Assessing the impact of land use and land cover on predicting in-stream total phosphorus using GIS and machine learning models

This study presents a spatially explicit framework for predicting in-stream total phosphorus (TP) concentrations by combining a convolutional neural network (CNN) with explainable AI techniques. A CNN–dense neural network (CNN–DNN) model was trained on five-year average in-stream TP data from 608 monitoring stations across the four major river basins of the Republic of Korea, using 22 detailed land use and land cover (LULC) classes, a digital elevation model (DEM), and slope. The model outperformed proportion-only regression, highlighting the added value of spatial LULC representation. Unlike previous TP prediction studies that relied primarily on proportional LULC metrics, our approach preserves the spatial arrangement of LULC patches through CNN-based feature extraction and integrates SHapley Additive exPlanations (SHAP) and Integrated Gradients (IG) to provide spatially explicit attribution of patch-level influences. This enables detection of near-stream clustering effects and landscape configurations that cannot be captured by proportion-only models. Model interpretation with SHAP revealed that residential zones generally reduced TP, while commercial and public utility areas increased it when clustered near streams. Paddy fields consistently exerted the strongest positive influence in agricultural watersheds, modulated by slope and hydrologic connectivity. Forests largely buffered TP, though mixed stands and adjacent industrial parcels occasionally increased nutrient loads. To enhance robustness, IG analysis with three baselines revealed consistent attribution patterns that reinforced SHAP findings. Our results reveal that spatial watershed structure (LULC, DEM, and slope) alone can reproduce long-term TP patterns, suggesting that watershed structure is a primary control on phosphorus loads and that land-use planning can serve as an effective long-term management tool independent of short-term meteorological variation. Scenario-based simulations further showed that forest-to-residential conversions slightly lowered TP, whereas industrial and agricultural conversions—particularly to paddy fields—substantially increased TP, with central clusters producing the largest impacts. Overall, the CNN–SHAP–IG framework provides local interpretability and domain-wide robustness, representing a methodological advancement over proportion-based studies and offering a practical decision-support tool for watershed management and spatially informed land-use planning to mitigate phosphorus pollution.