Gradient Characteristics and Nonlinear Driving Mechanisms of “Production–Living–Ecological” Space Evolution in Mountainous Villages: A Case from Taiji Town, Chongqing

Abstract

The evolution of “Production–Living–Ecological” spaces (PLESs) in mountainous rural areas is shaped by complex interactions between terrain gradients and socio-economic factors. However, existing research lacks a targeted exploration of their evolution and driving mechanisms at the town scale. This study takes Taiji Town in Chongqing, China, as a case study and identifies land use data for mountainous rural areas. Based on this, “Production–Living–Ecological” attributes are assigned to each land use class, terrain gradients are delineated using the Terrain Niche Index, and the gradient-specific characteristics and spatiotemporal distribution patterns of PLES evolution in mountainous rural areas are analyzed. Additionally, the nonlinear driving mechanisms of PLES evolution are explored by incorporating variables such as terrain gradient, geographical location, social development, and ecological landscape. The results show that the evolution of PLES in Taiji Town generally follows a trend of decreasing production space, expanding living space, and steadily increasing ecological space. Furthermore, topographic constraints form a bottleneck in the evolution of production space in mountainous rural areas, with some production space boundaries extending into higher-gradient areas. Analysis of the driving mechanisms reveals that the interactions between land use degree evolution and elevation, as well as between land use degree evolution and slope, are key factors influencing the evolution of PLES, with significant differences across villages with varying topographic conditions. This study provides a scientific basis and methodological reference for observing spatial evolution and optimizing spatial planning at the town scale.

1. Introduction

Land, as a fundamental spatial carrier for human survival and development, contains rich ecological and economic benefits [1]; its evolution is an important reference for the sustainable development of rural production, life, and ecology [2]. Land use research centered on “production–living–ecological” spaces (PLESs) is a crucial foundation for ensuring the high-quality development of rural areas and optimizing the urban–rural landscape [3].

Against the backdrop of economic globalization, territorial spatial planning, and urban–rural integration, rural economies, social development, and spatial patterns are undergoing significant changes, and the basic forms of rural spaces are continuously reshaped [4]. In this context, scholars have conducted studies on the coordinated development of rural PLES at various scales, including municipal [5], regional [6], and county levels [7]. These studies have focused on the identification [8], evolutionary characteristics [9], ecosystem service values [10], coupling coordination [11], and land use simulation [12] of rural PLES. Among these, research on the evolution of PLES indicates that unbalanced development of rural production spaces [13], the decline of ecological functions [14], and the encroachment of living spaces on production and ecological spaces [15] are common issues in current rural PLES. It is suggested that regional development cycle characteristics should be considered in adapting territorial spatial planning [16], and by summarizing the development patterns of different regions, more refined insights can be provided for future PLES construction. However, due to the limitations in the accuracy of historical data, many studies at the scale of towns and villages have relied on methods such as historical landscape characteristics (HLCs) [17], participatory mapping [18], and point data [19] to identify PLES, and have explored evolutionary relationships and PLES evaluation systems. Currently, in the field of rural PLES research, there is a lack of efficient and accurate land use identification methods at the town scale. Moreover, there is an urgent need to develop innovative quantitative methods to explore the specific and targeted evolutionary patterns of PLES at the town scale.

Meanwhile, topography, as the most fundamental geographical factor, serves as a constraint on human activities and ecological selectivity in the distribution of different spatial types in mountainous areas [20]. Therefore, many poverty-stricken areas are associated with mountainous terrain restrictions, and their spatial distribution shows significant imbalance and specificity [21,22]. In this context, scholars have focused on land transfer [23], landscape evaluation [24], rural type assessment [25], spatial reconstruction [26,27], ecosystem health [28], and natural disaster risk research [29] in mountainous villages. The results of these studies indicate that the rate of land use change in mountainous villages decreases with the elevation gradient [30], and that altitude, topography, and socio-economic factors are the most important determinants of local land development rates and regional patterns. As a result, mountainous villages, as geographic spatial units that are highly dependent on natural, geographical, and ecological systems [31], exhibit land use changes characterized by the interaction and gradient evolution of PLES. Among them, research on PLES at the town and village scales presents even more unique challenges [32]. Land use in these areas faces distinct challenges due to fixed limitations such as remote locations, sensitive ecosystems, and marginalization, which contribute to the difficulty of addressing the imbalance in rural development [33]. However, existing research on the nonlinear driving mechanisms of PLES evolution in mountainous areas under complex site conditions is still insufficient. Therefore, investigating the multi-temporal evolution characteristics of rural land use from the perspective of PLES and incorporating gradient effects to explore the evolutionary patterns and driving mechanisms in representative case areas can provide more comprehensive references and methodologies for regional planning, governance strategies and sustainable development.

The aim of this study is to explore the gradient characteristics and nonlinear driving mechanisms of PLES evolution in mountainous villages at the case scale, in order to provide effective methods for connecting territorial spatial planning and promote comprehensive rural revitalization and sustainable development. Specifically, this study addresses the following questions: (1) How to efficiently and accurately establish a recognition pathway for land use data at the town scale in mountainous villages? (2) What are the characteristic gradient patterns in the evolution of the functional attributes of PLES in mountain villages? (3) What are the nonlinear driving mechanisms behind the evolution of PLES in mountain villages?

2. Research Design

2.1. Study Area

The study area is located in Taiji Town, Qianjiang District, in southeastern Chongqing (Figure 1), encompassing seven administrative villages (communities): Taiji Community, Lizi Village, Shicao Village, Jintuan Village, Xinlu Village, Luzi Village, and Taihe Village. Situated within China’s Wuling Mountain contiguous poverty-stricken area, the town has experienced prolonged slow development due to long-standing constraints imposed by its geographic remoteness and mountainous terrain. Designated as a Chongqing municipal “Key Town for Rural Revitalization Support,” Taiji Town serves as a typical and representative case among impoverished mountainous towns in southwestern China. Over recent decades, with the progressive implementation of China’s comprehensive rural revitalization strategy, Taiji Town has achieved significant developmental outcomes, accompanied by pronounced structural transformations in its mountainous rural PLES. Analyzing the evolution of PLES in this context helps uncover the intrinsic linkage between national policy support and rural land use change, thereby providing a scientific basis for optimizing targeted assistance strategies and advancing sustainable rural development in similar impoverished mountainous towns across southwestern China.

2.2. Research Framework

Considering the differential characteristics of the integration of PLES in mountain villages, and based on previous research [34], the PLES of Taiji Town are divided into six spatial attributes: Production Space (PS), Living Space (LS), Ecological Space (ES), Production–Living Space (PLS), Production–Ecological Space (PES), and Living–Ecological Space (LES). From the perspective of the integration of PLES, dynamic research is conducted on five time periods: 1990, 2000, 2010, 2020, and 2024. The core path includes (Figure 2):

(1)

Land Use Identification and Transition Characteristics in Mountainous Villages: By utilizing multi-temporal remote sensing imagery, effective methods for accurately identifying land cover data in the town scale are explored to analyze the characteristics of land use transition in the mountain village.

(2)

Trends and Gradient Characteristics of the Evolution of PLES in Mountain Villages: Production, living, and ecological weights are assigned to different land types, and an “Elevation-Slope” terrain niche index is established. The spatial gradient distribution of the integrated PLES in Taiji Town is explored during the periods 1990–2000, 2000–2010, 2010–2020, and 2020–2024 (hereinafter referred to as Stage 1, Stage 2, Stage 3, and Stage 4). From the three perspectives of “increase–stable–decline”, the gradient evolution characteristics of the PLES under different time periods and gradients are revealed.

(3)

Nonlinear Driving Mechanisms of the Evolution of PLES in Mountain Villages: By combining multi-dimensional variables, the complex nonlinear relationships in the evolution of PLES are analyzed. The driving mechanisms of the integration of these spaces are explored, ultimately providing scientific decision-making basis for the sustainable development of mountain villages.

2.3. Data Source

The data used in this study, obtained through public channels, mainly includes Landsat remote sensing imagery for the year 1990, 2000, and 2010, Sentinel remote sensing imagery for 2020 and 2024, the Normalized Difference Vegetation Index (NDVI) and elevation data (

Table 1. Data source.

Data Source	Data Type	Years	Satellite	Time
Google Earth Engine	Remote Sensing Data	1990, 2000, 2010	Landsat 5	From 1 June to 30 September
		2020, 2024	Sentinel 2
	DEM	-	-	-
	NDVI	1990, 2000, 2010	Landsat 5	From 1 June to 30 September
		2020, 2024	Landsat 8

). To minimize the effects of factors such as seasonal vegetation variation, cloud cover, and climate change on individual imagery data, images with cloud cover lower than 20% were selected. At the same time, median compositing was applied to all imagery data from June to September for each year to ensure the quality of the selected data. The elevation (DEM) data were sourced from NASA’s SRTM elevation dataset, with a resolution of 30 m. The NDVI data is calculated based on remote sensing imagery from Landsat 5/8, with a resolution of 30 m. All of the above data were accessed via Google Earth Engine (GEE).

3. Methods

3.1. Methods for Exploring Land Use Identification and Transition Characteristics in Mountainous Villages

3.1.1. Land Cover Classification Based on Multi-Feature Assisted Random Forest

To address the issue of insufficient land use cover data in mountain village areas, the Random Forest algorithm was employed to identify land cover data for the mountain villages. By combining Landsat and Sentinel-2 imagery, the NDVI, Normalized Difference Water Index (NDWI), and Normalized Difference Built-up Index (NDBI) were calculated to classify land use types in the study area. Referring to the “Planning of Taiji Town and Settlement Construction in Qianjiang District (2016–2030)”, the main land use types of Taiji Town were summarized. Considering the resolution of remote sensing image grids, five land use categories were identified, including cropland, forest, water, residential land, and bare land. Training sample data were extracted from vector data annotated using ArcGIS Pro 3.4.2, and a Random Forest model consisting of 100 decision trees was constructed. The dataset was divided into 70% training samples and 30% test samples, with stratified sampling to ensure that the proportions of categories were consistent across both training and testing sets. After training the model, the overall accuracy and Kappa coefficient were used for evaluation (

Table 2. Classification Result Validation.

Year	Overall Accuracy	Kappa Coefficient
1990	0.935	0.917
2000	0.897	0.854
2010	0.929	0.908
2020	0.950	0.936
2024	0.949	0.934

), and the results were excellent. Considering data consistency and alignment, the classification results were standardized to a 30 m resolution and subjected to noise reduction before subsequent analysis.

3.1.2. Participatory Mapping Correction

Considering the details and accuracy of land use cover data identified through remote sensing imagery at the town scale, this study combines remote sensing imagery with visual interpretation corrections for the five identified land use types. Additionally, participatory mapping is employed, based on local residents’ knowledge and experience, for geographic information collection, and to supplement the collection of socio-economic development data for the village. From July 2024 to March 2025, three rounds of land use cover data correction and data collection were conducted.

In July 2024, a field survey was carried out to initially validate and correct the identification results, extracting major roads and rivers within the town.

From December 2024 to February 2025, a series of guest interviews were conducted with local government staff and villagers. A total of 10 individuals participated in the interviews, including one planner, two landscape designers, two local government officials, and five villagers. During this period, satellite imagery and GIS tools were used to present the preliminary classification results to the local government, combining their practical knowledge of land use and land cover types. They recalled land use changes over different periods and identified other key land types in the town, further refining the classification results. The refinement included secondary categories for forest land, such as “plantation forests,” secondary categories for water bodies, such as “ponds,” and the addition of “industrial, mining, and storage land,” “public administration and service land,” as well as “parks and green spaces.”

From February to May 2025, the PLES in Taiji Town were delineated and assigned functional scores (

Table 3. PLES Scoring for Taiji Town.

Types	Land Use Types	Description	Production Attribute	Living Attribute	Ecological Attribute
PS	Industrial, Mining, and Storage Land	Includes industrial land, mining land, storage land, etc.	5	0	0
ES	Forest	Includes tree forests, bamboo forests, shrub forests, other types of forests, etc.	0	0	5
	Water	Includes river water surfaces, pond water surfaces, ditches, wetlands, etc.	0	0	5
	Bare Land	Includes idle land, agricultural land for facilities, field ridges, sandy land, bare land, bare rocky ground, etc.	0	0	5
PES	Cropland	Includes paddy fields, irrigated fields, dry fields, etc.	5	0	3
	Plantation Forests	Includes forests planted on non-forested land or logged areas through artificial methods.	5	0	3
	Ponds	Includes facilities for water collection and irrigation on flat land, a combination of ditches and ponds.	3	0	1
PLS	Public Administration and Service Land	Includes educational land, healthcare land, social welfare land, cultural facility land, sports land, public utility land, etc.	1	3	0
	Residential Land	Includes rural residential land, etc.	3	5	0
ELS	Parks and Green Spaces	Includes parks, community gardens, squares, and green spaces used for recreation, beautification, and protection within village areas.	0	3	1

). Referring to the PLES identification and evaluation methods established by previous studies [35,36,37], existing research generally recognizes that land has multifunctional attributes, meaning that a single land use type can serve multiple functions, though these functions may vary in terms of their primary and secondary importance. Based on these studies and in consideration of the land use characteristics in Taiji Town, we used a graded scoring method to quantitatively assign values to the PLES functional attributes of each land type, with scores of 1, 3, and 5 representing weak, moderate, and strong functional intensities, respectively.

3.1.3. Land Use Transition Matrix

The changes in land use types over different periods are commonly measured using a Land Use Transition Matrix (LUTM), which generates information on the dynamic processes of mutual conversion between various land use categories at the beginning and end of specific time periods in a particular area [38]. Based on the previously identified and corrected land use cover data, the land use transition matrix for stages 1–4 of Taiji Town is calculated to display the land changes in Taiji Town. The generalized representation of the land use transition matrix is shown as follows [39]:

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

(1)

where S represents the area, n represents the number of land use types before and after the change, and i and j (i, j = 1, 2, …, n) represent the land use type before and after the change, respectively. S_ij represents the area of land type i before the change that is converted into land type j.

3.2. Methods for Evolution Trends and Gradient Characteristics of PLES Attributes

3.2.1. “Elevation-Slope” Terrain Niche Index

The terrain conditions that constrain land use distribution are often manifested through the combined effects of elevation and slope. The “Elevation-Slope” Terrain Niche Index (TNI) [40] is established using these two factors, and the terrain of Taiji Town is divided into five gradients using the natural break method, which is then used to statistically analyze the distribution of PLES across these gradients and for subsequent research. As the T value increases, the elevation and slope become steeper, and vice versa. Areas with high elevation and gentle slopes, or low elevation and steep slopes, generally obtain intermediate T values [41]. The formula for the TNI is expressed as follows:

$$TNI=\log \left[\left({\displaystyle \frac{E}{\overline{E}}}+1\right)\cdot \left({\displaystyle \frac{S}{\overline{S}}}+1\right)\right]$$

(2)

where E is the elevation of the pixel; $\overline{E}$ is the average elevation in the study area; S is the slope of the pixel; $\overline{S}$ is the average slope in the study area.

3.2.2. Multi-Temporal Observation of PLES Attribute Evolution

Based on multi-temporal PLES land use raster data from 1990 to 2024 and the terrain gradient of Taiji Town, the raster data for different years are compared pairwise to define three main types of PLES change (

Table 4. Calculation logic of PLES attribute evolution.

Starting Value	Ending Value	Assigned Value	Meaning
0	1, 3, 5	3	Increase (+)
1	3, 5
3	5
1, 3, 5	Same Value	2	Stable (=)
5	0/1/3	1	Decline (−)
3	0, 1
1	0
0	0	0	None

Note: In subsequent studies, the symbols −, =, and + will represent Decline, Stable, and Increase, respectively.

): (1) Increase—Refers to the transition from other types of space or lower to medium intensity production space to a space with stronger production attributes. (2) Stable—Refers to the unchanged land use between different years. (3) Decline—Refers to the transition from stronger attributes to weaker spatial attributes. Through this approach, combined with the policies from each period, the study further analyzes the more subtle dynamic evolution patterns of PLES under different gradient effects in Taiji Town during different time periods, and summarizes the evolution patterns of mountain village PLES driven by policy tendencies in different stages. These calculations were primarily implemented using Python 3.12.7 programming, with libraries such as Rasterio and NumPy.

3.3. Methods for Nonlinear Driving Mechanisms of PLES Evolution

3.3.1. XGBoost Model

XGBoost is a tree-based ensemble learning algorithm [42], known for its computational efficiency and predictive accuracy. It enhancing test set performance by iteratively optimizing training errors using a gradient boosting framework. It is widely used in machine learning applications [43]. This method optimizes solely based on data values, without the need for a predefined loss function form, effectively decoupling the selection of the loss function from the model optimization process [44]. It is also adept at determining the ranking of influential factors. The objective function formula is [45]:

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

(3)

In the equations above (u_i, v_i) is the spatial position of point i; m is the number of independent variables; y_i is the dependent variable; x_ik (k = 1, 2, …, m) is the independent variable; ε_i is a random variable; β₀ (u_i, v_i) is the intercept at point i; β_k (u_i, v_i) is the regression coefficient. If the regression coefficient is greater than 0, it indicates a synergistic relationship, and vice versa.

Three models were set up in the study, each aimed at predicting the potential impacts on the evolution of PLES based on production, living, and ecological spaces as target variables.

Regarding feature variables, the natural environment, landscape patterns, and social development together shape the structure and development trajectory of the human-environment relationship system [46], influencing changes in PLES. Considering data availability and representativeness, an indicator system with 12 potential influencing factors was established based on previous research (

Table 5. Description of the selection of target and feature variables.

Type	Dimensions	Variable Name	Resolution	Unit	Quantitative Method/Source
Target Variables	PLES Evolution	Production Space Evolution (PSE)	30 m × 30 m	-	The overall trend of the five years was calculated based on the Mann–Kendall trend test for various types of PLES land use.
		Living Space Evolution (LSE)		-
		Ecological Space Evolution (ESE)		-
Feature Variables	Terrain Gradient	Elevation (DEM)	30 m × 30 m	m	DEM data was processed using the ArcGIS Pro raster statistics tool.
		Slope		°	Processed and generated using the ArcGIS Pro slope tool.
		Terrain Niche Index (TNI)		-	Calculate the TNI based on ‘elevation-slope’.
	Geographical location	Distance to Main Roads (DTR) 1		m	Euclidean distance was calculated based on self-collected data.
		Distance to Water (DTW)		m
	Social Development	Land Use Degree Evolution (LUDE)		-	The overall LUD trend of the five years (1990, 2000, 2010, 2020, and 2024) was calculated using land use data identified in the previous phase, and the Mann–Kendall trend test was applied to assess the overall trend.
	Ecological landscape	Patch Density Evolution (PDE)		-	The moving window method in Fragstats 4.3 was used to process the PLES land use data for the years 1990, 2000, 2010, 2020, and 2024, and the Mann–Kendall trend test was applied to calculate the overall trend across these five years.
		Aggregation Index Evolution (AIE)		-
		Shannon’s diversity Index evolution (SHDIE)		-
		Normalized Difference Vegetation Index Evolution (NDVIE)		%	NDVI for the years 1990, 2000, 2010, 2020, and 2024 was calculated using Landsat 5/8 satellite data bands, and the overall trend across these years was assessed using the Mann–Kendall trend test.

¹ Regarding the distance to roads, the 30 m resolution remote sensing imagery in the past was not suitable for road extraction. Through field surveys, it was found that road changes in Taiji Town were primarily concentrated within the village settlement areas, while the county and township roads, as well as the village access roads, had not undergone significant changes. Thus, the study chose to map the distances to the county and township roads, as well as the village access roads, to calculate the distance to transportation networks.

) [47]:

(1)

Terrain gradient indicators: Under natural conditions, terrain gradient-related factors such as elevation (DEM), slope, and TNI influence the distribution of PLES.

(2)

Geographical location indicators: The distance to main roads and Water (DTR, DTW) is a key determinant of regional population concentration and industrial distribution.

(3)

Social development indicator: In economics, changes in development levels and population size often alter land use degree (LUD), thereby affecting land use efficiency [48]. Therefore, land use degree and human activity intensity are used to represent socio-economic factors [49]. However, based on previous studies on LUD classification and calculation methods [50], the land use level at the pixel level cannot be accurately or truly reflected in mountainous villages, where forest and cultivated land account for the majority of the area. Therefore, based on previous research, this study develops a method for calculating LUD at the pixel level in mountainous villages using a moving window approach (Supplementary Materials S1). This method accurately reflects the land use degree evolution (LUDE) at the pixel level. The choice of window size is based on standards from relevant studies [51]. Using 30 m land use data, three different window sizes—90 m (3 × 3), 150 m (5 × 5), and 210 m (7 × 7)—are compared. It is ultimately concluded that the 150 m window size best reflects the LUD of Taiji Town.

The formula for calculating LUD is:

$$k=\sum _{i=1}^n{A}_i\times C_i$$

(4)

In the formula, k represents the comprehensive land use index, where k ∈ [1,2,3,4]; A_i is the land use classification index for level i (1 point: “bare land”; 2 points: “forest”, “water”; 3 points: “cropland”, “plantation forests”, “ponds”; 4 points: “residential land”, “industrial, mining, and storage land”, “Public Administration and Service Land”, “parks and green spaces”.); C_i is the percentage of the land area of level i land use classification within each window; n is the total number of land use classifications.

(4)

Ecological landscape indicators: Ecological landscapes, which provide various ecosystem services and cultural benefits, influence the structure and function of ecosystems [52]. Based on previous research on land use under gradient effects [53], Normalized Difference Vegetation Index Evolution (NDVIE) along with three landscape pattern indicators—Patch Density Evolution (PDE), Aggregation Index Evolution (AIE), and Shannon’s Diversity Index Evolution (SHDIE)—were selected to represent the landscape ecological factors of PLES.

Considering the limitations of land use in mountainous areas and the complex influences of gradient factors such as slope and elevation, interaction terms (LUDE×SLOPE, LUDE×DEM) were additionally introduced to more accurately simulate the evolution of PLES under different factors.

The study uses Python 3.12.4 and the scikit-learn library for model training and evaluation. The three models for the evolution of PLES only consider the actual range of evolution for data processing. The dataset is divided into a training set (60%), a validation set (20%), and a test set (20%). Model performance is evaluated using R², RMSE, and MAE (

Table 6. Model performance evaluation.

Type	Training Set	Validation Set	Test Set
PES	R2: 0.519	R2: 0.461	R2: 0.455
	RMSE: 0.433	RMSE: 0.460	RMSE: 0.454
	MAE: 0.326	MAE: 0.345	MAE: 0.342
LES	R2: 0.533	R2: 0.490	R2: 0.496
	RMSE: 0.473	RMSE: 0.490	RMSE: 0.494
	MAE: 0.383	MAE: 0.391	MAE: 0.401
ESE	R2: 0.492	R2: 0.441	R2: 0.443
	RMSE: 0.391	RMSE: 0.412	RMSE: 0.413
	MAE: 0.291	MAE: 0.305	MAE: 0.306

).

3.3.2. SHAP and GeoShapley Combined Explanation

SHAP (SHapley Additive exPlanations) is a tool used for explaining model predictions. Based on the Shapley values from game theory, it decomposes predictions into local and global feature contributions that are interpretable. When combined with XGBoost, SHAP can effectively estimate Shapley values, revealing the local variation in feature variables’ impact on the target scalar and identifying key turning points. However, traditional SHAP methods do not explicitly consider spatial dependencies, limiting their application in research where spatial context is critical.

GeoShapley is an extended SHAP method that integrates features from X and Y into a unified geographical spatial variable and quantifies the interaction (spatial effect) between this variable (GEO), and all the other feature variables, specifically designed to address the limitations of traditional SHAP methods [54]. It provides local interpretability and deeper insights into spatial heterogeneity. These methods are implemented using Python’s SHAP and GeoShapley libraries. The formula for GeoShapley value calculation is shown below [55]:

$${\varphi }_i=\sum _{S\subseteq F\setminus \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\right|F|-|S|-1)!}{\left|F\right|!}}[f(S\cup \{i\})-f(S)]$$

(5)

In the formula, ϕ_i represents the GeoSHapley value for feature i; S is the subset of features that does not include i; F is the complete set of features; |S| is the size of the subset S; |F| is the size of the feature set F; f(S) is the model prediction using only the features in the subset S; f (S ∪ {i}) is the prediction with the feature iii added to the subset S.

The study combines SHAP, GeoShapley, and XGBoost to enhance the interpretability of the modeling results and strengthen the analysis of spatial dependencies. In addition, to better understand the spatial heterogeneity of intrinsic location effects, we visualize the SVC values, quantifying the magnitude and specific location of each influencing factor on the evolution of PLES dimensions.

4. Results

4.1. Land Use Identification Results and Transition Characteristics in Mountainous Villages

Under the constraints of complex site conditions, forest and cropland are the primary land use types in the villages of Taiji Town (

Table 7. Land use area of Taiji Town by year.

Land Class	1990 (ha)	2000 (ha)	2010 (ha)	2020 (ha)	2024 (ha)
Cropland	2929.5	2119.23	1959.03	1669.41	1377.72
Forest	2669.76	3954.15	4049.1	4274.55	4490.37
Plantation Forests	0	0	0	3.24	0.99
Water	113.31	86.58	88.83	87.3	85.77
Ponds	0	0	0	6.84	5.4
Residential Land	57.6	87.48	160.92	209.34	253.44
Industrial, Mining, and Storage Land	0	0.18	3.87	9.99	22.05
Public Administration and Service Land	2.7	3.51	3.6	4.14	4.86
Parks and Green Spaces	0	0	0	0.45	0.72
Bare Land	496.89	18.63	4.41	4.5	28.44

). Consequently, from 1990 to 2024, land use changes in Taiji Town were largely centered around the conversion of these two land types. Overall trends (Figure 3) show that cropland was predominantly converted into forest land: between 1990 and 2024, cropland decreased from 2929.5 ha to 1377.72 ha, with an average annual decrease of approximately 45.67 ha. In contrast, forest land increased by a factor of 1.68, from 2669.76 ha to 4490.37 ha by 2024. This demonstrates that the two rounds of the Grain-for-Green program implemented in China have had significant benefits [56], playing a remarkable role in promoting the ecological civilization construction in Taiji Town, Chongqing, which is located in the Wuling Mountain ecological barrier area.

Simultaneously, residential land expanded at a rapid rate of 4.4 times, growing from 57.60 ha to 253.44 ha. This expansion was primarily driven by the conversion of cropland, with cropland converting to residential land in each phase as follows: 39.69 ha, 93.51 ha, 83.61 ha, and 59.22 ha, respectively. In terms of land use types, Taiji Town’s land use structure and hierarchy diversified year by year. Modern agricultural and industrial land uses, such as industrial, mining, and storage land, as well as ponds, increased progressively. For instance, industrial, mining, and storage land grew from 0.18 ha in 2000 to 22.05 ha in 2024, a 122.5-fold increase, although it still accounted for a small proportion of the total land area. After 2020, with the implementation of various policies, rural infrastructure development and public services became increasingly prioritized. As a result, urban-like land uses such as public administration and service land, as well as parks and green spaces, began to emerge in rural areas.

4.2. Trends and Gradient Characteristics of the Evolution of PLES in Mountain Villages

4.2.1. Trends of the Evolution of PLES in Mountainous Villages

Based on the previously assigned PLES attribute scores, further exploration of the overall distribution characteristics of the PLES in Taiji Town was conducted (Figure 4a). Regarding PS, the proportion of land with strong production attributes gradually decreased, while the areas with moderate and weak production attributes progressively increased. For LS, land use continued to expand, showing a distributed expansion trend. Unlike the “first concentrate, then fill” logic commonly observed in flatland areas, Taiji Town expanded incrementally along its original edges, with fewer new patches, and the expansion rate slowed after 2010. In ES, areas with ecological functions occupied a high proportion, and over time, there was a noticeable trend of a decrease in areas with moderate ecological attributes, while areas with strong ecological attributes gradually increased (Figure 4b).

4.2.2. Gradient Characteristics of the Evolution of PLES in Mountainous Villages

Further analysis of the changes in the functional attributes of PLES over four stages reveals the evolutionary patterns of PS, LS, and ES in different gradients of Taiji Town (Figure 5).

From the area proportion of each stage, in Stage 1, PS shows a significant “Decline” (1384.56 ha) followed by a smaller “Increase” (574.29 ha). In contrast, ES exhibits a strong growth differential, with “Decline” (605.34 ha) and “Increase” (1353.69 ha). After Stage 2, the amplitude of increase and decrease gradually narrows, and LS enhancement becomes increasingly significant.

From the spatial distribution across gradients, PS is mainly concentrated in gradients 1–3, where strong PS decreases annually, while moderate PS increases annually. In gradient 5, PS initially decreases and then increases, suggesting that cropland in low-gradient areas decreases and gradually expands to higher-gradient areas. LS is mainly concentrated in gradients 1–3, but the proportion of LS in each gradient shows a year-on-year increase. ES has a relatively high proportion across all gradients, with gradients 1–4 showing annual increases in ES, and the increase in area clearly outweighing the decrease. In gradient 5, ES shows relatively small fluctuations across stages, but from Stage 2, the overall trend shows a greater decline than increase, indicating an overall ES decline (Figure 6).

4.3. Nonlinear Driving Mechanisms of the Evolution of PLES in Mountainous Villages

First, a comparison was made prior to and after the introduction of interaction terms in the spatial regression prediction model (Figure 7). Before the introduction of interaction terms, the contributions of the feature variables in the evolution of PLES were highly differentiated, with the driving force mainly relying on the contribution of the key variable, LUDE. After the introduction of interaction terms, the interaction effects between gradient factors (such as DEM and SLOPE) and LUDE were significantly enhanced, along with the interaction effects involving geographic location (GEO). The preliminary conclusion is that the evolution of PLES in mountain villages is a complex process closely related to terrain, ecological carrying capacity, and human activities. A summary plot produced using GeoShapley was used to provide an overall interpretation of the mountain village PLES evolution model, revealing the contributions of each feature variable to the model’s predictions (Figure 8).

Based on the variable types and the sequence of the bee Summary plot, the discussion will proceed in the following order: single feature variables, followed by manually crossed variables, and finally GEO interaction variables.

4.3.1. Relationship Between Single Feature Variables and PLES Evolution

PSE: The top three features contributing to PSE are LUDE, AIE, and PDE. Figure 9a shows the relationship between these three indicators and PS evolution. LUDE’s influence on PS follows a nonlinear pattern of “suppression–promotion–suppression.” In the model predictions, maintaining a moderately stable LUD helps enhance PS, while excessive decrease or increase leads to a suppressive effect. In contrast, AIE and PDE exhibit a generally linear trend, with their enhancement showing a significant negative correlation with PS growth.

LSE: The top three features contributing to LS evolution are SHDIE, LUDE, and DTR. Figure 9b shows the relationship between these three indicators and LSE. SHDIE and LUDE are linearly positively correlated with the probability of LS enhancement, meaning that higher values increase the likelihood of LS enhancement in the model predictions. DTR exhibits a nonlinear relationship: when DTR is small (<50 m), it has a significant positive contribution to LS enhancement; in the 50–250 m range, it shows a primarily negative contribution; and when DTR > 250 m, the sample size is limited, and the overall effect is positive.

ESE: The top three features contributing to ESE are LUDE, SHDIE, and AIE. Figure 9c shows the relationship between these three indicators and ES evolution. LUDE and AIE show an overall linear trend: higher LUD increases the likelihood of ES decline, while increasing AI helps enhance ES. Areas where SHDI excessively declines or increases tend to suppress ES enhancement, with the most significant positive effect occurring when SHDI is close to stable.

4.3.2. Interaction of Manually Crossed Variables and PLES Evolution

The interaction relationships of LUDE × DEM and LUDE × SLOPE were observed from the SHAP dependence plots (Figure 10).

PSE: The interactions between LUDE and DEM (LUDE × DEM) and LUDE and slope (LUDE × SLOPE) rank 1st and 3rd, respectively (Figure 10a). In LUDE × DEM, the prediction of PS is primarily concentrated in the 500–900 m elevation range. Within the 500–600 m range, the overall effect is negative, where an increase in LUD significantly suppresses the expansion of PS. Notably, in areas where LUD decreases, the expansion of PS is promoted. In the 600–900 m range, the overall effect is positive, and an increase in LUD promotes the expansion of PS. For LUDE × SLOPE, the prediction of PS mainly concentrates in areas with slopes less than 35°, and the overall effect is negative. However, in high-gradient areas with slopes greater than 35°, despite the smaller sample size, there is a positive contribution to the prediction of PS, with a positive correlation with the increase in LUD. The observed causal relationship may be directly linked to the terraced agricultural production systems in mountainous rural areas (Figure 1). In these regions, areas at lower elevations with gentler slopes are increasingly being converted to residential use, which may constrain the development of PS. Moreover, given the extreme scarcity of arable land and the natural constraints of steep terrain, the conversion of unsuitable sloped land into stable and usable terraced fields—while meeting production needs, reducing slope runoff, and preventing soil erosion—represents a highly adaptive and intelligent land use strategy in southwestern mountainous villages.

LSE: The interaction between LUDE and DEM (LUDE × DEM) and LUDE and slope (LUDE × SLOPE) rank 1st and 4th, respectively (Figure 10b). In LUDE × DEM, LS is mainly concentrated in the 500–800 m elevation range. In the 500–550 m range, the increase in LUD has a significant positive relationship with LS expansion. In the 550–600 m range, the increase in elevation suppresses LS enhancement. Above 700 m, the SHAP values fluctuate around 0, indicating a weak correlation and a weakening relationship with LUDE. In LUDE × SLOPE, LS is mainly concentrated in areas with slopes less than 30°, where the increase in LUD significantly promotes LS expansion. In high-slope regions (greater than 30°), the increase in LUD has a suppressive effect on LS expansion.

ESE: The interaction between LUDE and DEM (LUDE × DEM) and LUDE and slope (LUDE × SLOPE) rank 1st and 4th, respectively (Figure 10c). In LUDE × DEM, ES is mainly concentrated in the 500–1000 m elevation range. In the 500–600 m range, elevation has a significant suppressive effect on ES enhancement, and reduced LUD promotes the decline of ES in low-elevation areas. From 600 to 800 m, the effect of elevation on ES enhancement shifts from weak to positive, and areas with reduced LUD promote the prediction probability of ES. Above 800 m, the total amount of ES decreases, and areas with enhanced LUD promote ES enhancement. In LUDE × SLOPE, the increase in LUD in areas with slopes below 10° promotes ES enhancement. Between 20 and 40° slopes, LUD has a weak moderating effect, while in areas with slopes greater than 40°, the effect is suppressive, and the relationship with LUD is weak.

4.3.3. Relationship Between GEO Interaction Variables and PLES Evolution

This section further analyzes the GEO interaction variables. As shown in Figure 11, geographic location demonstrates significant heterogeneity. Regarding the interactions between geographic location and other factors, the analysis particularly focuses on three key spatial interaction features that are most relevant to PLES.

PSE: Figure 11a presents the top three interacting features related to the evolution of PS: LUDE × SLOPE × GEO, SLOPE × GEO, and LUDE × DEM × GEO. LUDE × SLOPE × GEO is highly correlated with the enhancement of PS, primarily concentrated in the eastern part of the town, including Lizi Village, Shicao Village, Jintuan Village, and Xinlu Village. Compared to other villages, these areas have more flat terrain, making land development easier. Therefore, the areas with a promoting effect are mainly located in regions with higher slopes and greater distances from the village centers. SLOPE × GEO is also highly correlated with the enhancement of PS, especially in the higher slope areas of Shicao Village, Jintuan Village, and Xinlu Village. By comparing these two factors, it is proved that the intervention of LUDE leads to the influence of slope on the evolution of PS, which forms certain spatial heterogeneity due to differences in LUDE across different villages. LUDE × DEM × GEO is highly correlated with the enhancement of PS, primarily in Shicao Village, Jintuan Village, Xinlu Village, and Taihe Village.

LSE: Figure 11b presents the top three interacting features related to the evolution of LS: DEM × GEO, LUDE × DEM × GEO, and DTR × GEO. DEM × GEO is highly correlated with the enhancement of LS, primarily distributed across Taiji Community, Lizi Village, and Taihe Village, with a relatively scattered distribution. LUDE × DEM × GEO is highly correlated with the enhancement of LS, particularly in Taiji Community, Lizi Village, and Taihe Village, and is significantly associated with the increase in LUD around the center of each village (or community). DTR × GEO is highly correlated with the enhancement of LS, particularly near the village roads in Lizi Village and Taiji Community. In summary, the first two interactions show that, under varying altitudes across different villages, altitude affects the evolution of LS differently. However, with the intervention of LUDE, the enhancement of LS shifts towards the village and town center areas with better public service provision.

ESE: Figure 11c presents the top three interacting features related to the evolution of ES: SHDIE × GEO, DEM × GEO, and LUDE × SLOPE × GEO. SHDI × GEO and DEM × GEO are distributed relatively evenly across the villages. SHDIE × GEO corresponds with areas where SHDI is decreasing, while DEM×GEO is primarily concentrated in villages with a higher proportion of relatively high-altitude areas in the town, such as Luzhi Village and Xinlu Village. Overall, the interaction of SHDI and DEM, with respect to geographic location, has a more significant impact on ES, and the distribution of these effects is more balanced across the villages. LUDE × SLOPE × GEO is also highly correlated with the enhancement of ES, particularly in the high-slope areas of Taihe Village and Xinlu Village.

5. Discussion

5.1. Comparison with Other Related Studies

Although previous studies have concluded and widely acknowledged that topography is a key factor influencing rural spatial distribution and land use structure, with research focusing on how the spatial patterns of mountainous villages are influenced by factors such as elevation, slope, distance to main roads, and indirect effects of socio-economic, environmental, and infrastructural factors [57,58], especially in areas with intense human activity [59], the correlation between topographic gradient variation and spatial patterns in the PLES space has been explored, including scale, shape, distribution, and structure characteristics [60]. However, challenges such as the difficulty of obtaining rural data and insufficient granularity have led to research gaps, and systematic studies on the driving mechanisms of PLES evolution at the town scale remain underdeveloped. This is particularly the case in mountainous villages, where the complex and dynamic changes pose difficulties in rural spatial planning and in formulating village development strategies.

We believe that to conduct localized and fine-grained land spatial planning, it is essential to develop corresponding research tools that allow for a more specific and detailed examination of the evolution of PLES in mountainous villages. This study approaches the topic from the perspective of the integration of production, life, and ecology. It provides a specific case study of land use transformation from 1990 to 2024, simulating the gradient evolution relationships from the three perspectives of “increase, stable, and decline,” and examining the nonlinear driving mechanisms.

5.2. Ecological Gains, Residential Expansion, and Diversified Production Are the Development Trends of the PLES in Mountainous Villages

Since 2002, with the implementation of policies such as “Grain for Green” and “Mountain Closure for Forest Regeneration,” the cultivated land in Taiji Town has decreased by 2.13 times between 1990 and 2024, while forest land has increased by 1.68 times. This profound change reflects the urgent demand for ecological restoration and soil and water conservation in the process of rural revitalization in China. Based on this conclusion and the related data provided by the town government during the field research, it is evident that, despite the continuous reduction in cultivated land and the increasing diversification of land use types and levels, such as the growing use of industrial and agricultural land, the number of town enterprises has increased significantly, without a noticeable decrease in total grain production. This demonstrates that, under the constraints of mountain terrain, the introduction of modern agricultural technologies and adjustments in land use structure are essential pathways for enhancing agricultural productivity and promoting diversified industrial development. These measures play a critical role in the poverty alleviation efforts of mountain villages.

However, based on the statistical data collected during the survey (

Table 8. Socio-economic Development Statistics of Taiji Town.

Year	1990	2000	2010	2020	2024
Total Population (persons)	12,806	12,723	12,808	12,828	12,674

Note: Collected from interviews with the Taiji Town Government.

), it is worth noting that, despite the rapid expansion of residential land (which increased 4.4 times), the total population of the village has not significantly increased. On one hand, this reflects the growing mobility of the rural population, and the expansion of residential land serves to improve the living standards and infrastructure for rural residents, thereby enhancing the living conditions in mountain villages. On the other hand, under the constraints of the homestead system, there may be an oversupply of construction land, with some homesteads far exceeding the actual needs of the resident population, resulting in a situation where there is “a small population but large land occupancy.” Therefore, how to improve land use efficiency while ensuring the quality of life for residents has become an urgent issue. This deserves more attention from policymakers and planners in future planning and management.

5.3. The Upward Shift in the PS Boundary Caused by Topographic Constraints Constitutes a Bottleneck in the Spatial Evolution of Mountainous Villages

Through an analysis of the functional attributes of PLES in Taiji Town from 1990 to 2024, this study reveals the evolutionary patterns of functional transformation in mountainous rural areas. Based on the changes in PLES attributes across different stages, the following major trends are summarized:

PS: The proportion of areas with strong production attributes has gradually decreased over time, while those with medium and weak production attributes have steadily increased. The spatial distribution of PS across terrain gradients has become increasingly diversified. Cropland in low-gradient areas has continuously declined, and PS has progressively shifted toward higher-gradient zones.

LS: The proportion of LS across all gradients has shown continuous growth. Unlike the “aggregation–infill” pattern commonly observed in plains [61], LS in this mountainous area has expanded gradually along existing settlement edges, with few new patches emerging. Although the LS in all gradients has increased year by year, areas with favorable topographic conditions (Gradients 1–3) remain the primary zones of rural LS.

ES: The distribution of ES in Taiji Town shows strong regional and temporal characteristics. Under the overall trend of continuous ecological gain, land with medium ecological attributes has decreased annually, while land with strong ecological attributes has increased. Overall, the trend of ES expansion is more pronounced. Particularly in gradients 1 to 4, ES has increased year by year, with the growth rate in most gradients significantly exceeding the rate of decline. However, in the more restrictive Gradient 5, ES has experienced a net decrease since 2000, with degradation areas outweighing the expansion areas.

These findings suggest that topographic factors, as key constraints on human utilization and transformation of land resources in natural environments, play a crucial role in shaping and evolving rural settlement patterns and multifunctionality. Under the combined influence of terrain conditions, higher-quality land in Taiji Town has been gradually allocated to LS and PS. However, with the passage of time, land dominated by production functions has progressively shifted upward along terrain gradients, forming differentiated combinations of topographic levels. In future rural planning and spatial layout, it is essential to comprehensively consider site conditions, the interrelationships among spatial functions, and differences in environmental carrying capacity. Only by addressing the potential ecological risks and resource mismatches arising from gradient-based PLES carrying capacities can the sustainable development of PLES be ensured.

5.4. The Interaction Between LUDE and Elevation/Slope Is a Core Driving Factor in the Evolution of PLES in Mountainous Villages

In the final part of the study, a nonlinear regression prediction of the evolution of PLES in mountain villages was conducted based on the XGBoost model. The study further explored the impact of various feature variables and their interactions through the use of interaction variables (LUDE × DEM, LUDE × SLOPE). The following key conclusions were drawn:

Single Feature Variables: PSE: Significantly influenced by social development and ecological landscape-related indicators. Both excessive reduction and enhancement of LUD may increase the likelihood of suppression in PS evolution. Areas with higher landscape aggregation and patch density tend to suppress the growth of PS. LSE: Significantly influenced by ecological landscape, social development factors, and geographical location-related indicators. As SHDI and LUD increase, the probability of enhancing LS also increases. When the distance to the main road is greater than 50 m, the probability of enhancement outweighs suppression; between 50 and 250 m, suppression is greater than enhancement; and above 250 m, although the sample size is small, the overall trend still shows enhancement. ESE: Significantly influenced by social development and ecological landscape-related indicators. Areas with increased LUD and AI promote the enhancement of ES. However, regions with excessive reduction or enhancement of SHDI suppress ES enhancement, with a relatively pronounced promoting effect observed when SHDI is close to its stable state.

Interaction Effects: The most significant overall impact on PLES comes from the interactions between land use degree evolution and elevation (LUDE × DEM), as well as land use degree evolution and slope (LUDE × SLOPE), with LUDE × DEM showing a high contribution rate in all three model predictions. PSE: In areas with low elevation, slopes below 35°, and increasing LUD, the overall effect is suppressive. In contrast, in areas with medium to high elevation and slopes above 35°, although the total area is smaller, increasing LUD promotes PS. LSE: In areas with low elevation, slopes below 30°, and increasing LUD, the overall effect is enhancing. However, in areas with medium to high elevation and slopes above 30°, the enhancement of LS is suppressed by increased LUD. ESE: In low elevation areas, the reduction in LUD suppresses ES evolution. In medium elevation areas, lower LUD promotes the enhancement of ES, while above high elevation, the total ES decreases, but areas with increasing LUD promote ES enhancement. As for slope, in areas with slopes below 10°, the increase in LUD promotes ES enhancement. In the 20–40° range, the effect of LUD is weaker, while in areas with slopes above 40°, increased LUD inhibits ES enhancement.

In all these cases, different villages (or communities) display spatial heterogeneity due to their unique geographical locations and terrain conditions. There are distinct variations in geographic environment, resource distribution, socio-economic characteristics, and land use types, leading to differing characteristics or impacts.

Overall, in mountain villages, settlement location and land use types are more influenced by natural factors. Unlike in plain areas, the evolution of PLES in mountain villages exhibits clear geographic dependency and gradient effects.

5.5. Limitations

First, due to the limited precision of temporal data at the town scale, and to ensure data alignment for the multi-temporal analysis of PLES evolution, this study adopted a unified 30 m grid resolution. This may fail to fully capture spatial changes at finer scales.

Secondly, in the discussion of driving mechanisms, the variables used to reflect PLES evolution remain insufficient. In particular, with regard to social development indicators, we obtained only village-level panel data through surveys; directly incorporating these data into the model as feature variables would greatly undermine the reasonableness of the model’s interpretation. Therefore, this study only selected LUD as a proxy for social development dimensions, which introduces a certain imbalance in the variable system. The performance evaluation results of the XGBoost model support this concern, with R² values generally ranging from 0.44 to 0.50, suggesting the potential omission of key explanatory variables. However, despite multiple rounds of field research and data retrieval, no micro-level social development gridded data (such as household income, population migration rates, etc.) could be obtained. As a result, the explanatory power of the analysis concerning the core driving mechanisms may be weakened, which represents an important direction for future work by the research team.

Finally, this study did not conduct a systematic evaluation or propose optimization pathways based on the results of PLES evolution. This will be a major focus of future research.

6. Conclusions

This study takes Taiji Town in Chongqing as a case to explore the gradient effects and driving mechanisms of the evolution of PLES in mountain villages. Through multi-temporal data analysis and gradient division, the study reveals the trends of land use changes and the gradient distribution characteristics of PLES in Taiji Town under complex site conditions. The results show that the evolution of PLES in mountain villages is influenced by both topographical factors and human activities. Under the influence of various factors, the evolution of PLES presents significant spatial heterogeneity. Against the backdrop of the ongoing rural revitalization strategy, how to optimize the resource allocation in mountain villages and scientifically plan the layout of PLES has become a key factor in improving resource utilization efficiency and promoting sustainable development. This study provides a new perspective for understanding the evolution patterns of PLES space in impoverished mountainous villages influenced by complex terrain, and offers a reference and analytical pathway for spatial planning and policy formulation in similar mountainous rural areas of southwestern China.