Next Article in Journal
The Effect of Water Treatment Processes on the Performance of Proton Exchange Membrane Water Electrolysis
Previous Article in Journal
Chromosomally Encoded Resistance and Virulence Determinants Are Selectively Enriched in Hospital Wastewater Effluent Despite Reduced Total ARG Abundance
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Water
Volume 18
Issue 10
10.3390/w18101212
water-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Landscape Controls on Coupled Water–Air Pollution in an Urbanized Watershed: A GeoSHAP Analysis of the Liaohe River Basin, China

by
Sixue Shi
,
Tingshuang Zhang
and
Miao Liu
*
CAS Key Laboratory of Forest Ecology and Silviculture, Institute of Applied Ecology, Chinese Academy of Sciences, Shenyang 110016, China
*
Author to whom correspondence should be addressed.
Water 2026, 18(10), 1212; https://doi.org/10.3390/w18101212
Submission received: 21 April 2026 / Revised: 13 May 2026 / Accepted: 14 May 2026 / Published: 17 May 2026
(This article belongs to the Section Urban Water Management)
Browse Figures
Versions Notes

Abstract

Landscape pattern is closely associated with pollution in rapidly urbanizing watersheds, but most studies still focus on single pollutants or single environmental media. This study developed a watershed-based framework to compare coupled water and air pollution in the Liaohe River Basin, China. A total of 156 hydrologically connected sub-basins were used as common spatial units. Landscape metrics were calculated for 2000, 2010, and 2020. Total nitrogen and total phosphorus loads were simulated using the Soil and Water Assessment Tool, while annual mean PM2.5 and O3 concentrations were aggregated from gridded products to the same sub-basin scale. Coupling coordination degree was used to identify relative co-pollution patterns within the aquatic and atmospheric systems. GeoXGBoost with spatial block cross-validation was used to evaluate predictive performance, and GeoSHAP was used to interpret model-based predictor contributions. The aquatic coupled pollution index was predicted more accurately than the atmospheric index, indicating a stronger landscape association with nutrient coupling. Cropland proportion was the most stable predictor of aquatic coupling, whereas forest proportion was the most stable predictor of atmospheric coupling. These results suggest that water-oriented management should focus on cropland structure and ecological buffering, while air-oriented management should emphasize forest continuity and fragmentation control. The framework provides a spatially explicit basis for differentiated watershed management and territorial spatial planning.

1. Introduction

Point-source pollution has been substantially reduced in China during the last decade. Industrial discharge standards, wastewater treatment upgrades, and strengthened regulation have lowered many conventional pollutant emissions [1,2]. Yet environmental improvement remains incomplete. In many rapidly urbanizing watersheds, nitrogen and phosphorus enrichment in rivers and reservoirs persists, while PM2.5 and O3 composite pollution continues to affect public health and regional sustainability [3,4]. This problem is also evident in the Liaohe River Basin, Liaoning section. From 2000 to 2020, mean TN load remained above 2.0 t/km2, while mean TP load declined from 0.88 to 0.61 t/km2. Over the same period, mean PM2.5 declined from 47.62 to 37.48 μg/m3, whereas mean O3 increased from 91.17 to 93.56 μg/m3. These contrasting trends indicate that coupled pollution cannot be fully understood from a single pollutant or a single environmental medium. These two problems are often studied separately, yet both are shaped by land-use composition and landscape configuration, although hydrological and atmospheric processes operate through different transport pathways. In the aquatic system, landscape pattern affects nutrient export through source distribution, runoff generation, hydrological connectivity, and ecological buffering. In the atmospheric system, landscape pattern may be associated with pollutant co-occurrence through local emission contexts, surface roughness, dry deposition, microclimatic conditions, and precursor-related land uses. Therefore, this study does not assume a shared physical mechanism for water and air pollution. Instead, it uses landscape pattern as a common spatial descriptor to compare their associations with land-use structure at the sub-basin scale.
Landscape pattern is widely used to describe the composition and spatial arrangement of land-cover types within a territorial system [5]. In the aquatic environment, landscape composition and configuration affect diffuse nutrient pollution through source distribution, runoff generation, and transport connectivity [6,7]. Studies based on the source–sink framework have shown that the spatial arrangement of cropland, forest, wetland, and built-up land can strongly influence nutrient export to receiving waters [8,9]. Related planning and management practices, such as best management practices (BMPs), low-impact development (LID), and nature-based solutions (NbS), further suggest that landscape intervention can improve water quality when ecological buffering and hydrological connectivity are properly managed [10]. In the atmospheric environment, landscape pattern also matters. Forest and grassland can reduce PM2.5 through deposition and microclimatic regulation, whereas built-up land often intensifies PM2.5 and O3 pollution through anthropogenic emissions and altered local meteorology [11,12]. Urban blue–green structure, forest continuity, and built-up morphology have therefore become important variables in air-pollution studies [13]. Taken together, these studies show that landscape patterns affect both water and air pollution, but the mechanisms are not identical across media. Existing evidence is still largely pollutant-specific and rarely comparable across environmental media at spatially consistent governance units.
This scientific gap is not merely academic. It also constrains territorial spatial planning, which requires evidence that is cross-media, unit-consistent, and spatially explicit. Historically, China’s environmental governance relied on multiple spatial plans, including environmental protection, land use, and urban master plans [14,15]. However, problems such as poor coordination among plans and fragmented spatial regulation became increasingly prominent. Since 2017, the reform of “multiple plans integration” has advanced, and territorial spatial planning has been established as the master plan for national spatial development, integrating major function-oriented zoning, land-use planning, and urban/rural planning into a unified system [16]. This reform creates an important opportunity for pattern-based ecological governance, but several research limitations remain. First, many studies still rely on administrative units and therefore do not reflect watershed-based transport structure or cross-boundary spillover [17,18]. Second, water and air pollution are usually examined separately, and integrated comparison under one landscape framework remains rare [19,20]. Third, although machine learning has improved the detection of nonlinear relationships, many studies still depend on ordinary random train–test splits and global interpretation methods, which are not well suited to spatial dependence or spatial heterogeneity [21,22]. For spatially structured watershed data, neighboring sub-basins often share similar land-use patterns, pollution levels, and geographic contexts. If these neighboring units are randomly split between training and testing sets, model performance may be overestimated, and spatial dependence may be misattributed to local predictors. Spatial block cross-validation is therefore needed to support more reliable GeoSHAP interpretation. As a result, the evidence required for territorial spatial planning is still insufficient. The novelty of this study lies in combining cross-media pollution comparison, SWAT-delineated sub-basins as common spatial units, and spatially validated interpretable machine learning within one watershed framework.
To address these gaps, we selected the Liaohe River Basin (Liaoning section, LN-LRB) in Northeast China as a case study. This region is both a major grain-production area and part of the old industrial base of China. It includes forested mountains in the east, cropland-dominated plains in the north and west, and urbanized areas in the central and southern parts of the basin. This strong spatial heterogeneity creates clear gradients in ecological background, agricultural nutrient pressure, and urban atmospheric pollution. It therefore provides a suitable setting for testing whether aquatic and atmospheric coupled pollution are associated with different landscape predictors under a common sub-basin framework. We proposed three hypotheses. First, aquatic coupled pollution would be more strongly associated with cropland-related composition metrics, because nutrient generation and transport are constrained by hydrological pathways within sub-basins. Second, atmospheric coupled pollution would show lower predictability and stronger spatial heterogeneity, because it is influenced not only by local landscape structure but also by extra-local atmospheric processes. Third, configuration metrics would show greater temporal variation than composition metrics, because configuration is more sensitive to urban expansion form, boundary reorganization, and landscape fragmentation over time. These hypotheses were tested through a watershed-based analytical framework that integrates non-point source pollution simulation, air-pollution assessment, landscape metrics, GeoXGBoost modeling, and GeoSHAP interpretation. The specific objectives were: (1) to establish an integrated framework that uses sub-basins as common analytical units for water and air pollution assessment; (2) to quantify the nonlinear and spatially heterogeneous associations of landscape metrics with coupled pollution indicators in the aquatic and atmospheric systems; (3) to identify the dominant landscape predictors and their temporal changes in each system; and (4) to derive differentiated implications for territorial spatial planning through cross-system comparison.

2. Materials and Methods

2.1. Study Area and Sub-Basin Delineation

The Liaohe River Basin (Liaoning section) is located in Northeast China and covers the core area of the Central Liaoning urban agglomeration (Figure 1). It includes a broad transition from forest-dominated mountains in the east to cropland-dominated plains in the north and west, with major urban and industrial areas concentrated in the central and southern parts of the basin. This east–west contrast creates strong differences in land-use intensity, ecological buffering capacity, and pollutant generation. Such heterogeneity makes the basin suitable for examining how landscape patterns are associated with coupled water and air pollution at the sub-basin scale. It also links the study to territorial spatial planning because the basin combines agricultural production, industrial transformation, urban expansion, and ecological protection within one watershed governance unit.
Sub-basins were delineated with the Soil and Water Assessment Tool (SWAT, Version 2012) based on the digital elevation model. Flow direction and flow accumulation were derived from the DEM, and the stream network was generated using a predefined drainage-area threshold. Watershed outlets and monitoring-station locations were used to check the drainage structure. This procedure produced 156 hydrologically connected sub-basins. The sub-basins were contiguous within the study area; their areas ranged from 112.92 to 1123.66 km2, with a median area of 380.36 km2.

2.2. Data Sources and Preprocessing

The data sources and information are presented in Table 1. The land use types were reclassified as cropland, forest land, grassland, built-up area, water area, barren, and wetland. We verified the accuracy of our classification by comparing it with ground truth data and other sources, achieving an overall accuracy of more than 90%. The climatic data included the annual average temperature, total annual precipitation, annual mean maximum temperature, and total annual evaporation. Daily climate data from 2000 to 2020 were collected from the Liaoning Meteorological Administration. Meteorological observation data were obtained from the China Meteorological Data Service Center (http://data.cma.cn/en). These data were formatted as station-based daily weather inputs for the SWAT model and used directly in the hydrological and nutrient simulations. PM2.5 and O3 concentration data were obtained from the Zenodo dataset (https://zenodo.org). Daily water quality data were obtained from the Bureau of Ecology and Environment of Liaoning Province.
All raster and gridded datasets were projected to a common coordinate system and summarized to the 156 SWAT-delineated sub-basins. Landscape metrics were calculated from 30 m land-use data. SWAT-simulated TN and TP loads were summarized at the sub-basin scale, while annual mean PM2.5 and O3 concentrations were aggregated from 1 km gridded products using zonal statistics. This preprocessing ensured that all predictors and response variables were aligned to the same spatial support before model fitting.

2.3. Study Methods

This study followed the workflow illustrated in Figure 2. First, 156 hydrological catchment units were created for mapping and analysis based on the SWAT model. Catchments are appropriate units because many ecological processes are organized along drainage networks [22,23,24]. Seven landscape metrics were calculated at both class and landscape levels for each sub-basin and each year (2000, 2010, 2020). Then, non-point source TN and TP loads were simulated using SWAT, and PM2.5 and O3 concentrations were aggregated to sub-basin scale. Watersheds are used here as common governance-support units for cross-medium comparison, rather than as physically closed atmospheric compartments. Composite pollution indices (CCDNPS and CCDAir) were constructed to capture coupled pollution hotspots. Finally, GeoXGBoost regression with spatial block cross-validation was applied to predict pollution indices from landscape metrics, and GeoSHAP was used to interpret model predictions, rank feature importance, and reveal spatial heterogeneity. These analyses were implemented in R (Version 4.5.1). The identified patterns were translated into spatial planning recommendations.

2.3.1. Landscape Metrics

To characterize territorial spatial pattern, we calculated seven commonly used landscape metrics at both landscape and class levels in Table 2. including patch density (PD), percentage of landscape (PLAND), largest patch index (LPI), edge density (ED), aggregation index (AI), Shannon’s diversity index (SHDI), and landscape shape index (LSI). These metrics represent different aspects of landscape composition, fragmentation, edge complexity, aggregation, and heterogeneity, and are widely used in ecological and planning studies. The seven base metrics were expanded into 21 predictors by calculating class-level metrics for the major land-cover classes used in interpretation (cropland, forest, and built-up land), together with selected landscape-level metrics (e.g., LPI, SHDI). Metrics were calculated for each sub-basin and for each benchmark year.

2.3.2. Aquatic and Atmospheric Pollution Indicators

The cross-system comparison was based on shared spatial support and shared landscape predictors. Landscape metrics represented source distribution, runoff pathways, and ecological buffering for the aquatic system, and represented land-surface structure, potential local emission context, deposition surfaces, and microclimatic background for the atmospheric system. Thus, the comparison should be interpreted as a landscape-based statistical comparison rather than as a unified mechanistic model of water and air pollution.
For the aquatic system, TN and TP loads were simulated using the SWAT model at the sub-basin scale. SWAT was selected because it provides a process-based representation of hydrological processes and has been widely used in watershed studies of diffuse nutrient pollution. Daily climate, land use, soil, and topographic data were integrated into the model, and outputs were summarized for the three benchmark years. The resulting TN and TP loads were used to represent the spatial pattern of nutrient pressure at the sub-basin scale.
For the atmospheric system, annual mean PM2.5 and O3 concentrations were obtained from gridded products and aggregated to the same sub-basin units. Sub-basins were used as a common landscape-governance unit rather than as physically closed atmospheric units. This choice improves cross-system comparability but may underrepresent extra-basin transport processes; therefore, atmospheric results should be interpreted as landscape associations at the sub-basin support scale rather than boundary-constrained atmospheric dynamics.
For each benchmark year (2000, 2010, 2020), we derived annual TN and TP export loads from the calibrated SWAT model and annual mean PM2.5 and O3 concentrations from gridded products, so that all response variables represented annual pollution conditions summarized under a consistent spatial support at the sub-basin scale. This comparison was based on spatial alignment, annual aggregation, and the same landscape predictor set; it did not imply physical equivalence between nutrient loads and air-pollutant concentrations.

2.3.3. Composite Pollution Index (Coupling Coordination Degree, CCD)

To capture the synergistic pollution status of two pollutants within each environmental system (aquatic: TN and TP; atmospheric: PM2.5 and O3), we employed the Coupling Coordination Degree (CCD) model. Unlike a simple average, the CCD model was used to identify areas where two pollutants are jointly elevated in relative terms, which is useful for recognizing co-pollution hotspots. The specific formula is given below:
C   =   2 ×   [ F ( X )   ×   G ( Y ) ( F ( X )   +   G ( Y ) ) 2 ] 1 2
T = α × F ( X ) + β × G ( Y )
C C D = ( C × T ) 1 2
where C is the coupling degree (0–1); F(X) and G(Y) are normalized TN (PM2.5) and TP (O3) in the aquatic and atmospheric systems, respectively; T is the comprehensive evaluation score reflecting the combined status of the two pollutants; α and β are the weights, set to 0.5 each (equal importance), assuming equal importance of both pollutants in contributing to the coupled pollution status; and CCD is the coupling coordination degree between the two pollutants, also ranging from 0 to 1. Equal weights were used because the study aimed to identify co-pollution hotspots rather than prioritize one pollutant over the other. Min-max normalization was conducted separately for each year, and CCD was therefore interpreted as a within-year relative indicator rather than as a measure of absolute interannual pollution intensity.
A higher CCD indicates that both pollutants are co-elevated in relative terms within the same year (i.e., high TN and TP loads, or high PM2.5 and O3 concentrations, after normalization), while a lower CCD indicates weaker relative co-elevation or lower coordinated pollution status. Based on the above procedure, we constructed two indicators: CCDNPS (aquatic system, using TN and TP) and CCDAir (atmospheric system, using PM2.5 and O3). Both served as response variables in the subsequent GeoXGBoost modeling. Global Moran’s I was calculated to provide a quantitative summary of the spatial autocorrelation of CCDNPS and CCDAir in each benchmark year.

2.3.4. GeoXGBoost Modeling

GeoXGBoost models were built separately for the water quality and air quality coupling coordination degrees in 2000, 2010, and 2020. GeoXGBoost was used because gradient boosting can capture nonlinear relationships and interactions between landscape metrics and coupled pollution indicators, while spatial information can be incorporated to account for broad geographic structure. Each year was modeled independently at the sub-basin scale. The response variable was the water quality coupling coordination degree, derived from total nitrogen (TN) and total phosphorus (TP), for the aquatic system, and the air quality coupling coordination degree, derived from O3 and PM2.5, for the atmospheric system. The initial predictor pool comprised 21 landscape metrics. Longitude and latitude derived from watershed polygons were also included to account for broad spatial structure. Predictor reduction was conducted in two steps.
Predictors with zero variance were first removed. We then examined pairwise correlations among landscape metrics using a reporting threshold of |r| ≥ 0.90 to identify redundancy. Strong overlap was repeatedly observed within the patch density (PD), edge density (ED), and aggregation index (AI) family. We therefore applied a fixed reduction rule at both the landscape and class levels: ED was retained as the representative configuration metric, whereas the corresponding PD and AI terms were excluded. Other metrics, including PLAND, LPI, LSI, and SHDI, were retained. This step was intended to reduce redundancy and improve interpretive parsimony while maintaining a consistent predictor set across years and between the two pollution systems.
Model performance was evaluated using spatial block cross-validation implemented in blockCV. Spatial folds were generated from watershed polygons rather than from randomly sampled records. We used five spatial folds and searched across candidate block sizes between 30 and 80 km under repeated random allocation until a complete five-fold partition was obtained. The selected spatial partition was then used for model fitting and validation. Predictive accuracy was assessed using out-of-fold R2, RMSE, and MAE. The spatial cross-validation results were used to evaluate predictive performance. After model evaluation, the final model was refitted using the full dataset for each year and pollution system, and the fitted models were then used for GeoSHAP-based interpretation. Thus, GeoXGBoost was used for prediction, whereas GeoSHAP was used to interpret model-predicted responses.

2.3.5. GeoSHAP Interpretation and Comparative Analysis

After the final GeoXGBoost model was fitted for each year and pollution system, GeoSHAP values were calculated to quantify the contribution of each predictor to the modeled coupling coordination degree. In this study, GeoSHAP was used because it provides both global importance rankings and local predictor contributions for spatially heterogeneous model predictions. SHAP values were interpreted as predictor contributions to model predictions, not as causal effects. Global importance was summarized using the mean absolute SHAP value. Longitude and latitude were retained in the model to account for broad spatial structure, but they were excluded from the main ecological interpretation so that the explanatory ranking focused on landscape variables.
GeoSHAP values were also mapped to characterize spatial heterogeneity in predictor contributions across sub-basins. To assess the direction of model association, we calculated the Spearman correlation between predictor values and their corresponding SHAP values. Positive correlations indicate that higher predictor values tend to be associated with higher predicted coupling coordination degree, whereas negative correlations indicate the opposite.
GeoSHAP dependence plots were used to examine nonlinear model associations between landscape metrics and the predicted coupling coordination degree. The transition intervals reported in this study were identified from fitted dependence curves as ranges where SHAP contributions changed direction, crossed zero, or showed clear slope attenuation. These intervals were treated as exploratory model-based sensitivity ranges rather than statistically validated ecological thresholds.
GeoSHAP outputs were summarized at three levels: temporal changes within the water system, temporal changes within the atmospheric system, and cross-system contrasts between the two coupling coordination degrees. This framework was used to identify stable dominant predictors, shifting secondary contributors, and their spatial expression.

3. Results

3.1. Landscape Pattern Dynamics

Between 2000 and 2020, the Liaohe River Basin experienced substantial changes in land-use structure and landscape configuration (Figure 3). Cropland area decreased from 35,033.1 to 33,595.2 km2, built-up land increased from 4861.2 to 6441.0 km2, and forest area remained nearly stable at about 21,100–21,300 km2. Grassland and wetland also decreased slightly, whereas water area changed little. Spatially, forest land remained concentrated in the eastern mountains, cropland dominated the western and northern plains, and built-up land clustered in the central and southern urbanized zones. During the same period, fragmentation increased at the landscape level, as mean PD increased from 0.501 to 0.601 and mean ED increased from 20.29 to 22.64, while mean AI decreased slightly from 96.87 to 96.52. These changes suggest that the basin underwent both compositional adjustment and configurational reorganization during rapid urbanization.
From an ecological perspective, the most important landscape shift was not the small increase in total forest area, but the rising fragmentation of forest and cropland patches in zones of urban expansion. This suggests that the modeled pollution patterns may be associated not only with land-use conversion itself, but also with increasing interface complexity among cropland, built-up land, and ecological land. Such configuration changes provided the landscape background for the coupled pollution responses examined below.

3.2. CCD Analysis

The two CCD indicators showed clear spatial structure in the Liaohe River Basin (Figure 4 and Figure 5). For the aquatic system, TN and TP loads showed different temporal trajectories but remained spatially associated. Mean TN load increased from 2.04 t/km2 in 2000 to 2.37 t/km2 in 2010 and then decreased slightly to 2.27 t/km2 in 2020. In contrast, mean TP load declined from 0.88 to 0.33 t/km2 between 2000 and 2010, before increasing to 0.61 t/km2 in 2020. Correspondingly, the mean CCDNPS increased slightly from 0.671 in 2000 to 0.691 in 2010, but declined to 0.602 in 2020. Spatially, high TN and TP loads were mainly concentrated in the western and northern parts of the basin. The corresponding CCDNPS hotspots were mainly distributed in the central-western sub-basins. The bivariate Moran’s I values for the spatial association between TN and TP loads were 0.536, 0.499, and 0.443 in 2000, 2010, and 2020, respectively, and all passed significance tests. These results indicate a persistent positive spatial association between TN and TP loads, supporting the identification of aquatic co-pollution clusters.
For the atmospheric system, PM2.5 first increased and then decreased, whereas O3 showed the opposite trend. Their co-occurrence pattern was better captured by CCDAir than by either pollutant alone. The mean CCDAir remained almost unchanged between 2000 and 2010, increasing only from 0.686 to 0.687, but then rose to 0.731 in 2020. High PM2.5–O3 coupling was mainly concentrated in the central and southern urbanized sub-basins. Despite opposite interannual trends in the two pollutants, their high-value co-occurrence still showed spatial aggregation in the central-southern urbanized sub-basins, and CCDAir also showed a clustered distribution. The Global Moran’s I values for CCDAir were 0.727, 0.693, and 0.490 in 2000, 2010, and 2020, respectively, and all passed significance tests. This indicates that atmospheric co-pollution showed significant positive spatial autocorrelation across the three benchmark years, although the degree of clustering weakened by 2020.
Overall, both coupled pollution systems were spatially organized, but their temporal and spatial patterns differed. Aquatic coupling increased slightly before declining, while atmospheric coupling became stronger by 2020. Spatially, aquatic co-pollution hotspots were mainly associated with central-western and northwestern agricultural sub-basins, whereas atmospheric co-pollution hotspots were concentrated in central-southern urbanized sub-basins. This contrast is consistent with the broad landscape transition from the forested ecological background in the east to cropland and urban land in the west and south.

3.3. GeoXGBoost Performance and GeoSHAP Interpretation of the Water and Air Quality Coupling Coordination Degrees

3.3.1. Landscape Predictors Showed Stronger Model Performance for the Aquatic System than for the Atmospheric System

The GeoXGBoost models retained clear predictive power under spatial block cross-validation, but predictive strength differed between the two systems. The water quality coupling coordination degree was predicted more accurately and more consistently than the air quality coupling coordination degree. In the water system, out-of-fold R2 values were 0.934, 0.884, and 0.942 in 2000, 2010, and 2020, with RMSE values of 0.066, 0.079, and 0.052. In the atmospheric system, the corresponding R2 values were 0.842, 0.784, and 0.818, with RMSE values of 0.081, 0.090, and 0.086. These differences indicate that landscape pattern had a tighter and more spatially coherent relationship with TN–TP coupling than with O3–PM2.5 coupling. In ecological terms, nutrient coupling remained closely tied to the internal source–sink structure of the sub-basin, whereas atmospheric coupling was associated with a broader set of processes that included, but was not limited to, local landscape structure.

3.3.2. Stable Cropland and Forest Predictors Characterized the Aquatic and Atmospheric Systems

The two systems differed first in their dominant predictor structure (Figure 6). In the water system, cropland proportion (PLAND_C) ranked first in all three years. Built-up edge density (ED_B) remained the second-ranked predictor, while forest proportion (PLAND_F), built-up proportion (PLAND_B), and landscape dominance (LPI) formed a smaller but persistent secondary layer of control. Taken together with the threshold behavior and SHAP spatial patterns, this ranking suggests that the water system was likely structured around a stable agricultural matrix. The leading role of PLAND_C did not simply reflect greater cropland area. It suggests that cropland proportion was associated with the spatial coincidence of nutrient generation, runoff production, and delivery pathways, thereby contributing to the modeled co-elevation of TN and TP within the same sub-basin. ED_B then modified that structure by changing interface density and hydrological connectivity between built-up land, cropland, and receiving waters. PLAND_F acted in the opposite direction, consistent with the buffering role of ecological land.
The atmospheric system followed a different pattern. Forest proportion (PLAND_F) ranked first in 2000, 2010, and 2020, but the secondary contributors reorganized more clearly through time. Built-up proportion (PLAND_B) was the second-ranked predictor in 2000, whereas cropland proportion (PLAND_C) rose to second place in 2010 and 2020. Edge-related metrics, especially ED and ED_F, remained influential, and shape-related complexity also entered the leading group in some years. This structure indicates that the modeled PM2.5–O3 coupling was not associated with a single fixed landscape type. Instead, it was most consistently associated with a relatively stable ecological background and then secondarily related to changing combinations of urban land, cropland, and interface structure. In other words, the water system showed a matrix-related pattern, whereas the atmospheric system showed a forest-background association with more variable secondary predictors.

3.3.3. Nonlinear Response Intervals in the Fitted Curves Showed Changing Landscape Associations

The fitted response curves for the water system showed clear nonlinear patterns rather than simple monotonic effects (Figure 7). Several predictors displayed relatively stable transition intervals across the three benchmark years. PLAND_C shifted from negative to positive SHAP contributions at roughly 45–55% cropland cover. This interval can be interpreted as a transition zone in the fitted response, suggesting that the modeled water pollution coupling became more sensitive once cropland approached a dominant share of the sub-basin area. Below this interval, the contribution of cropland was weaker and more context-dependent. Above it, the fitted response became more strongly positive, which is consistent with a more continuous agricultural source background. ED_B showed a rise-and-saturation pattern, with SHAP values increasing rapidly and then leveling off around 5–8. This curve shape suggests that increasing built-up edge density was associated with stronger nutrient coupling at lower values, but that the marginal effect weakened after interface density reached a relatively high level. PLAND_F showed an increasingly negative contribution above about 20–30%, suggesting that forest cover was associated with stronger buffering once ecological land reached a more continuous presence within the sub-basin. In 2010 and 2020, LPI also became positive above about 60–70, indicating that stronger landscape dominance was associated with higher modeled coupling after a dominant land-cover structure had emerged. Taken together, these fitted transition intervals suggest that the water-system response was jointly structured by source continuity, transfer connectivity, and ecological buffering.
The fitted response curves for the atmospheric system were more nonlinear and less regular than those of the water system (Figure 8). PLAND_F remained negative across the observed range, and its SHAP values declined more steeply near 50–60% forest cover. This transition interval suggests that the association between forest background and lower modeled atmospheric coupling became more apparent when forest occupied a more continuous share of the landscape. Compared with the water system, this higher transition interval is consistent with the idea that landscape associations with atmospheric coupling may be more evident under a broader and more continuous ecological background than hydrological buffering. In 2000, PLAND_B increased rapidly at low values and then stabilized, showing a rise-and-plateau pattern in the fitted response. This suggests that the modeled atmospheric system was especially sensitive to early built-up expansion. In 2010 and 2020, PLAND_C showed a similar pattern, implying that cropland became more strongly associated with atmospheric coupling in the later years. ED and ED_F showed turning zones around 15–22 m/ha, beyond which SHAP values declined more strongly. These intervals can be interpreted as fitted turning zones where the modeled association shifted from moderate landscape heterogeneity toward a pattern more consistent with fragmentation-related loss of ecological continuity. In other words, moderate edge density may still reflect a mixed landscape structure that supports ecological regulation, whereas higher edge density may correspond to reduced background continuity and a weaker regulating effect. Overall, the atmospheric response curves suggest model-based associations involving ecological background continuity and interface complexity, rather than a simple linear relationship of any single landscape variable.

3.3.4. Dominant Landscape Predictors Were Expressed as Persistent Basin-Wide Gradients

The dominant controls identified above were not spatially random (Figure 9). In the water system, PLAND_C remained the top-ranked predictor in all three years, and its contribution followed a highly stable east–west structure. Positive SHAP contributions were concentrated in the western and central-western sub-basins, while negative effects were concentrated in the more forested eastern mountain areas. This pattern indicates that cropland acted most strongly where agricultural land already formed a coherent landscape matrix. In the atmospheric system, PLAND_F also remained spatially stable through time. Negative contributions were concentrated in the eastern forest belt, whereas more positive contributions were distributed across central and western sub-basins with weaker forest background. These two dominant SHAP surfaces therefore define two different but equally persistent basin-wide gradients: an agricultural source gradient in the water system and a forest-background association gradient in the atmospheric system.
Overall, both systems were sensitive to landscape pattern, but the structure of their fitted responses differed. For the water quality coupling coordination degree, cropland-related composition showed the most stable association across years, while built-up edge structure and forest-related variables acted as secondary contributors. For the air quality coupling coordination degree, forest-related composition remained the most stable leading factor, whereas cropland, built-up land, and edge-related configuration showed greater temporal reorganization. These differences suggest that the same landscape metric may play different roles in the two environmental media. They also indicate that the turning intervals observed in the fitted curves are better interpreted as model-based transition zones than as generic threshold values for planning.

4. Discussion

4.1. Cropland Dominance Was Associated with Aquatic Coupling Through Source Continuity and Delivery Efficiency

The long-term dominance of cropland proportion in the water system indicates that TN–TP coupling was most consistently associated with the continuity of agricultural source areas, not merely with the presence of cropland itself. In a sub-basin where cropland occupies a limited share of the landscape, nutrient generation remains spatially important but is still moderated by ecological land, topographic interruption, and heterogeneity in runoff pathways [23,24]. Once cropland approaches or exceeds about half of the basin area, that balance changes. Cropland no longer functions as one landscape component among several; it becomes a dominant source matrix within which nutrient generation, runoff formation, and pollutant transfer are spatially aligned [25,26]. This helps explain why the strongest shift in model response occurred at approximately 45–55% cropland cover. That interval represents a model-based transition interval from mixed landscapes to source-dominated agricultural landscapes.
This interpretation also helps explain why the water system was more predictable than the atmospheric system. Hydrological export is constrained by topography, channel networks, and sub-basin boundaries. As a result, once a dominant source matrix is established, nutrient coupling tends to remain spatially coherent. The strong and repeated role of ED_B further supports this interpretation. Built-up edge density likely captured the reorganization of flow interfaces among impervious surfaces, cropland, and drainage corridors [27]. At low to moderate levels, increasing edge density would be expected to accelerate runoff concentration and shorten transfer pathways [28]. The saturation observed around 5–8 suggests that once a fitted transition interval of interface connectivity is reached, the system already functions as a highly connected transfer network, and further edge growth adds little new hydrological effect [29]. In this sense, built-up edge density was not a substitute for cropland dominance, but a modifier of the efficiency with which cropland-derived nutrients were delivered.
Forest proportion played a different role [30,31]. Its increasingly negative contribution above about 20–30% suggests that forest buffering became more apparent in the model only after ecological land reached sufficient continuity to alter runoff partitioning and pollutant interception at the basin scale. This transition interval is lower than the corresponding interval in the atmospheric system, which is also ecologically sensible. In watersheds, relatively modest forest presence can already modify infiltration, sediment retention, and slope-to-channel connectivity. The water system therefore reflects a source–transfer–buffer logic: cropland represents the strength of nutrient generation, built-up edges alter transfer efficiency, and forest cover is associated with lower coupled nutrient export once buffering land becomes sufficiently continuous.

4.2. Landscape Associations with the Atmospheric Coupled Pollution System

The atmospheric system showed a different ecological association pattern. Forest proportion remained the most important predictor in all three years, but its interpretation should be linked to broader land-surface conditions rather than to land-cover composition alone. In the atmospheric context, PLAND_F represents a background field. Forest may be associated with pollutant dry deposition, surface roughness, turbulence, temperature, humidity, and near-surface mixing conditions [32]. Its stable dominance therefore suggests that the PM2.5–O3 coupling was consistently associated with a relatively persistent ecological background rather than with any single source type. This helps explain why the forest transition interval occurred at a higher level, around 50–60%. Local or scattered forest patches may improve microenvironmental conditions, but a stronger landscape association with lower PM2.5-O3 coupling may require a broader and more continuous surface background before this pattern becomes evident in the model [33,34]. The transition interval therefore suggests a shift from local ecological influence on basin-scale atmospheric moderation.
The secondary predictors were more dynamic because the atmospheric system integrates several processes at once. The shift from PLAND_B in 2000 to PLAND_C in 2010 and 2020 suggests that modeled atmospheric coupling was not associated with one fixed landscape source. Early in the period, urban expansion and built-up concentration may have been more closely associated with precursor environments and pollutant co-occurrence [35,36]. Later, cropland background became more important, which may reflect stronger associations with agricultural emissions, dust, biomass burning, or broader land-surface effects once some urban emissions were reduced or spatially restructured [37]. This interpretation is consistent with recent evidence from the Central Liaoning urban agglomeration, where PM2.5 decreased during 2015–2020, O3 showed spatially banded high-value areas, and both pollutants were influenced by regional or short-distance transport. This change should not be interpreted as direct evidence that agricultural influence replaced urban influence. Rather, it indicates a rebalancing of the secondary source environment under a stable forest-background association.
The edge-related transition intervals deepen this interpretation. In both ED and ED_F, the turning zones around 15–22 suggest that modeled atmospheric coupling was associated with some degree of landscape heterogeneity, but not with unlimited fragmentation. Moderate edge density may be consistent with greater local mixing or contact between ecological and developed surfaces [38]. Beyond a certain point, however, fragmentation may reduce forest continuity, weaken the integrity of ecological sinks, and increase the instability of the background field [39,40]. The atmospheric system therefore appears to be associated with a balance between background continuity and interface complexity. This is why its response curves are more curved, more context-dependent, and less monotonic than those in the water system.

4.3. Different Transport Domains Were Associated with Different Predictor Hierarchies and Transition Intervals

The most important ecological contrast in this study lies not only in which variables ranked first, but in why the two systems organized themselves differently. The water system behaved as a more spatially bounded source-transfer system. Because nutrient export follows constrained hydrological pathways, landscape composition can directly organize source continuity and transfer efficiency within the same natural unit [41]. This produces a relatively stable hierarchy, a relatively clear source–buffer structure, and transition intervals that correspond to matrix transition or buffering activation.
The atmospheric system behaved more like a landscape–background association system. Landscape metrics still mattered, but they were linked to pollutant coupling through multiple intermediate processes, including deposition, turbulence, thermal conditions, precursor environments, and extra-local transport [42]. This produces a less stable secondary hierarchy and transition intervals that correspond more to the maintenance or collapse of ecological background structure than to simple source expansion. This difference also explains the model performance contrast. Landscape metrics describe hydrologically organized nutrient processes more directly than they describe the full atmospheric system, so stronger prediction in the water system is consistent with the more spatially bounded nature of nutrient export. However, the difference between the water-related transition interval for forest cover (about 20–30%) and the atmospheric interval (about 50–60%) should not be interpreted only as an ecological contrast. It may also reflect differences in response-variable construction. TN and TP were represented as SWAT-simulated loads, whereas PM2.5 and O3 were represented as gridded concentrations aggregated to sub-basins. In addition, CCD was calculated from year-specific normalized variables, and the aquatic models had higher predictive performance than the atmospheric models. Therefore, the two intervals are best interpreted as system-specific model sensitivity ranges rather than directly comparable ecological thresholds.
This distinction also helps explain why the same metric family cannot be assigned the same planning meaning across media. Cropland proportion in the water system signals source dominance and delivery risk. Cropland proportion in the atmospheric system signals a broader source environment whose association depends on forest background and interface conditions. Forest proportion in the water system acts mainly as a hydrological and biogeochemical buffer. Forest proportion in the atmospheric system acts as a background landscape predictor. Edge-related metrics in the water system modify runoff linkage, whereas in the atmospheric system they indicate the balance between useful heterogeneity and excessive fragmentation. A central contribution of this study is therefore not only that landscape pattern matters in both systems, but that the ecological meaning of landscape pattern is medium-specific.

4.4. Implications for Territorial Spatial Planning

A major practical implication of these findings is that planning should not rely only on whether a predictor has a positive or negative association. The more useful information lies in the intervals where the fitted model response changes rapidly. In the water system, cropland cover of near 50%, built-up edge density of around 5–8, and forest cover of around 20–30 define model-based high-sensitivity ranges. These are the ranges where sub-basins are most likely to shift from buffered mixed systems to source-dominated nutrient systems. In the atmospheric system, forest cover around 50–60 and edge-density values around 15–22 define model-based transition zones between relatively stable ecological background and increasingly fragmented atmospheric environments. These are not universal thresholds in a mechanistic sense, but they are operationally meaningful ranges within this basin and time frame. They identify where marginal changes in landscape pattern are likely to produce the largest change in model-predicted system response. These recommendations should be understood as planning heuristics derived from model-identified high-sensitivity intervals and spatial hotspots, rather than as prescriptive regulatory thresholds.
Based on the dominant environmental pressure and landscape metrics, three management zones are delineated: a water management zone targeting agricultural source control (PLAND_C transition intervals), an air management zone focusing on forest expansion and connectivity (PLAND_F and AI_F targets), and an ecological protection zone where forest fragmentation should be minimized (AI_F > 97, ED_F < 15). The specific planning interventions for each zone—including BMPs, cropland restructuring, forest connectivity improvement, and protection measures—are summarized in Figure 10.
This makes differentiated planning unavoidable. Water-oriented management should prioritize sub-basins where cropland is approaching or exceeding the matrix-transition range and where built-up interfaces are intensifying transfer efficiency. In these areas, reducing source continuity and restoring ecological buffers may have the greatest effect. Air-oriented management should focus more on preserving or rebuilding continuous forest background and preventing edge complexity from crossing into excessive fragmentation. In practice, this means that “more green” is not always an adequate planning objective. What matters is whether forest forms a sufficiently continuous background to be associated with low atmospheric coupling, and whether landscape edges remain within a range that preserves ecological function rather than eroding it.
For this reason, a unified water–air planning target is unlikely to work. The water system requires regulation of source structure and transfer pathways. The atmospheric system requires maintenance of ecological background and control of fragmentation. Their overlap lies in the need to treat landscape pattern as a functional ecological regulator rather than as a descriptive land-use statistic. That is the central planning lesson emerging from this study.

4.5. Limitations and Future Directions

Several limitations should be acknowledged. First, the models were built at the sub-basin scale (mean area 420 km2), which is suitable for watershed governance but may smooth finer within-basin heterogeneity. Future studies should explore multi-scale analysis (e.g., grid, sub-basin, and whole basin) as recommended by previous work [17]. Second, longitude and latitude were included to account for spatial structure, but they can also absorb part of the broad regional trend. Alternative approaches such as spatial eigenvector mapping could be tested in future research. Third, although the fixed ecological-cluster rule reduced redundancy among closely related metrics, cross-cluster correlation may remain. Fourth, this study focused on three benchmark years (2000, 2010, 2020). Future research should extend the framework to continuous time series, seasonal dynamics, and stronger integration with climate and emission data. Finally, the causality of identified relationships remains to be tested with process-based models or experimental designs. A further limitation is that the atmospheric models relied primarily on landscape metrics and broad spatial coordinates, without explicit emission inventories, meteorological transport variables, boundary-layer information, or atmospheric chemical processes. Therefore, the weaker and more dynamic atmospheric responses identified here should be interpreted as landscape-conditioned associations rather than full-system or mechanistic atmospheric explanations. Similarly, GeoSHAP-derived transition intervals should be treated as model-based sensitivity ranges within this basin and time frame, rather than as universal ecological thresholds.

5. Conclusions

This study developed a spatially explicit framework to examine how landscape pattern is associated with coupled water and air pollution in the Liaohe River Basin. By integrating watershed-based pollution indicators, GeoXGBoost, and GeoSHAP, the study showed that landscape pattern retained substantial predictive relevance for both CCDNPS and CCDAir under strict spatial validation.
The two systems showed distinct predictor structures. The aquatic coupled pollution system was consistently associated with cropland structure, while forest background, built-up edge density, and landscape organization acted as secondary contributors. The atmospheric coupled pollution system was more dynamic, with forest background remaining the most stable leading predictor, and cropland, built-up land, and edge-related variables changing in importance over time.
These results indicate that coordinated water–air governance should be differentiated by environmental medium. For aquatic management, priority should be given to cropland structure and its spatial relationship with ecological buffering land. For atmospheric management, coordinated management of forest background, built-up land, and landscape configuration is more important. More broadly, this study demonstrates that GeoSHAP can provide a useful model-interpretation tool for linking spatial ecological analysis with territorial spatial planning in rapidly urbanizing watersheds.

Author Contributions

Conceptualization, M.L. and T.Z.; methodology, S.S.; software, S.S.; validation, S.S. and T.Z.; formal analysis, M.L. and T.Z.; investigation, S.S. and T.Z.; resources, S.S.; data curation, S.S. and T.Z.; writing—original draft preparation, S.S. and T.Z.; writing—review and editing, M.L. and T.Z.; visualization, T.Z.; supervision, M.L.; project administration, S.S.; funding acquisition, S.S. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Liaoning Provincial Applied Basic Research Program, grant number 2025JH2/101330015, and the Liaoning Provincial Natural Science Foundation, grant number 2025-BS-0188. The APC was funded by the Liaoning Provincial Applied Basic Research Program, grant number 2025JH2/101330015.

Data Availability Statement

The datasets generated and analyzed during this study are available from the corresponding author upon reasonable request. The request should be accompanied by a clear and legitimate academic or research justification, and the corresponding author will review and respond to such requests in a timely and appropriate manner, adhering to the principles of academic transparency and cooperation.

Acknowledgments

During manuscript preparation, the authors used OpenAI (ChatGPT-5.5) for language polishing and improvement of readability. No generative AI tools were used to produce research data, perform analyses, or draw scientific conclusions. All text was reviewed and approved by the authors, who take full responsibility for the final content of the manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Bryan, B.A.; Gao, L.; Ye, Y.; Sun, X.; Connor, J.D.; Crossman, N.D.; Stafford-Smith, M.; Wu, J.; He, C.; Yu, D.; et al. China’s response to a national land-system sustainability emergency. Nature 2018, 559, 193–204. [Google Scholar] [CrossRef]
  2. Zhang, Q.; Zheng, Y.; Tong, D.; Shao, M.; Wang, S.; Zhang, Y.; Xu, X.; Wang, J.; He, H.; Liu, W.; et al. Drivers of improved PM2.5 air quality in China from 2013 to 2017. Proc. Natl. Acad. Sci. USA 2019, 116, 24463–24469. [Google Scholar] [CrossRef]
  3. Wu, J.; Wang, Y.; Liang, J.; Yao, F. Exploring common factors influencing PM2.5 and O3 concentrations in the Pearl River Delta: Tradeoffs and synergies. Environ. Pollut. 2021, 285, 117138. [Google Scholar] [CrossRef] [PubMed]
  4. Chen, L.; Dai, Y.; Zhi, X.; Xie, H.; Shen, Z. Quantifying nonpoint source emissions and their water quality responses in a complex catchment: A case study of a typical urban-rural mixed catchment. J. Hydrol. 2018, 559, 110–121. [Google Scholar] [CrossRef]
  5. Turner, M.G.; O’Neill, R.V.; Gardner, R.H.; Milne, B.T. Effects of changing spatial scale on the analysis of landscape pattern. Landsc. Ecol. 1989, 3, 153–162. [Google Scholar] [CrossRef]
  6. Xue, F.; Yang, Y.R.; Hu, Y.N. Integrating supply and benefits of ecosystem services into ecological security pattern: A spatial flow perspective. Ecol. Indic. 2024, 166, 112455. [Google Scholar] [CrossRef]
  7. Ding, J.; Jiang, Y.; Liu, Q.; Hou, Z.; Liao, J.; Fu, L.; Peng, Q. Influences of the land use pattern on water quality in low-order streams of the Dongjiang River Basin, China: A multi-scale analysis. Sci. Total Environ. 2016, 551, 205–216. [Google Scholar] [CrossRef]
  8. Zheng, S.; Zhou, X.; Singh, R.P.; Wu, Y.; Ye, Y.; Wu, C. The spatiotemporal distribution of air pollutants and their relationship with land-use patterns in Hangzhou City, China. Atmosphere 2017, 8, 110. [Google Scholar] [CrossRef]
  9. Wang, R.J.; Eckelman, M.J.; Zimmerman, J.B.; Li, X.; Ma, D.; Sun, P.; Sun, G. Discussing on “source-sink” landscape theory and phytoremediation for non-point source pollution control in China. Environ. Sci. Pollut. Res. 2020, 27, 44797–44806. [Google Scholar] [CrossRef] [PubMed]
  10. Boano, F.; Caruso, A.; Costamagna, E.; Ridolfi, L.; Fiore, S.; Demichelis, F.; Galvão, A.; Pisoeiro, J.; Rizzo, A.; Masi, F. A review of nature-based solutions for greywater treatment: Applications, hydraulic design, and environmental benefits. Sci. Total Environ. 2020, 711, 134731. [Google Scholar] [CrossRef]
  11. Shen, Z.; Zhang, Z.; Cui, L.; Xia, Z.; Zhang, Y. Coordinated change of PM2.5 and multiple landscapes based on spatial coupling model: Comparison of inland and waterfront cities. Environ. Impact Assess. Rev. 2023, 102, 107194. [Google Scholar] [CrossRef]
  12. Yang, L.J.; Wang, S.; Hu, X.; Shi, T. County-based PM2.5 concentrations’ prediction and its relationship with urban landscape pattern. Processes 2023, 11, 704. [Google Scholar] [CrossRef]
  13. Yang, H.O.; Leng, Q.; Xiao, Y.; Chen, W. Investigating the impact of urban landscape composition and configuration on PM2.5 concentration under the LCZ scheme: A case study in Nanchang, China. Sustain. Cities Soc. 2022, 84, 104006. [Google Scholar] [CrossRef]
  14. Liu, Y.; Fu, B.; Zhao, W.; Wang, S.; Deng, Y. A solution to the conflicts of multiple planning boundaries: Landscape functional zoning in a resource-based city in China. Habitat Int. 2018, 77, 43–55. [Google Scholar] [CrossRef]
  15. Fan, J.; Wang, Y.; Wang, C.; Chen, T.; Jin, F.; Zhang, W.; Li, L.; Xu, Y.; Dai, E.; Tao, A.; et al. Reshaping the sustainable geographical pattern: A major function zoning model and its applications in China. Earth’s Future 2019, 7, 25–42. [Google Scholar] [CrossRef]
  16. Liu, Y.; Zhou, Y. Territory spatial planning and national governance system in China. Land Use Policy 2021, 102, 105288. [Google Scholar] [CrossRef]
  17. Xia, H.; Yuan, S.F.; Prishchepov, A.V. Spatial-temporal heterogeneity of ecosystem service interactions and their social-ecological drivers: Implications for spatial planning and management. Resour. Conserv. Recycl. 2023, 189, 106767. [Google Scholar] [CrossRef]
  18. Huang, X.F.; Hua, W.W.; Dai, X.Y. Performance evaluation of watershed environment governance: A case study of Taihu Basin. Water 2022, 14, 158. [Google Scholar] [CrossRef]
  19. Locke, K.A. Modelling relationships between land use and water quality using statistical methods: A critical and applied review. J. Environ. Manag. 2024, 362, 121290. [Google Scholar] [CrossRef] [PubMed]
  20. Li, Y.; Huang, L.; Li, S.; Cao, M.; Tan, P.; Wang, Q.; Meng, H.; Yin, S.; Zhang, W. The role of urban green space landscape patterns in the synergistic prevention of PM2.5 and ozone pollution: A case study in Shenyang City, China. Atmos. Pollut. Res. 2024, 15, 102278. [Google Scholar] [CrossRef]
  21. Li, Z.Q. Extracting spatial effects from machine learning model using local interpretation method: An example of SHAP and XGBoost. Comput. Environ. Urban Syst. 2022, 96, 101845. [Google Scholar] [CrossRef]
  22. Roberts, D.R.; Bahn, V.; Ciuti, S.; Boyce, M.S.; Elith, J.; Guillera-Arroita, G.; Hauenstein, S.; Lahoz-Monfort, J.J.; Schröder, B.; Thuiller, W.; et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 2017, 40, 913–929. [Google Scholar] [CrossRef]
  23. Huang, J.; Li, Q.; Pontius, R.G., Jr.; Klemas, V.; Hong, H. Detecting the dynamic linkage between landscape characteristics and water quality in a subtropical coastal watershed, Southeast China. Environ. Manag. 2013, 51, 32–44. [Google Scholar] [CrossRef] [PubMed]
  24. Tu, J. Spatially varying relationships between land use and water quality across an urbanization gradient explored by geographically weighted regression. Appl. Geogr. 2011, 31, 376–392. [Google Scholar] [CrossRef]
  25. De Mello, K.; Valente, R.A.; Randhir, T.O.; dos Santos, A.C.A.; Vettorazzi, C.A. Effects of land use and land cover on water quality of low-order streams in Southeastern Brazil: Watershed versus riparian zone. Catena 2018, 167, 130–138. [Google Scholar] [CrossRef]
  26. Clément, F.; Ruiz, J.; Rodríguez, M.A.; Blais, D.; Campeau, S. Landscape diversity and forest edge density regulate stream water quality in agricultural catchments. Ecol. Indic. 2017, 72, 627–639. [Google Scholar] [CrossRef]
  27. Kaushal, S.S.; Belt, K.T. The urban watershed continuum: Evolving spatial and temporal dimensions. Urban Ecosyst. 2012, 15, 409–435. [Google Scholar] [CrossRef]
  28. Bell, C.D.; Tague, C.L.; McMillan, S.K. Modeling runoff and nitrogen loads from a watershed at different levels of impervious surface coverage and connectivity to storm water control measures. Water Resour. Res. 2019, 55, 2690–2707. [Google Scholar] [CrossRef]
  29. Oswald, C.J.; Kelleher, C.; Ledford, S.H.; Hopkins, K.G.; Sytsma, A.; Tetzlaff, D.; Toran, L.; Voter, C. Integrating urban water fluxes and moving beyond impervious surface cover: A review. J. Hydrol. 2023, 618, 129188. [Google Scholar] [CrossRef]
  30. Teixeira, Z.; Teixeira, H.; Marques, J.C. Systematic processes of land use/land cover change to identify relevant driving forces: Implications on water quality. Sci. Total Environ. 2014, 470, 1320–1335. [Google Scholar] [CrossRef]
  31. Lee, S.W.; Hwang, S.J.; Lee, S.B.; Hwang, H.S.; Sung, H.C. Landscape ecological approach to the relationships of land use patterns in watersheds to water quality characteristics. Landsc. Urban Plan. 2009, 92, 80–89. [Google Scholar] [CrossRef]
  32. Sicard, P.; Agathokleous, E.; Araminiene, V.; Carrari, E.; Hoshika, Y.; De Marco, A.; Paoletti, E. Should we see urban trees as effective solutions to reduce increasing ozone levels in cities? Environ. Pollut. 2018, 243, 163–176. [Google Scholar] [CrossRef] [PubMed]
  33. Zhai, C.; Bao, G.; Zhang, D.; Sha, Y. Urban forest locations and patch characteristics regulate PM2.5 mitigation capacity. Forests 2022, 13, 1408. [Google Scholar] [CrossRef]
  34. Zhang, X.D.; Du, J.; Zhang, L.M.; Huang, T.; Gao, H.; Mao, X.X.; Ma, J.M. Impact of afforestation on surface ozone in the North China Plain during the three-decade period. Agric. For. Meteorol. 2020, 287, 107979. [Google Scholar] [CrossRef]
  35. Duan, Z.X.; Yang, Y.; Wang, L.; Liu, C.; Fan, S.; Chen, C.; Tong, Y.; Lin, X.; Gao, Z. Temporal characteristics of carbon dioxide and ozone over a rural-cropland area in the Yangtze River Delta of eastern China. Sci. Total Environ. 2021, 757, 143750. [Google Scholar] [CrossRef]
  36. Lu, D.B.; Xu, J.; Yue, W.; Mao, W.; Yang, D.; Wang, J. Response of PM2.5 pollution to land use in China. J. Clean. Prod. 2020, 244, 118741. [Google Scholar] [CrossRef]
  37. Zhang, Q.; He, K.B.; Huo, H. Cleaning China’s air. Nature 2012, 484, 161–162. [Google Scholar] [CrossRef] [PubMed]
  38. Xie, W.D.; Wu, J.S. Effects of land use and urban landscape pattern on PM2.5 concentration: A Shenzhen case study. Acta Sci. Nat. Univ. Pekin. 2017, 53, 160–170. [Google Scholar]
  39. Lin, F.Y.; Chen, X.W. Effects of landscape patterns on atmospheric particulate matter concentrations in Fujian Province, China. Atmosphere 2023, 14, 787. [Google Scholar] [CrossRef]
  40. Chen, X.; Wei, F. Reducing PM2.5 and O3 through optimizing urban ecological land form based on its size thresholds. Atmos. Pollut. Res. 2025, 16, 102466. [Google Scholar] [CrossRef]
  41. Basu, N.B.; Van Meter, K.J.; Byrnes, D.K.; Van Cappellen, P.; Brouwer, R.; Jacobsen, B.H.; Jarsjö, J.; Rudolph, D.L.; Cunha, M.C.; Nelson, N.; et al. Managing nitrogen legacies to accelerate water quality improvement. Nat. Geosci. 2022, 15, 97–105. [Google Scholar] [CrossRef]
  42. Qu, K.; Yan, Y.; Wang, X.; Jin, X.; Vrekoussis, M.; Kanakidou, M.; Brasseur, G.P.; Lin, T.; Xiao, T.; Cai, X.; et al. The effect of cross-regional transport on ozone and particulate matter pollution in China: A review of methodology and current knowledge. Sci. Total Environ. 2024, 947, 174196. [Google Scholar] [CrossRef]
Water 18 01212 g001
Figure 1. Study area of the Liaohe River Basin, Liaoning section, China.
Figure 1. Study area of the Liaohe River Basin, Liaoning section, China.
Water 18 01212 g001
Water 18 01212 g002
Figure 2. Analytical framework linking landscape metrics, pollution indicators, and GeoSHAP interpretation. CCD, coupling coordination degree; GeoXGBoost, geographically informed extreme gradient boosting; GeoSHAP, geospatial shapley additive explanations.
Figure 2. Analytical framework linking landscape metrics, pollution indicators, and GeoSHAP interpretation. CCD, coupling coordination degree; GeoXGBoost, geographically informed extreme gradient boosting; GeoSHAP, geospatial shapley additive explanations.
Water 18 01212 g002
Water 18 01212 g003
Figure 3. Land-use composition and landscape configuration changes in the Liaohe River Basin from 2000 to 2020.
Figure 3. Land-use composition and landscape configuration changes in the Liaohe River Basin from 2000 to 2020.
Water 18 01212 g003
Water 18 01212 g004
Figure 4. Spatial distributions of TN load, TP load, and aquatic coupled pollution index (CCDNPS) at the sub-basin scale. TN, total nitrogen; TP, total phosphorus; CCDNPS, coupling coordination degree for aquatic non-point source pollution.
Figure 4. Spatial distributions of TN load, TP load, and aquatic coupled pollution index (CCDNPS) at the sub-basin scale. TN, total nitrogen; TP, total phosphorus; CCDNPS, coupling coordination degree for aquatic non-point source pollution.
Water 18 01212 g004
Water 18 01212 g005
Figure 5. Spatial distribution of PM2.5, O3, and atmospheric coupled pollution index (CCDAir) at the sub-basin scale. PM2.5, fine particulate matter with an aerodynamic diameter ≤ 2.5 μm; O3, ozone; CCDAir, coupling coordination degree for atmospheric pollution.
Figure 5. Spatial distribution of PM2.5, O3, and atmospheric coupled pollution index (CCDAir) at the sub-basin scale. PM2.5, fine particulate matter with an aerodynamic diameter ≤ 2.5 μm; O3, ozone; CCDAir, coupling coordination degree for atmospheric pollution.
Water 18 01212 g005
Water 18 01212 g006
Figure 6. Relative importance of landscape predictors based on mean absolute GeoSHAP values. GeoSHAP values represent predictor contributions to model predictions, not causal effects.
Figure 6. Relative importance of landscape predictors based on mean absolute GeoSHAP values. GeoSHAP values represent predictor contributions to model predictions, not causal effects.
Water 18 01212 g006
Water 18 01212 g007
Figure 7. Nonlinear response relationships between landscape predictors and aquatic coupled pollution (GeoSHAP dependence plots). Fitted curves indicate model-based associations.
Figure 7. Nonlinear response relationships between landscape predictors and aquatic coupled pollution (GeoSHAP dependence plots). Fitted curves indicate model-based associations.
Water 18 01212 g007
Water 18 01212 g008
Figure 8. Nonlinear response relationships between landscape predictors and atmospheric coupled pollution (GeoSHAP dependence plots). Fitted curves indicate model-based associations.
Figure 8. Nonlinear response relationships between landscape predictors and atmospheric coupled pollution (GeoSHAP dependence plots). Fitted curves indicate model-based associations.
Water 18 01212 g008
Water 18 01212 g009
Figure 9. Spatial gradients of dominant landscape predictors associated with aquatic and atmospheric coupled pollution.
Figure 9. Spatial gradients of dominant landscape predictors associated with aquatic and atmospheric coupled pollution.
Water 18 01212 g009
Water 18 01212 g010
Figure 10. Planning zones derived from model-based transition intervals and dominant landscape predictors. Purple indicates the water management zone, red indicates the air management zone, and green indicates the ecological protection zone; gray areas in the inset maps indicate non-target areas.
Figure 10. Planning zones derived from model-based transition intervals and dominant landscape predictors. Purple indicates the water management zone, red indicates the air management zone, and green indicates the ecological protection zone; gray areas in the inset maps indicate non-target areas.
Water 18 01212 g010
Table 1. Data sources, temporal coverage, and spatial resolution.
Table 1. Data sources, temporal coverage, and spatial resolution.
DataTimeData FormatData Source
Digital elevation model (DEM)-Grid (cell size, 10 × 10 m)http://www.gscloud.cn (accessed on 20 April 2025)
Soil map and properties-Grid (cell size, 1 × 1 km)Harmonized World Soil Database (https://iiasa.ac.at/models-tools-data/hwsd, accessed on 10 April 2025)
Land use map2000, 2010, 2020Grid (cell size, 30 × 30 m)Landsat 5 Thematic Mapper and Landsat 8 Operational Land Imager (https://earthexplorer.usgs.gov/, accessed on 16 May 2025)
Weather2000–2020Database file (DBF)China Meteorological Administration (https://data.cma.cn/, accessed on 21 May 2024)
Hydrology and water quality data2000–2020Database file (DBF)Local hydrographical station and environmental monitoring station
PM2.5 data2000, 2010, 2020Grid (cell size, 1 × 1 km)ChinaHigh PM2.5 data set (https://zenodo.org/record/6398971, accessed on 15 April 2025)
O3 data2000, 2010, 2020Grid (cell size, 1 × 1 km)ChinaHigh O3 data set (https://zenodo.org/records/13342827, accessed on 15 April 2025)
Table 2. Landscape metrics and their ecological interpretations.
Table 2. Landscape metrics and their ecological interpretations.
MetricLevelMain MeaningEcological Interpretation in This Study
PLANDClassRelative area of a given land typeRepresents the dominance of cropland, forest, or built-up land within a sub-basin
LPILandscapeRelative area of the largest patchReflects landscape dominance and the presence of a controlling spatial matrix
PDLandscape/classNumber of unit areaIndicates fragmentation intensity
EDLandscape/classTotal edge length per unit areaRepresents interface density and edge-related ecological exchange
AILandscape/classDegree of aggregation among similar patchesRepresents spatial cohesion and continuity of a land type
LSILandscape/classShape complexity of the landscape or classCaptures geometric irregularity and spatial complexity
SHDILandscapeLandscape diversityReflects overall landscape heterogeneity and compositional complexity
Notes: PLAND, percentage of landscape; LPI, largest patch index; ED, edge density; PD, patch density; AI, aggregation index; LSI, landscape shape index; SHDI, Shannon’s diversity index.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Shi, S.; Zhang, T.; Liu, M. Landscape Controls on Coupled Water–Air Pollution in an Urbanized Watershed: A GeoSHAP Analysis of the Liaohe River Basin, China. Water 2026, 18, 1212. https://doi.org/10.3390/w18101212

AMA Style

Shi S, Zhang T, Liu M. Landscape Controls on Coupled Water–Air Pollution in an Urbanized Watershed: A GeoSHAP Analysis of the Liaohe River Basin, China. Water. 2026; 18(10):1212. https://doi.org/10.3390/w18101212

Chicago/Turabian Style

Shi, Sixue, Tingshuang Zhang, and Miao Liu. 2026. "Landscape Controls on Coupled Water–Air Pollution in an Urbanized Watershed: A GeoSHAP Analysis of the Liaohe River Basin, China" Water 18, no. 10: 1212. https://doi.org/10.3390/w18101212

APA Style

Shi, S., Zhang, T., & Liu, M. (2026). Landscape Controls on Coupled Water–Air Pollution in an Urbanized Watershed: A GeoSHAP Analysis of the Liaohe River Basin, China. Water, 18(10), 1212. https://doi.org/10.3390/w18101212

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop