the shap library.

3. Results

3.1. Spatio-Temporal Evolution of Land Use Transformation

3.1.1. Changes in Land Use in Study Area

The two predominant land use categories within the study area were FL and CL, collectively accounting for between 92.42% and 94.85% of the total area. FL experienced the largest decline, decreasing by 617.76 km². This was followed by CL, which decreased by 462.87 km². The proportions of GL and UL remained stable at 2.5% and 0.05%, respectively. WB area increased by 124.89 km², a growth of 15.10%. BL area expanded rapidly, increasing by 967.12 km², with a growth rate of 276.41%. Regarding phased changes, all six land use categories showed notably greater change magnitudes during 2000–2010 compared to 2010–2020 (Figure 2).

3.1.2. Characteristics of Land Use Transformation

A total of 1815.70 km² underwent land use transfer during the study period, accounting for 4.00% of the study area. The predominant transitions involved outflows from CL and FL and inflows to BL (

Table 3. Land use transfer matrix at each stage (Unit: km²).

Years	Type	CL	FL	GL	WB	BL	UL	Loss
2000–2010	CL	6038.33	175.02	7.63	83.18	349.86	0.11	615.80
	FL	280.19	35,716.31	100.97	69.76	203.93	2.86	657.70
	GL	6.54	74.60	1035.91	2.17	14.45	0.51	98.28
	WB	10.65	10.64	1.97	796.51	7.52	0.07	30.84
	BL	8.34	3.80	0.18	1.33	336.23	0.00	13.65
	UL	0.04	2.10	1.54	0.42	0.68	18.99	4.79
	Gain	305.77	266.16	112.29	156.85	576.44	3.55	1421.06
2010–2020	CL	6191.00		0.27	1.14	152.71		154.12
	FL		35,748.60	6.36	0.74	232.47		239.57
	GL		3.65	1130.49		14.98		18.63
	WB				949.90	20.99		20.99
	BL	0.25	4.01	13.55	0.46	895.25		18.26
	UL					0.60	21.95	0.60
	Gain	0.25	7.66	20.18	2.34	421.76	0.00	452.18
2000–2020	CL	5888.73	172.02	7.79	84.23	501.24	0.11	765.39
	FL	276.80	35,488.81	114.57	70.50	420.50	2.84	885.20
	GL	6.30	74.13	1023.77	1.98	27.49	0.51	110.41
	WB	10.44	10.19	1.97	790.95	13.74	0.07	36.41
	BL	7.92	3.55	0.10	1.33	336.97		12.92
	UL	0.04	2.09	1.54	0.42	1.28	18.42	5.37
	Gain	301.50	261.98	125.98	158.46	964.25	3.53	1815.70

). Specifically, the largest transferred area was from CL to BL (27.61% of total transfers), followed by FL to BL (23.16%) and FL to CL (15.24%). This indicates that encroachment on cropland and forests during industrial development and urban expansion was the primary driver of land use transformation. Examining phased changes, the 2000–2010 period recorded 1421.06 km² of transfers—representing an era of rapid transformation—dominated by CL to BL (24.62%), FL to CL (19.72%), and FL to BL (14.35%). Although the magnitude of change decreased during 2010–2020, transitions became more concentrated, predominantly featuring FL to BL (51.41%) and CL to BL (33.77%) (Figure 3).

This study further applied the SDE and gravity center models to analyze the spatial characteristics and spatiotemporal evolution of land use categories, with results shown in Figure 4. From 2000 to 2020, the SDE for all land use categories covered most areas, demonstrating relatively stable spatial differentiation patterns. The major axes consistently exhibited a northwest-southeast orientation, indicating greater dispersion of land use distributions along this direction due to the region’s topographic characteristics of higher elevation in the southwest and lower elevation in the northeast. Temporally, SDE underwent significantly greater changes during 2000–2010 than 2010–2020, confirming the earlier decade as the period of more intensive land use transitions. Spatially, BL’s ellipse shifted substantially southward while its minor axis elongated, reflecting more balanced urban-rural development along the north–south dimension driven by accelerated urbanization in southern hilly areas (Figure 4E). Gravity center analysis revealed southward migration for CL, FL, and BL (Figure 4A,B,E), northward migration for GL (Figure 4C), and eastward migration for WB and UL (Figure 4D,F).

3.2. Eco-Environmental Response to Land Use Transformation

3.2.1. Spatio-Temporal Evolution of EEQI

The overall EEQI for the study area was 0.7635, 0.7511, and 0.7472 in 2000, 2010, and 2020, respectively, consistently classified as high level. This indicates generally favorable EEQ conditions across the region. Temporally, EEQI exhibited a slight declining trend over the study period, though the rate of decrease decelerated (Figure 5D). Spatially, grid cells classified as high and highest levels accounted for 19% and 53% of the total area in 2000, while low and lowest levels covered 14% and 2% (Figure 5A). By 2010, the proportion of highest-level grids decreased by 4 percentage points to 49%, and lowest-level grids increased by 1 percentage point to 3% (Figure 5B). In 2020, highest-level grids further declined by 1 percentage point to 48% (Figure 5C). Geographically, EEQI values were notably higher in the southwestern areas and lower in urbanized zones.

3.2.2. Hot Spots of EEQI

The spatial agglomeration dynamics of EEQI exhibited distinct evolutionary patterns: hot spots demonstrated a qualitative transition from 90% confidence expansion to 95% confidence emergence, while cold spots revealed a paradoxical trend of areal contraction with persistent 99%-confidence resistance (Figure 6). Structurally, hot spots underwent optimization as 90%-confidence hot spot area showed fluctuating growth, indicating expanding ecological improvement. Notably, the inaugural emergence of 95%-confidence hot spots in 2020 marked a transition from quantitative to qualitative improvement in core ecological zones like southwestern Zhejiang (Lishui), evidencing successful conservation in key ecological functional areas. Conversely, cold spots displayed entrenched persistence: 99%-confidence cold spots declined minimally (−0.15%), 95%-confidence cold spots decreased by 0.09%. This pattern signifies a deep-seated lock-in effect of ecological degradation in intensively developed regions.

3.3. Driving Factors of EEQI

3.3.1. Relative Importance Analysis

In this study, 70% of the samples were allocated for training the XGBoost model, with the remaining 30% reserved for validation. Automated hyperparameter tuning was performed using GridSearchCV, ultimately identifying the optimal parameter set yielding the highest score (lowest negative mean squared error). The optimized parameters were: learning_rate = 0.05; max_depth = 5; n_estimators = 500. Model accuracy was evaluated by calculating the coefficient of determination (R²) and Root Mean Squared Error (RMSE) for both training and validation sets, where higher R² and lower RMSE values indicate superior performance. As shown in Figure 7, the XGBoost model achieved R² = 0.83 (RMSE = 0.09) on the training set and R² = 0.79 (RMSE = 0.11) on the validation set, demonstrating high predictive accuracy.

The histogram in Figure 8 presents SHAP-based feature rankings, indicating the relative importance of driving factors on EEQI during the study period, where higher SHAP values denote stronger influence. Ranked by mean absolute SHAP value in descending order, the factors are NDVI, GDP, TMP, PD, SLP, ELV, ASP, RD. NDVI emerged as the most influential factor (SHAP = 0.1226), accounting for over 59% of the total impact, followed by per-unit-area GDP (11.25%) and temperature (10.62%). In the accompanying beeswarm plot, each point represents a sample’s SHAP value, with the x-axis indicating effect direction: left of the vertical line signifies negative impacts and right indicates positive impacts. Red/blue coloring denotes high/low feature values, respectively. Distribution patterns reveal factor effects: NDVI and SLP exhibit positive influences (characterized by right-clustered red points and left-clustered blue points), while PD and ASP show negative influences (right-clustered blue points and left-clustered red points). GDP, TMP, ELV, and RD demonstrate complex nonlinear relationships, evidenced by scattered distributions across both SHAP domains.

3.3.2. Partial Dependence Relationship Between Driving Factors and EEQI

To further elucidate the nonlinear relationships between driving factors and EEQI, Figure 9 presents SHAP dependence plots for eight factors, incorporating scatter points, fitted curves, and density histograms. The x-axis represents feature values of driving factors, while the y-axis denotes corresponding SHAP values. Confidence intervals were estimated using a bootstrap method [48], with fitted curves generated through moving window-based locally estimated scatterplot smoothing (LOWESS) regression [49]. Results reveal: ELV exhibits significant nonlinearity, positive effects occur within 154.34–325.99 m and 627.68–1041.70 m ranges, but negative impacts prevail outside these thresholds. Notably, when ELV exceeds 1041.70 m, SHAP values decrease with increasing elevation (Figure 9A). SLP positively influences EEQI between 19.79 and 35.69°, with negative effects below 19.79° or above 35.69° (Figure 9B). ASP shows a negative correlation with EEQI (Figure 9C). TMP demonstrates positive effects within 149.39–183.54 (units: 0.1 °C), transitioning to negative impacts beyond this range (Figure 9D). NDVI maintains a positive correlation overall, though negative effects emerge when NDVI < 0.65, shifting to positive impacts when NDVI > 0.65 (Figure 9E). Per capita GDP and PD consistently suppress EEQI, while RD enhances it (Figure 9F–H).

3.3.3. Interactions of Driving Factors

To elucidate the mechanisms through which driving factors influence EEQI, this study evaluated interaction effects among these factors (Figure 10). Results revealed the strongest interaction between per-unit-area GDP and NDVI, followed by PD-NDVI, TMP-NDVI, and TMP-GDP interactions, indicating NDVI’s pronounced interactive role with other variables. Further investigation through interaction plots (Figure 11) visualized NDVI’s interplay with other factors: the x-axis represents alternate factor values, while the right y-axis shows interaction SHAP values. Point coloration reflects NDVI magnitude (red: high NDVI; blue: low NDVI). Samples were stratified into high-value (NDVI > 0.68) and low-value (NDVI ≤ 0.68) groups based on the NDVI median, with separate LOWESS curves fitted for each (red: high group; blue: low group). Interaction effects demonstrated weak positive correlations with ELV and ASP (Figure 11A,C), suggesting synergistic positive impacts on EEQ. The steepest positive slope occurred when per-unit-area GDP exceeded 1187 with low NDVI values (Figure 11E). Conversely, negative correlations emerged with TMP and PD (Figure 11D,F), particularly pronounced in the low-NDVI/PD interaction. RD-NDVI interaction was negligible (Figure 11G). Notably, SLP-NDVI interactions exhibited divergent trends across groups (Figure 11B): high-NDVI samples showed positive correlation with SLP, while low-NDVI samples displayed negative correlation, with an interaction shift threshold near SLP = 30°. This implies cooperative positive EEQ effects for high NDVI on steep slopes (>30°), but antagonistic effects for low NDVI, underscoring the critical importance of vegetation enhancement in steep terrain for ecological resilience.

4. Discussion

4.1. Interpretation of Results and Theoretical Contributions

This study established a “process-effect-driver” analytical framework, revealing the mechanisms of land use transformation and EEQ responses in China’s southern hilly area under the coupled context of “high vegetation cover baseline + rapid urbanization pressure.” This framework advances research on territorial spatial transformation and EEQ in typical hilly ecosystems. By integrating XGBoost and SHAP methods, we overcame limitations of traditional statistical models in identifying nonlinear relationships, quantifying contribution rates and pathways of EEQI drivers in hilly areas, thereby providing theoretical support for adaptive governance of human-land systems. Specific contributions are threefold:

Deciphering ecological effects of land use transformation in hilly areas. Rapid expansion of BL is the core driver of the slight decline in EEQ, primarily through its encroachment on CL and FL. The spatial heterogeneity of EEQI within the region was significant. The result highlights urbanization’s squeeze on ecological spaces and the success of key ecological zone policies, with hotspots marking a shift to substantive restoration.

Unveiling nonlinear mechanisms of EEQI drivers. This study effectively deciphered complex relationships that are difficult to capture with traditional linear models: NDVI is the primary driver of EEQ in hilly areas, exhibiting a significant positive correlation with EEQI and a threshold effect—specifically, when NDVI > 0.65, the positive influence strengthens. This threshold corresponds to the minimum vegetation coverage required for effective soil conservation, microclimate regulation, and biodiversity maintenance. Socioeconomic factors demonstrated a dual-edged effect: per capita GDP and PD generally suppressed EEQ, yet the impact of GDP showed inflection points across intervals, suggesting potential room for coordination between economic development and ecological protection. The study also revealed pronounced nonlinear characteristics in natural topographic factors, providing precise target areas for ecological restoration on steep slopes.

Quantifying nonlinear interaction effects: The NDVI-GDP interaction showed the strongest contribution, indicating vegetation cover can buffer economic development’s ecological costs. The SLP-NDVI synergy proved critical: when SLP>30°, high NDVI substantially amplified EEQ gains. This discovery deepens understanding of the “topography-vegetation” coupling mechanism and provides a theoretical anchor for soil-water conservation and biodiversity protection in hilly area.

4.2. Comparative Analysis

The mechanisms of EEQ response to territorial spatial transformation and its drivers in China’s southern hilly area, as revealed in this study, engage with existing literature through multidimensional dialog while highlighting regional specificity and theoretical innovation. Key comparative insights are:

This study found that the rapid expansion of BL led to a slight decline in EEQ, with cold spots concentrated predominantly in urban belts. This finding is consistent with research in multiple urbanized regions, which identifies construction land expansion as a core driver of EEQ degradation [29]. Meanwhile, ecological policies were observed to exert a buffering effect: despite the encroachment of BL, hotspot areas increased, confirming the effectiveness of policies such as the Grain for Green Program and the development of ecological functional zones. This aligns with the findings of related studies [50,51], underscoring the significant benefits of ecological restoration initiatives on EEQ. Regarding regional heterogeneity, unlike studies conducted in arid areas [52], this region—characterized by high vegetation coverage—experiences simultaneous urbanization pressure and ecological restoration, presenting a more complex “squeeze-restoration” equilibrium model.

Traditional linear models, such as geographical detector and GWR, often overlook threshold effects and complex interactions among driving factors [53,54]. This study overcame these limitations through the XGBoost-SHAP model, which not only confirmed the dominant role of NDVI in influencing EEQI but also identified a significant threshold effect. This finding is consistent with multiple studies [55]. Previous research has shown that high vegetation coverage enhances ecosystem stability by improving soil-water conservation and carbon sequestration capacity, thereby exerting a positive influence on EEQ [56,57]. Regarding socioeconomic factors, existing studies have discussed the effects of GDP and PD on EEQI, revealing generally negative impacts [58,59,60], which aligns with the results of this study. In contrast, RD was found to have a positive effect on EEQI, which can be attributed to improved infrastructure and management practices in hilly areas over the past two decades. Within the context of Zhejiang’s “Beautiful Countryside” initiative, higher road density is often associated with ecological management, nature-based tourism development, and village renovation projects that include greening and environmental protection measures [61]. This suggests that under specific policy conditions, infrastructure construction can be synergistically aligned with environmental improvement. Moreover, the complex nonlinear characteristics of topographic factors identified in this study have also been observed in recent related research [62].

This study revealed that when SLP > 30°, the positive contribution of high NDVI to EEQI is significantly amplified. This SLP threshold represents a critical gradient where human disturbance diminishes and natural vegetation becomes dominant. It also indicates that vegetation restoration on steep slopes can effectively reduce erosion and enhance ecological resilience, highlighting the importance of slope-vegetation coupling. Related studies have similarly identified a range of 25–69.7° within which SLP exerts improving effects on EEQ, consistent with the findings of this research [63].

4.3. Policy Evolution and Phased Characteristics of EEQ Response

Due to data availability constraints, this study primarily considered natural and socioeconomic drivers of EEQI without quantifying policy impacts. Nevertheless, Zhejiang’s hilly areas underwent profound territorial spatial transformation during 2000–2020, where phased policy orientations significantly shaped land use patterns and EEQI dynamics. Based on documented policy practices, we interpret policy influences on EEQI as follows:

2000–2010: Preliminary balance between gentle slope development and ecological protection. The policy focus centered on resolving conflicts between CL preservation and urbanization through gentle slope development to expand construction space. Specifically, pilot initiatives launched in 2005 promoted industrial park development on gentle slopes (6–15°). While reducing encroachment on plain CL, this triggered substantial FL-to-BL conversion, becoming a primary contributor to EEQI decline. Concurrently, preliminary ecological conservation measures emerged: the 2008 Zhejiang Land Use Master Plan designated ecological source areas in hilly area as “provincial ecological barriers and riverhead protection zones”, enabling counties with high forest coverage (e.g., Longquan, Suichang) in southwestern Zhejiang to maintain stable EEQ.

2010–2020: Policy upgrading under ecological priority. Policies shifted toward ecological value realization and refined governance, arresting EEQI degradation. The “Slope Village-Town” initiative (2015 pilot) adopted a “point-based layout, vertical development” model under the principle of “zero cropland occupation, minimal agricultural land conversion”, pursuing triple objectives of conserving cultivated land, ecology, and development. Simultaneously, deepened forest tenure reforms introduced innovations like forest-right mortgages, promoting forest area expansion and demonstrating the beneficial effects of vegetation cover on steep slope ecological restoration.

Over the past two decades, Zhejiang has implemented policies such as the “Thousand-Village Demonstration and Ten-Thousand-Village Improvement” project to synergistically advance land consolidation, ecological restoration, and rural revitalization. This approach is now being promoted in other regions across China. Zhejiang’s experience demonstrates that integrating ecological priorities into land use planning and policy implementation can effectively mitigate environmental degradation while supporting sustainable development. Practices such as slope-based development controls, forest tenure reforms, and ecological protection redlines offer valuable insights for other hilly areas in southern China and beyond—particularly those undergoing rapid urbanization and ecological transition. Moreover, this model provides a reference for similar regions globally, demonstrating a viable pathway to balance ecological conservation with economic development.

4.4. Limitations and Future Research Directions

This study systematically revealed the spatiotemporal response characteristics of EEQI to territorial spatial transformation and its nonlinear driving mechanisms in China’s southern hilly area, yet several limitations persist: Insufficient integration and quantification of policy drivers obscure direct contributions of policy interventions and their coupling effects with land use transformation. The decadal interval may overlook short-term dynamics and transitional phases, while future scenario projections (e.g., land use evolution under climate change) remain unexplored. Analysis at the 1 km grid scale potentially overlooks finer ecological processes and mechanisms at the parcel level. The composite index approach cannot disentangle pathways and tradeoffs/synergies of land use transformation impacts on specific ecosystem services.

Future research should prioritize: constructing policy indicator systems using text mining and causal analysis to quantify policy intensity and orientation; capturing short-term EEQI fluctuations and simulating future trajectories under diverse scenarios (ecological/economic priority, balanced development) to inform proactive spatial planning; conducting high-resolution case studies in typical watersheds to validate macro-scale mechanisms; integrating EEQI with ecosystem service models (e.g., InVEST) to quantify impacts on service flows and spatial tradeoffs/synergies, enabling precision conservation.

5. Conclusions

This study established an integrated analytical framework to decipher the spatio-temporal evolution and driving factors of EEQ in China’s southern hilly area. The core findings are: (1) Rapid BL expansion, at the expense of CL and FL, was the primary driver of a slight but persistent decline in EEQI, with clear spatial disparities between urbanizing zones (degradation) and ecological conservation areas (improvement). (2) Driving factors exhibited complex nonlinear relationships, with NDVI identified as the dominant positive factor, exhibiting a critical threshold effect (>0.65). (3) Significant factor interactions were identified, most notably the coupling mechanism between slope and vegetation: on steeper slopes (>30°), high vegetation cover substantially amplified ecological benefits.

The application of the XGBoost-SHAP framework proved highly effective in uncovering these complex, nonlinear mechanisms, offering a superior approach compared to traditional linear models for human-land system research. These insights highlight the critical importance of protecting and restoring vegetation on steep slopes and suggest that economic development can be better coordinated with ecological conservation. The findings provide a scientific basis for precise land-use planning and ecological restoration strategies not only in the study area but also in other hilly regions facing similar development pressures.