Adaptive and Differentiated Land Governance for Sustainability: The Spatiotemporal Dynamics and Explainable Machine Learning Analysis of Land Use Intensity in the Guanzhong Plain Urban Agglomeration

Abstract

Urban agglomerations underpin regional economic growth and sustainability transitions, yet the spatial heterogeneity and drivers of land use intensity (LUI) remain insufficiently resolved in inland settings. This study develops a high-resolution framework—combining a 1 km hexagonal grid with XGBoost-SHAP—to (i) map subsystem-specific LUI evolution, (ii) identify dominant drivers and nonlinear thresholds, and (iii) inform differentiated, sustainable land governance in the Guanzhong Plain Urban Agglomeration (GPUA) over 2000–2020. Composite LUI indices were constructed for human settlement (HS), cropland (CS), and forest (FS) subsystems; eleven natural, socioeconomic, urban–rural, and locational variables served as candidate drivers. The results show marked redistributions across subsystems. In HS, the share of low-intensity cells declined (86.54% to 83.18%) as that of medium- (12.10% to 14.26%) and high-intensity ones (1.22% to 2.56%) increased, forming a continuous high-intensity corridor between Xi’an and Xianyang by 2020. CS shifted toward medium-intensity (32.53% to 50.57%) with the contraction of high-intensity cells (26.62% to 14.53%), evidencing strong dynamism (55.1% net intensification; 38.5% net decline). FS transitioned to low-intensity dominance by 2020 (59.12%), with stability and delayed growth concentrated in conserved mountainous zones. Urban–rural gradients were distinct: HS rose by >20% (relative to 2000) in cores but remained low and stable in rural areas (mean < 0.20); CS peaked and stayed stable at fringes (mean ≈ 0.60); FS shifted from an inverse gradient (2000–2010) to core-area recovery by 2020. Explainable machine learning revealed inverted U-shaped relationships for HS (per capita GDP) and CS (population density) and a unimodal peak for FS with respect to distance to urban centers; model performance was strong (HS R² up to 0.82) with robust validation. Policy recommendations are subsystem-specific: enforce growth boundaries and prioritize infill/polycentric networks (HS); pair farmland redlines with precision agriculture (CS); and maintain ecological redlines with differentiated conservation and afforestation (FS). The framework offers transferable, data-driven evidence for calibrating thresholds and sequencing interventions to reconcile land use intensification with ecological integrity in rapidly urbanizing contexts.

1. Introduction

Urban agglomerations have become the predominant spatial frameworks for national economic advancement and global competitiveness, accommodating more than half of the global population and generating nearly 80% of the world’s gross domestic product [1,2]. As central engines of regional growth and the prevailing form of contemporary urbanization, shifts in land use patterns and land use intensity (LUI) within these agglomerations exert profound effects on regional sustainability, ecological integrity, and national food security [3]. LUI not only denotes the types of land use but also quantifies the degree of human appropriation of natural ecosystems within specific land cover classes, and it is therefore a key indicator for evaluating land use sustainability [4].

The dynamics of LUI emerge from multifaceted interaction among drivers operating across spatial and temporal scales. These include natural endowments (e.g., soils, topography) that constrain or enable utilization [5]; socioeconomic processes (e.g., population growth, economic activity) that underpin intensification [6]; locational attributes (e.g., accessibility to transport networks and urban centers) that shape opportunities and development potential [7]; and policy interventions (e.g., land use regulation, urban planning) that govern spatiotemporal trajectories [8].

A growing body of literature examines how intensive land use (ILU) influences urban development, agricultural productivity, and ecological quality [9,10,11]. In urban contexts, evidence from Chinese and European cities indicates that ILU and compact urban forms can enhance efficiency and economic performance [9,10,12]. However, when intensification exceeds optimal thresholds, livability and broader social outcomes tend to deteriorate [12,13], and reductions in urban green space alongside shifts in urban morphology become evident [12,14]. In agriculture, intensification through more efficient per-area inputs and higher multiple-cropping indices remains central to boosting yields and meeting food demand [15,16]. Ecologically, ILU places pressure on ecosystems and biodiversity by diminishing habitat area and connectivity [11,17,18].

Despite these advances, refined, system-specific research into LUI remains limited—particularly work that identifies causal mechanisms and compares heterogeneous drivers across urban, agricultural, and forest subsystems [19,20]. The multidimensional nature of LUI demands analyses that move beyond land cover transitions to capture within-class management intensity [19]. For instance, urban LUI involves building density, floor area ratio, population capacity, and economic activities [21]; agricultural LUI encompasses cropping intensity, agrochemical inputs, and mechanization [22]; and forest LUI reflects silvicultural practices, logging, and management regimes [23]. Addressing these dimensions requires a shift toward the underlying drivers, resource efficiencies, and input–output relationships [20].

Conventional approaches to analyzing LUI drivers—such as linear or logistic regression [24]—offer limited explanatory power in high-dimensional settings characterized by nonlinearity, thresholds, and interactions [25]. They frequently violate key assumptions (e.g., linearity, independence) [26] and struggle to recover complex cross-factor effects [27]. Recent advances in artificial intelligence and data availability have catalyzed the adoption of machine learning methods in land use research, enabling flexible nonlinear modeling and scalable data integration [28]. Among these, the XGBoost-SHAP framework shows particular promise for disentangling mechanisms and interpreting feature effects [29]; however, its application in LUI remains emergent [30], and comprehensive frameworks for subsystem-level investigations (human settlement, cropland, forest) are still scarce [31].

The Guanzhong Plain Urban Agglomeration (GPUA)—a strategic hub at the Loess Plateau–Qinling Mountains interface and a key node in both the Belt and Road Initiative and China’s Western Development Strategy—has experienced rapid yet uneven growth [32]. Intensifying human activities have sharpened land development, which has led to ecological conservation tensions [33], underscoring the need for science-based, adaptive governance [34,35]. Neglecting LUI obscures critical human–environment linkages and risks suboptimal policy design [36].

To address these gaps, this study develops a 1 km hexagonal grid LUI framework to assess LUI across human settlement, cropland, and forest subsystems in the GPUA from 2000 to 2020 and integrates the XGBoost-SHAP analytical approach to reveal nonlinear contributions, thresholds, and interaction effects among diverse drivers. Our objectives are threefold: (1) to map subsystem-specific spatiotemporal LUI evolution in the GPUA; (2) to identify and interpret dominant drivers and heterogeneous effects using explainable machine learning; and (3) to provide policy recommendations for differentiated, efficient, and sustainable land management in inland urban agglomerations.

The remainder of this paper proceeds as follows. Section 2 details the study area, analytical framework, data, and driver analysis. Section 3 reports the results on spatiotemporal LUI characteristics (structure, distribution), subsystem variation (urban–rural gradients, change patterns), and key drivers. Section 4 discusses synergistic changes and policy responses in the GPUA, nonlinear mechanisms and thresholds, and implications for sustainable LUI policy. Section 5 concludes this paper.

2. Materials and Methods

2.1. Study Area

The Guanzhong Plain Urban Agglomeration (GPUA)—the second largest urban agglomeration in western China—extends from 33.35° N to 36.72° N and 104.57° E to 112.22° E [37] (Figure 1). It spans three provinces (Shaanxi, Gansu, and Shanxi), with Xi’an as its core, and comprises 11 municipalities: Xi’an, Baoji, Shangluo, Xianyang, Weinan, Tongchuan, Qingyang, Pingliang, Tianshui, Linfen, and Yuncheng. Situated in the lower Yellow River Basin at the Loess Plateau–Qinling Mountains transition, the GPUA exhibits a west-to-east descending topography, forming a U-shaped basin—mountain-bounded on three sides and opening eastward [38]. The landscape is dominated by alluvial plains, loess tablelands, and piedmont alluvial fans. The Wei River system provides the principal surface water support for agriculture and urban expansion [39].

The region has a warm temperate monsoon climate, ranging from semi-humid to semi-arid. The mean annual temperature is 9–13 °C, and the mean annual precipitation is 500–800 mm, with falls concentrated from July to September [40]. In 2020, the GPUA’s permanent resident population reached 42.35 million (2.79% of China’s total), with an urbanization rate of 60.06% [41]. Covering 107,100 km² (1.12% of national land area), the agglomeration generated CNY 2.19 trillion GDP (2.20% of national GDP). The GPUA offers a representative inland mega-region where rapid urbanization intersects with pronounced ecological constraints. Its cross-provincial governance, diverse landforms, and strong hydrological dependence create marked spatial heterogeneity in land use processes, making it an appropriate case for analyzing land use intensity (LUI) dynamics and subsystem-specific trajectories (human settlement, cropland, forest) in large urban agglomerations.

2.2. Analytical Framework and Data

This study adopts a structured analytical framework (Figure 2) comprising three core components: (i) a 1 km hexagonal grid-based assessment of LUI across human settlement (HS), cropland (CS), and forest (FS) subsystems; (ii) spatiotemporal analysis to identify structural shifts and distribution patterns; and (iii) XGBoost–SHAP modeling to identify nonlinear driver effects, thresholds, and interactions. This design responds to calls for high-resolution, subsystem-specific approaches that generate reproducible evidence regarding human–environment dynamics [32,37].

Land cover data for 2000, 2010, and 2020 were sourced from the China Land Cover Dataset (CLCD) [42], ensuring consistency in classification and temporal coverage.

2.2.1. Subsystem Classification

Land use intensity (LUI) was evaluated on a 1 km hexagonal grid, selected for its near-isotropic geometry, reduced edge effects, and lower sampling bias relative to rectangular tessellations [43]. Each hexagon was assigned to one of three subsystems according to the dominant land cover type: human settlement (HS; impervious/built surfaces indicating concentrated human activity) [44], cropland (CS; agricultural land for food production) [45], and forest (FS; forests, shrublands, or grasslands providing ecosystem services) [46]. This subsystem stratification enables a targeted analysis of within-class dynamics and mitigates the limitations of aggregate land use assessments [47].

2.2.2. Selection and Calculation of Land Use Intensity Indicators

Subsystem-specific indicators were specified to operationalize LUI, with emphasis on human intervention and sustainability considerations [48]. A multi-criteria decision-making (MCDM) composite index framework was adopted [49,50], with equation forms following Fu et al. [51]. Eight indicators per subsystem were chosen based on human activity intensity, urban attributes, crop production and intensity, forest management and reserve, and data availability (

Table 1. Indicators and data sources used to construct composite land use intensity (LUI) indices for human settlement, cropland, and forest subsystems.

Subsystem	Indicator	Source	Spatial Resolution
Human Settlement Systems	Population Density (PD)	Global Human Settlement Layer (GHSL), European Commission [52]	100 m
	Built-up Volume (BV)	Global Human Settlement Layer (GHSL), European Commission [52]	100 m
	Nighttime Light (NTL)	Global Annual Simulated VIIRS Nighttime Light Dataset (1992–2023) [53]	500 m
Crop Systems	Cropping Intensity (CI)	Annual Dynamic Dataset of Global Cropping Intensity (2001–2019) [54]	250 m
	Grains Production (GP)	Global Wheat Yield 4 km [55]	4 km
	Nitrogen Fertilizer Inputs (NFIs)	History of Anthropogenic Nitrogen Inputs (HaNI) [56]	300 m
Forest Systems	Forest Management (FM)	Annual Maps of Global Forest Management Types (2001–2020) [57]	250 m
	Forest Reserve (FR)	Boundary Data of National Nature Reserves [58]	-

). All indicators were harmonized to the 1 km grid and standardized to enhance comparability and reduce scale-related bias; equal weights were applied, and scores were normalized to a 0–1 range to enable cross-subsystem and intertemporal analyses.

1.

Human Settlement Subsystem (HS)

Within the HS subsystem, land use intensity (LUI) was operationalized using three indicators of anthropogenic activity and inputs: population density (PD), building volume (BV), and nighttime light intensity (NTL). PD and BV were obtained from the Global Human Settlement Layer (GHSL) produced by the European Commission’s Joint Research Centre, and NTL was derived from the annual NPP-VIIRS nighttime light (NTL) composites. In combination, these indicators capture complementary dimensions of human presence (population), built environment stock (building mass), and energy/economic activity (luminosity), providing an integrated proxy for urban development intensity at the grid cell level.

The LUI of the human settlement system for each grid cell was calculated as follows:

$$LU{I}_{HS}={\displaystyle \frac{P{D}_i-P{D}_{\min }}{P{D}_{\max }-P{D}_{\min }}}+{\displaystyle \frac{B{V}_i-B{V}_{\min }}{B{V}_{\max }-B{V}_{\min }}}+{\displaystyle \frac{NT{L}_i-NT{L}_{\min }}{NT{L}_{\max }-NT{L}_{\min }}}$$

(1)

where $LU{I}_{HS}$ represents the land use intensity value for the human settlement system in a given grid cell, $PD$ is population density (people/m²), $BV$ is building volume (m³), and $NTL$ denotes nighttime light intensity.

2.

Cropland Subsystem (CS)

Within the CS subsystem, land use intensity (LUI) was assessed using three indicators that represent agricultural management and input–output dynamics: cropping intensity (CI), grain yield (GY), and nitrogen fertilizer input (NFI). CI was obtained from the global 250 m dynamic cropping intensity dataset (2001–2019); GY from the GlobalWheatYield4km product; and NFI from the History of Anthropogenic Nitrogen Inputs (HaNi) dataset. All layers were harmonized to the 1 km analysis grid and aligned to the study years. Together, these indicators characterize production outcomes and input intensity, providing an objective basis for assessing agricultural land use intensity and its relevance to sustainable management.

The LUI of the cropland system for each grid cell was calculated as follows:

$$LU{I}_{CS}={\displaystyle \frac{C{I}_i-C{I}_{\min }}{C{I}_{\max }-C{I}_{\min }}}+{\displaystyle \frac{G{P}_i-G{P}_{\min }}{G{P}_{\max }-G{P}_{\min }}}+{\displaystyle \frac{N{I}_i-N{I}_{\min }}{N{I}_{\max }-N{I}_{\min }}}$$

(2)

where $LU{I}_{CS}$ is the land use intensity value for a single grid cell in the cropland system, and $CI$, $GP$, and $NI$ represent cropping intensity, Grain Production (ton/m²), and nitrogen fertilizer input (kg/m²), respectively.

3.

Forest Subsystem (FS)

Within the forest (FS) subsystem, land use intensity (LUI) was characterized using two indicators that jointly capture resource use and conservation dimensions: forest management (FM) and forest reserve (FR). FM was derived from the Global Forest Management Map by classifying forest areas into five management types—agroforestry systems, short-rotation artificial forests (less than 15 years), long-rotation artificial forests (greater than 15 years), commercial natural regeneration forests, and non-commercial natural regeneration forests—to represent a gradient of human intervention. FR denotes the presence or extent of designated forest reserves within each analysis unit, reflecting protection status. In combination, FM and FR provide a nuanced appraisal of forest land use intensity consistent with contemporary forest management typologies.

Additionally, a forest reservation level was incorporated to capture the status and effectiveness of ecological protection measures [59]. Higher reservation levels indicate stronger conservation regimes and, by design, lower expected human use intensity [60]. The land use intensity of the forest system for each grid cell was calculated as follows:

$$LU{I}_{FS}={\displaystyle \frac{F{M}_{\max }-F{M}_i}{F{M}_{\max }-F{M}_{\min }}}+{\displaystyle \frac{F{R}_i-F{R}_{\min }}{F{R}_{\max }-F{R}_{\min }}}$$

(3)

where $LU{I}_{FS}$ denotes the land use intensity value for an individual grid of the forest system. $FM$ represents the forest management type, while $FR$ indicates forest protection levels. Protection levels are assigned as follows: national-level forest reserves = 0.1; provincial-level forest reserves = 0.5; municipal-level forest reserves = 0.7; and other cases = 1.

2.2.3. Analysis of Driving Factors

Land use intensity (LUI) dynamics reflect the joint influence of natural, social, economic, technological, and policy processes operating across multiple spatial and temporal scales [61]. Conventional parametric regressions, although widely used, have limited capacity to represent the nonlinearities, high-dimensional interactions, and threshold effects that characterize land use systems. Accordingly, this study employs gradient-boosted decision trees (XGBoost) to model LUI—an approach well suited to large, high-dimensional tabular data and flexible, nonparametric response surfaces [62,63]. To enhance the interpretability and support effect attribution, Shapley Additive Explanations (SHAP) are used to (i) identify and rank the dominant drivers of LUI, (ii) quantify their marginal contributions, and (iii) diagnose threshold and interaction patterns within and across the HS, CS, and FS subsystems.

Analyses are conducted at the 1 km grid cell level within the Guanzhong Plain Urban Agglomeration (GPUA). Separate machine learning regression models are trained for the human settlement (HS), cropland (CS), and forest (FS) subsystems for the years 2000, 2010, and 2020. In each model, the subsystem-specific LUI serves as the response variable, and a multi-source set of candidate drivers (natural, socioeconomic, locational, and policy proxies) constitutes the predictors. This design enables the subsystem- and period-specific estimation of nonlinear relationships and possible threshold effects between LUI and its driving factors while maintaining comparability across time and land system contexts.

4.

Dataset assembly and preprocessing

The analytical dataset comprises multi-period observations of LUI and associated socioeconomic–environmental covariates for the Guanzhong Plain Urban Agglomeration (GPUA). A 1 km hexagonal tessellation was employed to enhance spatial isotropy and reduce edge/sampling bias relative to rectangular grids [43]. The response variable consists of subsystem-specific LUI measures for the human settlement (HS), cropland (CS), and forest (FS) subsystems in 2000, 2010, and 2020. The explanatory set includes eleven drivers (X1–X11, indexed by year, e.g., X1_2000 … X11_2020) spanning demographic pressure, economic development, accessibility, and biophysical/environmental conditions. All variables were harmonized to the 1 km grid and standardized to ensure cross-indicator comparability and mitigate unit inconsistencies [26].

5.

Experimental Design and Model Configuration

For each subsystem–year model, observations were randomly partitioned into training (80%) and test (20%) sets. Model performance on the training data was assessed via 5-fold cross-validation (CV), reporting the mean ± standard deviation of evaluation metrics across folds. Generalization was appraised by comparing the held-out test performance with the corresponding CV mean (i.e., the CV–test discrepancy), following recommended practices for estimating out-of-sample error [64]. Random partitions were seeded to ensure reproducibility.

Predictive performance was quantified using the following: (i) the coefficient of determination (R²) to summarize explanatory power; (ii) root mean squared error (RMSE) to reflect error magnitude; and (iii) mean absolute error (MAE) to capture average predictive deviation.

The baseline regressor was eXtreme Gradient Boosting (XGBoost), selected for its capacity to handle high-dimensional tabular data, model nonlinear relationships and interactions, and mitigate issues related to predictor multicollinearity [30]. Hyperparameters were tuned using cross-validated search on the training set; final models were refit on the full training data before test evaluation. To provide transparent, post hoc explanations of model outputs, SHAP (Shapley Additive Explanations) was employed to decompose predictions into marginal feature contributions, enabling the identification and ranking of key drivers, as well as the examination of nonlinearities and potential thresholds [29].

6.

Robustness Checks

To evaluate the stability and uncertainty of the results beyond point estimates, a nonparametric bootstrap was conducted with B = 1000 resamples drawn from the training set. For each bootstrap sample, the model was retrained using the same hyperparameters and evaluated on the fixed test set, and percentile-based 95% confidence intervals were computed for R², RMSE, and MAE [65].

The robustness of feature explanations was assessed in parallel. Within each bootstrap replicate, features were ranked by mean absolute SHAP values; rank distributions were then summarized across replicates. Cross-metric consistency was quantified using Spearman’s rank correlation (ρ) between SHAP-based rankings and XGBoost gain-based importance rankings [66]. This procedure provides distributional evidence on performance uncertainty and on the stability of driver attribution.

7.

Variable Definitions and Expected Effects

Land use intensity (LUI) dynamics reflect the joint action of natural, socioeconomic, technological, policy, and institutional processes operating across multiple spatial and temporal scales [4]. In urban agglomerations, these processes are commonly organized into four driver classes: natural environmental constraints, socioeconomic pressures, locational accessibility, and policy frameworks [5,6,7,8]. To ensure theoretical coherence and interpretability, eleven explanatory variables (X1–X11) were specified across these four dimensions (

Table A1. List of driving factors for LUI analysis.

Driving Factors	Theoretical Rationale	Code	Variable	Expected Effect on LUI *	Source
Natural Factors	These factors exert a profound influence on the type, pattern, and potential of land use over longer time scales. However, their changes may not be significant within relatively short periods (e.g., a few years or decades), instead acting more as a relatively stable “baseline” condition. Natural factors primarily determine the range of suitability and limiting conditions for land use in a specific region.	X1	Precipitation	Variable: Adequate levels may positively enable higher-intensity in cropland and forest subsystems, but extremes (e.g., drought/flooding) constrain overall LUI [92].	The data were sourced from the “1960–2020 China 1 km Resolution Precipitation Dataset,” which provides monthly precipitation data for China at a spatial resolution of 0.008333° (approximately 1 km), covering the period from 1960 to 2020. This dataset is available from geodata.cn.
		X2	Temperature	Variable: Optimal ranges support intensification in agricultural and settlement uses, while extremes limit suitability across subsystems [92].	The data were obtained from the Dataset of Daily Near-Surface Air Temperature in China from 1979 to 2018 (CDAT), which encompasses daily maximum, minimum, and average temperatures (T_max, T_min, T_avg) at a spatial resolution of 0.1° over the period 1979–2018. The dataset is publicly available via Zenodo.
		X3	Elevation	Negative: Higher elevations typically constrain LUI intensification, particularly in human settlements and croplands due to accessibility and climatic challenges [74].	The data were sourced from the FathomDEM v1.0—Eurasia & Africa dataset, which utilizes the WGS84 coordinate system (EPSG:4326), with elevations referenced to EGM08 (EPSG:3855), in units of centimeters, formatted as GeoTIFF, and organized in 1° × 1° tiles. Version 1.0 was released on the Zenodo platform in February 2025.
		X4	Slope	Negative: Steeper slopes reduce land suitability for intensive uses, limiting development in cropland and settlement subsystems [93].	Slope data for the study area were derived using GIS software from the FathomDEM v1.0—Eurasia & Africa dataset.
Socioeconomic Factors	These factors reflect the intensity of demand for and the capacity to transform land resources, influenced by the stage of social development, economic structure, technological level, and market demand.	X5	Population Density	Positive: Higher density increases pressure for land intensification, especially in human settlement subsystems [94].	The data were obtained from the WorldPop China 1 km Population Density Dataset, which provides gridded population density estimates in persons per square kilometer, in GeoTIFF (.tif) format, at a spatial resolution of 30 arc-seconds (approximately 1 km). The dataset is published by WorldPop.
		X6	Gross Domestic Product	Positive: Greater economic output drives land conversion and higher LUI across urban agglomerations [95].	The data were sourced from the China GDP Spatial Distribution Kilometer Grid Dataset, which includes multiple time points such as 1995, 2000, 2005, 2010, 2015, and 2019. The dataset features a spatial resolution of 1 km × 1 km grids, with units in CNY 10,000 per square kilometer, in grid format, and employs the Krassovsky ellipsoid with the Albers projection system. This dataset was released by the Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, and is accessible via the RESDC platform (DOI: 10.3974/geodb.2014.01.07.V1).
		X7	GDP Per Capita	Positive: Indicates economic prosperity that accelerates LUI intensification through investment and urbanization [93].	The data were derived from the “Downscaled Gridded Global Dataset for Gross Domestic Product (GDP) per Capita at Purchasing Power Parity (PPP) over 1990–2022,” published in Scientific Data in 2025 (DOI: 10.1038/s41597-025-04487-x). This dataset spans the period from 1990 to 2022, encompassing GDP per capita (PPP) values for various intervals. It consists of gridded data at a 5 arc-minute resolution and is accessible via the Zenodo platform (DOI: 10.5281/zenodo.10976733).
Policy Factors (Urban–rural Integration Factors)	These factors guide, regulate, incentivize, or constrain the behavior of land use actors, thereby significantly influencing the type, structure, intensity, and spatiotemporal evolution of land use. Such factors often exhibit strong guidance and mandatory characteristics, and in this study, they specifically pertain to urban–rural integration policies.	X8	Urban–Rural Population Distribution	Moderating: Balanced distributions may reduce rural LUI intensification while promoting efficient urban use [92].	The data were sourced from the Global Urban and Rural Settlement Dataset (GURS), which provides global coverage from 2000 to 2020 at a 100 m resolution and was published in Scientific Data in 2024 (DOI: 10.1038/s41597-024-04195-y). The dataset is available via the Zenodo platform.
		X9	Urban–Rural Gradient Distribution	Moderating: Gradients (e.g., in infrastructure or income) can constrain excessive LUI in transitional zones, fostering sustainable integration [96].	Data processing was based on the GHS-SMOD R2023A dataset, which forms part of the Global Human Settlement Layer (GHSL) operated by the European Commission’s Joint Research Centre (JRC). This dataset utilizes the Degree of Urbanisation model (recommended by the United Nations), with a spatial resolution of 1 km and employing the World Mollweide projection. The data are accessible via data.europa.eu.
Locational Factors	These factors indirectly or directly regulate the type and intensity of land use by influencing opportunity costs, transportation costs, ease of information access, and agglomeration effects.	X10	Road Length	Positive: Greater road length (or density) enhances accessibility, promoting higher LUI in connected areas [97].	The data were sourced from the statistical yearbooks of the cities within the Guanzhong Urban Agglomeration for the years 2000, 2010, and 2020.
		X11	Distance to the City Center	Negative: Greater distance reduces LUI due to lower accessibility and higher costs [98].	The data were derived from the CLCD land use dataset, with parameters assigned according to the travel impedance of different land use types (e.g., permeability and barriers). Cost rasters were then computed using GIS software to serve as a proxy for spatial accessibility.

* Note: Expected effects are hypothesized based on general patterns in urban agglomeration studies and may vary by subsystem (e.g., human settlements vs. croplands). Potential multicollinearity (e.g., between X6 and X7) should be tested in modeling. All variables assume grid-level aggregation for the Guanzhong Urban Agglomeration.

), drawing on the established literature for the Chinese urban context [67]. Table A1 reports variable definitions, data sources, temporal coverage, hypothesized signs by subsystem, etc.

Natural factors establish baseline conditions and constraints for LUI, often exerting long-term stabilizing or limiting effects. This dimension includes four indicators (X1–X4), expected to negatively correlate with LUI intensification in sensitive subsystems like forests and croplands due to environmental vulnerabilities (e.g., reduced development in steep terrains). Socioeconomic factors reflect human-induced changes, typically driving rapid LUI increases through urbanization and economic growth; they comprise three indicators (X5–X7) and are anticipated to positively influence LUI, particularly in human settlement subsystems. Policy factors, specifically urban–rural integration, encompass governmental mechanisms that regulate land use via planning, incentives, and constraints, shaping trajectories toward balanced development; represented by two indicators (X8–X9), these are expected to moderate LUI, potentially reducing intensity in rural areas while promoting efficient urban use. Locational factors capture spatial accessibility to key features (e.g., infrastructure), facilitating land conversion; with two indicators (X10–X11), they are hypothesized to positively affect LUI by enhancing connectivity and economic viability.

3. Results

3.1. Spatiotemporal Characteristics of Land Use Intensity

3.1.1. Structure of Land Use Intensity

Over the past two decades, the distribution of land use intensity (LUI) has shifted markedly across all three subsystems (Figure 3).

• Human settlement (HS) subsystem. Low-intensity grids remained predominant but declined modestly from 86.54% (2000) to 83.18% (2020; −3.36 percentage points, pp). Medium-intensity areas increased from 12.10% to 14.26% (+2.16 pp) and high-intensity areas from 1.22% to 2.56% (+1.34 pp), indicating steady densification, particularly within intermediate zones.
• Cropland (CS) subsystem. A pronounced structural reallocation is evident. Low-intensity cropland rose from 40.85% (2000) to 50.10% (2010) before decreasing to 34.90% (2020). Meanwhile, medium-intensity cropland increased from 32.53% to 50.57% over the period, whereas high-intensity cropland decreased from 26.62% to 14.53%. Overall, this pattern is consistent with a shift toward more balanced cultivation intensities by 2020.
• Forest (FS) subsystem. The earlier dominance of high-intensity forest (52.68% in 2000; 57.59% in 2010) was reversed by 2020, when low-intensity forest rose to 59.12%, indicating a broad reduction in anthropogenic disturbance consistent with large-scale restoration and afforestation initiatives.

3.1.2. Spatial Distribution of Land Use Intensity

In 2000, the GPUA exhibited a clear center–periphery (concentric) gradient in land use intensity (Figure 4). Within HS, high-intensity cells were concentrated in Xi’an and secondary cores—Xianyang, Weinan, Baoji, Linfen, and Yuncheng—with medium-intensity HS extending along major corridors and low-intensity HS prevalent on rural fringes. CS’s high-intensity zones formed a belt along the Xi’an–Xianyang–Weinan–Baoji axis; medium-intensity cropland dominated the central plain; and low-intensity cropland occupied marginal uplands. For FS, high-intensity areas were concentrated in the steep, densely forested southern Qinling Mountains and in isolated pockets across the northern foothills, while low-intensity FS was primarily distributed around the periphery of these core forests—along the northern slopes, within the gullied sectors of Loess Plateau gullies, and in riparian restoration zones.

By 2010, the LUI configuration exhibited notable shifts (Figure 4). Within HS, the high-intensity area expanded westward into Xianyang and northward, forming near-continuous urban belts; medium-intensity HS further penetrated suburban corridors. In CS, high-intensity cropland retreated along metropolitan fringes—coincident with urban expansion—while remaining extensive within core agricultural tracts; medium- and low-intensity cropland remained continued to dominate hilly terrain and intermontane basins. The FS distribution displayed broad stability: core high-intensity forests persisted within steep, protected mountain reserves, whereas low-intensity FS remained concentrated along the margins of these areas, consistent with steady stable conservation outcomes over this interval.

By 2020, HS formed a continuous corridor between Xi’an and Xianyang, with medium-intensity HS infilling the peri-urban area. Along metropolitan fringes, urban expansion coincided with the conversion of portions of high-intensity CS to HS; nevertheless, extensive high-intensity cropland persisted in remote grain-producing plains, and medium- and low-intensity CS remained prevalent in transitional landscapes. FS again exhibited broad stability: high-intensity forests were largely confined to major ecological barriers (e.g., Qinling and Bashan ranges), while low-intensity FS continued to encircle these cores—a pattern consistent with sustained afforestation and protection initiatives.

3.2. Subsystem-Specific Variations in Land Use Intensity

3.2.1. Spatiotemporal Structural Dynamics: Variations in Land Use Intensity Along the Urban–Rural Gradient

Within urban agglomerations, the urban–rural gradient denotes a continuum from densely built cores to sparsely populated peripheries, characterized by systematic gradients in urbanization level, population density, building intensity, and ecosystem service provision [68,69]. Using Global Human Settlement (GHS) products, this study employed metrics of urbanization class, population distribution, and mixed land use attributes [70,71] to delineate the Guanzhong Plain Urban Agglomeration (GPUA) into three zones: urban center, urban fringe, and rural area (Figure 5). The resulting zoning is consistent with concentric-ring models of urban morphology [72]. Under this framework, land use intensity (LUI) in the human settlement (HS) and forest (FS) subsystems exhibits a radial decay pattern from the core outward.

Subsystem-specific LUI gradient dynamics are summarized as follows:

Human Settlement System (HS): HS LUI declines markedly from the urban core to rural peripheries. Over 2000–2020, the mean HS LUI in the urban center rose by over 20% (relative to 2000), consistent with continued infill and construction activities. The urban fringe registered a ~10% rise over the same period. By contrast, rural HS LUI remained low (mean < 0.2) and largely unchanged, indicating that high-intensity settlement was effectively concentrated within peri-urban belts.

Cropland System (CS): CS LUI peaks prominently in the urban fringe (mean ≈ 0.6), consistent with greater market access, transport connectivity, and service availability that enable more intensive management. Fringe CS LUI remained broadly stable over 2000–2020. By contrast, the mean CS LUI in the urban core and rural zones declined by ~5–10% relative to 2000 levels, a pattern consistent with the urban conversion of arable land in core areas and a shift toward more extensive practices in remote agricultural areas.

Forest System (FS): From 2000 to 2010, FS LUI exhibited an inverse core—periphery gradient—rising from the urban core through the fringe to the rural hinterland—while absolute levels declined across all zones, indicating reduced anthropogenic disturbance. By 2020, the pattern reversed: FS LUI in the urban center exceeded its 2000 baseline (consistent with peri-urban afforestation and urban greening initiatives); the fringe remained broadly comparable to 2010, suggesting a stable balance between utilization and conservation; and the rural zone continued a gradual increase, in line with large-scale ecological restoration programs (e.g., Grain-for-Green).

3.2.2. Spatiotemporal Distribution Patterns: Dynamics of Land Use Intensity Change

To examine heterogeneity in LUI trajectories, analysis was restricted to grid cells with unchanged land cover classes over 2000–2020. For each subsystem, intensity changes were classified into nine trajectory types based on directional shifts—decrease, no change, or increase—across the 2000–2010 and 2010–2020 intervals (

Table 2. Classification of LUI change trajectories from 2000 to 2020.

2000–2010 Change	2010–2020 Change	Category Code	Type Description
Decrease	Decrease	1	Continuous decline type
Decrease	No Change	2	Stabilized after decline type
Decrease	Increase	3	Rebound after decline type
No Change	Decrease	4	Delayed decline type
No Change	No Change	5	Long-term stable type
No Change	Increase	6	Delayed growth type
Increase	Decrease	7	Fluctuating decline type
Increase	No Change	8	Stabilized after growth type
Increase	Increase	9	Continuous growth type

, Figure 6). This typology follows established approaches to temporal land use dynamics and supports a nuanced, within-class interpretation of subsystem-specific evolution.

In the human settlement (HS) subsystem, the majority of HS grids (86.5%) exhibited no net change in LUI (Category 5), indicating stable development intensity within established urban fabric. Sustained or accelerated intensification (Categories 6, 8, 9) affected 10.5% of cells, concentrated around major centers—Xi’an, Xianyang, and Weinan—where expansion approached saturation. Fluctuating or declining patterns (Categories 1–4, 7) were rare (<3%) and spatially localized, primarily in Baoji, Yuncheng, and Linfen, with small occurrences in the central district of Xi’an; these areas are consistent with redevelopment cycles or temporary contractions.

The cropland (CS) subsystem exhibited pronounced dynamism: only 1.7% of grid cells showed no net change (Category 5). Over half of the CS area (55.1%) underwent net intensification (Categories 3, 6, 9), concentrated in agriculturally productive zones—southern Baoji, western Xianyang, southern Weinan, and northwestern Yuncheng—and consistent with increases in management intensity and policy incentives. By contrast, 38.5% experienced a net decline (Categories 1, 2, 4, 7), prevalent in Tianshui, Linfen, and northern Xianyang; many of these followed fluctuating-decline trajectories (Category 7) and prevailed, often located in areas characterized by stronger environmental constraints and proximity to expanding urban fronts.

Forest subsystem LUI was predominantly stable: 59.1% of grid cells exhibited long-term equilibrium (Category 5), concentrated in the mountainous areas of southern Shaanxi (e.g., Shangluo, southern Baoji) and along the Qinling foothills under established conservation designations. Delayed growth (Category 6) accounted for 26.2%, prominent across the Qinling and Bashan Mountains, consistent with progressive restoration associated with programs such as Grain-for-Green. Transient changes (Categories 2–4) comprised 12.8%, occurring mainly along the northern margins of Linfen and peripheral urban fringes, coinciding with topographic constraints, policy implementation timelines, and disturbance episodes.

3.3. Drivers of Land Use Intensity

Changes in land use intensity (LUI) arise from the interaction of natural, socioeconomic, technological, and policy factors operating across multiple spatial and temporal scales [63]. Identifying the dominant drivers and their mechanisms is essential for diagnosing current land use dynamics, informing regulatory interventions, and projecting future trends within the GPUA. Guided by established land system frameworks, eleven explanatory variables (X1–X11) were specified across four domains—natural conditions (X1–X4), socioeconomic attributes (X5–X7), urban–rural integration (X8–X9), and locational factors (X10–X11)—to capture the multidimensional determinants of LUI (Figure 7).

A panel of approximately 83,000 1 km grid cells with 11 explanatory variables was used to train gradient-boosted tree (XGBoost) models with early stopping and row/column subsampling. Separate models were estimated for the human settlement (HS), cropland (CS), and forest (FS) subsystems for 2000, 2010, and 2020.

The HS subsystem achieved the highest explanatory power, with R² = 0.76–0.82 (

Table A2. Performance metrics and cross-validation results for XGBoost models of land use intensity across human settlement, cropland, and forest subsystems in Guanzhong Urban Agglomeration (2000–2020).

Exp	Year	Sample Size	R2 (Final)	RMSE	MAE	CV R2 (Mean ± Std)	Gap (R2)	RMSE/MAE
E1 CS2000	2000	83,129	0.6562	0.09914	0.05054	0.6499 ± 0.00584	+0.00634	1.962
E2 CS2010	2010	83,347	0.6441	0.10345	0.05600	0.6509 ± 0.01123	−0.00687	1.847
E3 CS2020	2020	83,347	0.5915	0.11281	0.06263	0.5972 ± 0.00643	−0.00572	1.800
E4 FS2000	2000	83,129	0.7203	0.13971	0.07417	0.7112 ± 0.00690	+0.00917	1.883
E5 FS2010	2010	83,347	0.5380	0.155996	0.08834	0.5391 ± 0.00409	−0.00111	1.766
E6 FS2020	2020	83,347	0.6002	0.17505	0.10684	0.6003 ± 0.00362	−0.00006	1.638
E7 HS2000	2000	83,129	0.7604	0.01223	0.00421	0.7714 ± 0.00813	−0.01106	2.906
E8 HS2010	2010	83,347	0.8167	0.01483	0.00533	0.8203 ± 0.00743	−0.00359	2.783
E9 HS2020	2020	83,347	0.8085	0.01357	0.00526	0.8077 ± 0.00469	+0.00079	2.581

), indicating the strong capture of urbanization-related variation. The FS subsystem peaked in 2000 (R² = 0.72), followed by a decline in 2010 and a partial recovery by 2020 (R² = 0.60). The CS subsystem showed a gradual decrease in fit over time (R²: 0.66 → 0.59, 2000–2020). Across experiments, cross-validation variability was low (SD ≈ 0.003–0.011), consistent with stable generalization and limited overfitting.

Overall, the models provide reliable, high-resolution evidence on nonlinear relationships, interactions, and threshold patterns among LUI drivers within the Guanzhong Plain Urban Agglomeration.

SHAP (Shapley Additive Explanations) was applied to the XGBoost models to decompose predictions and quantify, for each variable, its marginal contribution, effect direction, and across-cell distributional in the human settlement (HS), cropland (CS), and forest (FS) subsystems over 2000–2020 (Figure 8).

• Human Settlement Subsystem (HS)

Socioeconomic variables—particularly population density (X5) and per capita GDP (X7)—exhibited the largest positive SHAP contributions to HS LUI, indicating a strong association between agglomeration/economic activity and urban intensification. Proximity to city centers (X11) was likewise positively associated with HS LUI. Measures of urban–rural integration (X8, X9) increased in importance over time, consistent with the rising influence of peri-urban expansion processes.

• Cropland Subsystem (CS)

At the beginning of the period, biophysical constraints—most notably slope (X4)—had the largest absolute SHAP attributions, indicating a strong (negative) association with CS LUI. By 2020, socioeconomic pressures had become dominant, with population density (X5), proximity to urban markets (X11), and gross domestic product (X6) exhibiting the strongest positive contributions. Accessibility (road length, X10) also contributed positively, consistent with the increasing integration of cropland systems into regional economic networks.

• Forest Subsystem (FS)

Natural environmental variables—precipitation (X1), elevation (X3), and slope (X4)—exhibited the largest positive SHAP contributions to FS LUI, indicating that biophysical suitability is associated with higher forest use intensity, on average. By contrast, indicators of human activity and market access—population density (X5), GDP (X6), per capita GDP (X7), and road length (X10)—were negatively associated with forest LUI, consistent with lower intensity in more accessible, development-oriented settings. Distance to urban centers (X11) emerged as a salient determinant with a nonlinear (unimodal) relationship: FS LUI peaked at intermediate distances, whereas both near-core and remote locations were linked to lower intensity.

SHAP dependence plots (Figure 9 and Figure 10) indicate pervasive nonlinear and threshold-like relationships. In the HS subsystem, per capita GDP shows a concave (inverted U) association with LUI, with diminishing marginal gains and eventual declines beyond a critical income level. In the CS subsystem, population density exhibits a similar inverted U pattern: moderate densities are associated with higher cultivation intensity, whereas excessive densities correspond to land competition and conversion pressures that lower LUI. In the FS subsystem, precipitation is linked to a positive, nonlinear response of LUI, while distance to urban centers displayed a unimodal (peaked) effect—LUI is the greatest at intermediate distances and declines in both near-core and remote locations.

To assess the reliability and generalizability of the XGBoost models for the HS, CS, and FS subsystems (years 2000, 2010, 2020), three complementary diagnostics were implemented:

• Overfitting assessment. Final held-out test performance was compared with the 5-fold CV mean. Across the nine models, final–CV R² gaps were small (predominantly within ±0.011) (Table A2). Five of the nine final models fell within the CV mean 95% CI; the remaining four showed marginal deviations (two slightly above—indicative of optimistic splits; two slightly below—suggesting conservative splits).
• Uncertainty quantification. A nonparametric bootstrap (B = 10,000 resamples from the training set) was used to compute 95% confidence intervals (CIs) around CV means for R², RMSE, and MAE. CI widths were generally narrow for FS and HS (high precision). The widest interval occurred for CS-2010 (≈0.018 absolute width in R²; ~2.7% relative variability), consistent with stronger temporal variation in driver interactions that year.
• Stability of model explanations. Agreement between SHAP importance (mean absolute values) and XGBoost gain-based importance was evaluated using Spearman’s ρ. Rank concordance was high for FS (ρ ≈ 0.84–0.92), moderate to weak for HS (ρ ≈ 0.56 to 0.31, reflecting correlated predictors), and deteriorating for CS (ρ: 0.92 [2000] to−0.19 [2010] to 0.02 [2020]), indicating evolving mechanisms, increased feature interdependence/interaction, and potential zero-inflation in later years.

A consolidated summary is provided in

Table A3. Summary of XGBoost model performance, cross-validation metrics, and robustness indicators for land use intensity subsystems in Guanzhong Urban Agglomeration (2000–2020).

Exp	Final R2	CV Mean R2	Gap	95% CI of CV Mean R2	Final ∈ CI	SHAP–Gain ρ (p)
E1 CS2000	0.656	0.650	+0.006	[0.645, 0.655]	No (↑)	0.918 (6.7 × 10−5)
E2 CS2010	0.644	0.651	−0.007	[0.643, 0.660]	Yes	−0.191 (0.574)
E3 CS2020	0.592	0.597	−0.006	[0.593, 0.602]	No (↓)	0.018 (0.958)
E4 FS2000	0.720	0.711	+0.009	[0.706, 0.717]	No (↑)	0.836 (1.3 × 10−3)
E5 FS2010	0.538	0.539	−0.001	[0.536, 0.543]	Yes	0.918 (6.7 × 10−5)
E6 FS2020	0.600	0.600	~0.000	[0.597, 0.603]	Yes	0.909 (1.1 × 10−4)
E7 HS2000	0.760	0.771	−0.011	[0.765, 0.778]	No (↓)	0.564 (0.071)
E8 HS2010	0.817	0.820	−0.004	[0.816, 0.827]	Yes	0.164 (0.631)
E9 HS2020	0.808	0.808	+0.001	[0.804, 0.811]	Yes	0.309 (≈0.35)

Notes. CI bounds are 5-fold bootstrap percentiles over folds; “↑/↓” mark final above/below the CI.

. Overall, the diagnostics indicate no substantive evidence of overfitting, stable out-of-sample generalization, and interpretable driver attributions while also highlighting subsystem-specific complexities—most notably in CS—that warrant cautious interpretation.

4. Discussion

Prior urban land use research frequently aggregates indicators at administrative unit scales, obscuring intra-urban heterogeneity, including subsystem-specific policy effects and geomorphological influences such as terrain reshaping and ecosystem fragmentation [73,74]. To address this limitation, a 1 km hexagonal grid is employed to evaluate land use intensity (LUI) in the Guanzhong Plain Urban Agglomeration (GPUA) over 2000–2020, disaggregated into the human settlement (HS), cropland (CS), and forest (FS) subsystems. This design enhances spatial resolution and subsystem specificity, enabling the detection of localized geomorphological responses—e.g., slope-mediated LUI thresholds in FS—that are typically smoothed in coarser, unit-based analyses [75,76].

Methodologically, combining XGBoost with SHAP overcomes the limitations of linear regression by recovering nonlinearities, interactions, and threshold effects (e.g., income-related tipping points in HS intensification) that parametric or uninterpreted black-box models may miss [77]. SHAP provides consistent, post hoc attributions of driver importance and effect direction, refining the understanding of heterogeneous mechanisms and subsystem interplay often hidden in aggregate studies [78,79].

Empirically, LUI changes in the GPUA align with rapid urbanization and integration: HS and CS intensification coincides with heightened ecological pressures (e.g., increased erosion risk), whereas policy interventions—most notably Grain-for-Green—are associated with FS recovery and reduced disturbance [80,81,82]. These results extend coastal-focused evidence by documenting inland-specific geomorphological adaptations and inform targeted, subsystem-aware land-governance strategies for sustainability.

4.1. Synergistic Land Use Intensity Change and Policy Responses in the Guanzhong Plain Urban Agglomeration

This study advances LUI comprehension in the GPUA through high-resolution spatial analysis and interpretable machine learning, uncovering subsystem-specific trajectories shaped by policy, socioeconomic shift, and environmental factors.

In the HS subsystem, economic growth and population agglomeration have expanded high-intensity zones in urban cores and peripheries, with medium-intensity areas extending to the fringe to bolster urban–rural linkages, while low-intensity zones contract amid the rural–urban transition. Declines in central HS LUI signal counter-urbanization and functional transformation, such as renewal into ecological or service-oriented spaces [73,76].

For the CS subsystem, policy-driven shifts toward cultivation are evident: high-intensity persist in core farmlands, but medium-intensity zones grow under protection policies. Spatial–temporal volatility increases urban encroachment and resource constraints, moderated by “cropland occupation–compensation balance” and “permanent basic farmland delimitation” initiatives [75,80].

In the FS subsystem, restoration goals reduce high-intensity areas in conservation zones, expanding lower-intensity forest for enhancing overall ecosystem resilience against climate variability and human disturbances. Yet, fluctuations at the urban fringes underscore vulnerabilities, where competing pressures from development necessitate buffer-zone policies that balance timber management, recreational use, and strict preservation, ensuring forests serve as vital carbon sinks and habitat corridors [81,82].

4.2. Nonlinear Mechanisms and Threshold Effects in LUI Dynamics

By leveraging XGBoost integrated with SHAP analysis, this study provides a robust characterization of the nonlinear and threshold-dependent interactions governing LUI across subsystems; this approach not only quantifies marginal contributions of drivers but also visualizes dependence patterns, offering actionable insights into how LUI responds to varying driver intensities.

In the HS subsystem, socioeconomic drivers such as per capita GDP and population density predominate, exhibiting an inverted U-shaped relationship with LUI that indicates critical thresholds: initial growth phases accelerate urban intensification through infrastructure investments, but beyond certain economic tipping points, further increases lead to efficiency constraints, saturation, or shifts toward green urbanism due to regulatory caps and environmental feedbacks [74,79]. Urban–rural integration metrics further impose nonlinear effects, where moderate connectivity enhances LUI by facilitating resource flows, but extremes—either isolation or over-integration—can disrupt equilibrium, emphasizing the pivotal role of spatial planning in modulating these dynamics.

For the CS subsystem, the evolving balance between biophysical and socioeconomic controls reveals temporal shifts in dominance: slope and other topographic factors initially constrain LUI by limiting mechanization, but by 2020, population density and urban proximity emerged as key influencers, reflecting intensified demands for food production near markets. SHAP dependence plots illustrate an inverted U-shaped response, with LUI peaks at moderate population densities and intermediate urban distances—optimal for access to inputs and markets—before declining under excessive pressures from overcrowding or remoteness-induced abandonment [75,76].

In the FS subsystem, natural environmental factors—precipitation, elevation, and slope—remained the principal positive determinants of LUI. Conversely, indicators of human activity and accessibility (population density, GDP, road length) generally enhance forest intensity, underscoring the protective effect of ecological isolation. Crucially, a pronounced “peak effect” characterized the relationship between distance to urban centers and FS LUI: forests at moderate isolation exhibited the highest intensity, whereas those too close to urban areas suffered disturbance, and those excessively distant from them faced management challenges.

Collectively, these insights from explainable machine learning establish an empirical foundation for adaptive governance, enabling policymakers to anticipate tipping points and design interventions that reconcile LUI intensification with long-term ecological and socioeconomic sustainability.

4.3. Policy Implications for Sustainable LUI Optimization

Drawing on the synergistic LUI changes described in Section 4.1 and the nonlinear mechanisms and threshold effects elucidated in Section 4.2, policies in urbanizing regions such as the GPUA must be adaptive, evidence-based, and subsystem-specific to avert maladaptive outcomes, including surpassing economic tipping points in HS or excessive intensification in CS and FS that exacerbates ecological pressures [73,80]. This necessitates a multi-level governance framework that aligns national directives with localized execution, emphasizing resilience in the face of climatic and demographic uncertainties while utilizing high-resolution data to promote beneficial intensification and restrain it where harmful. International assessments reinforce these priorities: global projections show that unmanaged urban expansion threatens biodiversity and ecosystem services, highlighting the importance of growth management and compact or polycentric urban forms that raise land use efficiency while curbing peripheral sprawl [3,83]. At the same time, European experience cautions that densification must be paired with deliberate greening to sustain livability and ecological performance, implying an efficiency–amenity balance in mature metropolitan areas [84,85].

Where economic expansion has broadened high-intensity zones yet poses risks of saturation beyond identified thresholds, the enforcement of urban growth boundaries and incentives for infill redevelopment are crucial to boost land use efficiency, mitigate inefficient sprawl, and safeguard peripheral areas; concurrently, advancing polycentric urban networks via targeted infrastructure investments and holistic urban–rural integration initiatives can rectify spatial disparities, support equitable growth, alleviate geomorphological disturbances such as urban heat islands, and facilitate the controlled escalation of urban land intensification for sustainability [78,79]. These measures are consistent with international evidence that compact and polycentric strategies can deliver productivity gains while containing land take—provided they are complemented by urban green infrastructure to offset densification externalities [83,84,85].

Where policy-led transitions toward balanced farming practices have expanded medium-intensity zones amid volatility from urban incursion, upholding farmland redlines should be augmented by precision agriculture technologies—such as smart irrigation and soil monitoring—to elevate yields without compromising the environment. Fostering peri-urban agri-tourism and ecologically converting marginal lands (e.g., steep slopes) can additionally harmonize productivity with restoration, securing food supplies while curbing erosion and biodiversity decline, thus regulating over-intensification in line with inverted U responses to population density and urban proximity [75,76]. This aligns with international practice in sustainable intensification, which emphasizes yield increases alongside reduced externalities through agronomic redesign and targeted inputs [86,87], with precision technologies and site-specific management improving resource use efficiency near urban markets [88].

Where restoration efforts have diminished high-intensity zones to favor low-intensity forests and bolster resilience, adherence to ecological redlines ought to encompass differentiated management approaches—from urban green belts and eco-tourism corridors to rigorous conservation in remote reserves—reinforced by sustained afforestation and fair compensation mechanisms for local stakeholders; this strategy preserves ecosystem services, minimizes anthropogenic stresses, and improves habitat quality and carbon sequestration in sensitive terrains, thereby constraining forestry intensification to evade disturbance peaks linked to urban adjacency [81,82]. International evaluations of protected-area effectiveness similarly demonstrate that well-sited, well-enforced reserves reduce deforestation and disturbance pressures [89,90], while China’s large-scale restoration programs (e.g., Grain-for-Green) illustrate how tenure- and incentive-based mechanisms can deliver measurable ecological recovery at landscape scales [91].

5. Conclusions

This study delivers a high-resolution assessment of land use intensity (LUI) dynamics in the Guanzhong Plain Urban Agglomeration (GPUA) from 2000 to 2020, integrating a 1 km hexagonal grid with XGBoost-SHAP to reveal subsystem-specific patterns, nonlinear driver mechanisms, and policy-relevant thresholds. By addressing the study’s threefold objectives, our findings offer robust, evidence-based insights into sustainable land management for inland urban agglomerations.

First, regarding the spatiotemporal evolution of subsystem-specific LUI, marked redistributions were observed across HS, CS, and FS. In HS, low-intensity areas declined from 86.54% to 83.18%, with gains in medium- (12.10% to 14.26%) and high-intensity areas (1.22% to 2.56%), forming continuous high-intensity corridors (e.g., Xi’an–Xianyang) by 2020. CS shifted toward medium-intensity (32.53% to 50.57%), with high-intensity contraction (26.62% to 14.53%), exhibiting dynamism (55.1% net intensification, 38.5% net decline, 1.68% stability). FS transitioned to low-intensity dominance (59.12% by 2020), with stability (59.1%) and delayed growth (26.15%) prevailing in conserved mountainous zones. Urban–rural gradients showed HS declining from core (>20% increase) to rural areas (<0.20 stable), CS peaking in fringes (~0.60 stable), and FS reversing from inverse patterns (2000–2010 declines) to core recovery by 2020.

Second, to identify and interpret dominant drivers and heterogeneous effects via explainable machine learning, models demonstrated strong performance (HS R² = 0.76–0.82; FS 0.72 declining to 0.60; CS 0.66 to 0.59), with robustness confirmed by narrow cross-validation gaps (±0.011) and bootstrap confidence intervals. SHAP attribution highlighted HS driven by population density, per capita GDP, and urban proximity; CS shifting from slope constraints to socioeconomic and accessibility factors; and FS positively influenced by precipitation/elevation/slope but negatively by human indicators. Nonlinearities included inverted U shapes in HS (GDP thresholds) and CS (population density peaks) and unimodal peaks in FS (urban distance optima).

Third, policy recommendations for differentiated, efficient, and sustainable land management emphasize subsystem-targeted strategies: for HS, enforce growth boundaries and polycentric networks to control sprawl and raise efficiency; for CS, integrate precision technologies and redlines to balance yields with erosion mitigation; for FS, apply ecological redlines and afforestation to reduce disturbance and enhance resilience. Cross-cutting measures should leverage thresholds to anticipate tipping points, fostering integrated governance that reconciles intensification with ecological integrity. The international parallels suggest that this paper’s policy guidance has transferability beyond China to inland urban agglomerations facing similar growth–environment tensions.

While our framework provides robust, data-driven insights into LUI processes, limitations include reliance on available datasets and potential for finer-scale climate integration. Future research could incorporate higher-resolution data, additional drivers (e.g., climate change impacts), and cross-agglomeration comparisons to broaden applicability in urbanization contexts.