Advancing regional water quality modelling: integrating spatial machine learning with large-scale catchment data

Intensified agriculture increases nutrient loads in waterbodies, threatening aquatic ecosystems and human health. Estimating nutrient concentrations is challenging due to the limited spatial and temporal coverage of national monitoring networks. Explainable machine learning can address this by linking nutrient concentrations to upstream catchment characteristics. We trained Random Forest models to predict total nitrogen (TN) and total phosphorus (TP) concentrations in streams at almost 900 in-stream locations using upstream catchment-scale covariates. To characterise each of the upstream catchment areas, we used openly accessible global and local environmental datasets. As a novel approach, we additionally incorporated spatial covariates, including coordinates and buffers, and we tested how models would perform with fewer but more meaningful covariates. To assess covariate importance to the prediction target, we employed the SHapley Additive exPlanations (SHAP) method. TN predictions were accurate, while TP predictions were poor. Models with reduced covariates achieved similar accuracy to baseline models and decreased overfitting. The inclusion of spatial covariates provided only minimal improvement of the prediction accuracy scores themselves; however, they demonstrate potential in capturing spatial structure and supporting regionalisation, and in some cases, they outranked their corresponding full-catchment covariate versions in SHAP covariate importance. In conclusion, utilising catchment characteristics and machine learning can yield a robust regional model for TN, enabling the reliable estimation of TN loads in streams.