Nationwide Modeling of River Ammonia-Nitrogen Concentrations in Taiwan Using Machine Learning

Decades of river water quality monitoring in Taiwan have revealed a clear trend of deterioration, with ammonia-nitrogen (NH₃–N) concentrations at several monitoring stations frequently exceeding the regulatory thresholds established by the Ministry of Environment. This degradation arises from the combined influence of natural biogeochemical processes and diverse anthropogenic pressures. Accurately modeling the spatial variability of river water quality is therefore both challenging and essential for protecting riverine ecosystems and public health. In recent years, data-driven machine learning (ML) approaches have demonstrated strong capability in capturing complex, nonlinear relationships in both surface and subsurface water systems. In this study, we develop a predictive model for riverine NH₃–N concentrations using an artificial neural network (ANN) trained on an extensive suite of multivariate datasets compiled across multiple government ministries. Model performance is rigorously evaluated through three-fold cross-validation, confirming that the ANN effectively captures the primary spatiotemporal variability of NH₃–N and provides reliable predictive accuracy. To further interpret the model, SHAP analysis is conducted to identify key predictors. The results show that average precipitation in November, the extent of land undergoing human modification, the density of food-product and animal-feed manufacturing activities, and the forest land-use group are among the most influential drivers of NH₃–N concentrations. Identifying such dominant variables is crucial for guiding evidence-based river water quality management and for formulating targeted pollution mitigation strategies.