Geographically weighted machine learning for predicting the spatial distribution of groundwater nitrate nitrogen (NO3-N) concentration

This study was conducted on Jeju Island, South Korea, where groundwater serves as a critical water resource and is highly susceptible to contamination from both natural and anthropogenic influences. This study aims to predict the spatial distribution of groundwater nitrate nitrogen (NO 3 -N) concentrations using geographically weighted random forest (GWRF). The predictive performance and robustness of the GWRF were compared against five conventional machine learning models (CMLMs). Shapley Additive Explanations (SHAP) analysis was employed to quantify the influence of key input variables, categorized into land surface, geological, and anthropogenic factors, on the model predictions. The result showed the superior performance of the GWRF in predicting the spatial distribution of groundwater NO 3 -N on Jeju Island. Compared to CMLMs, the GWRF provided more accurate and spatially unbiased predictions due to consideration of spatial variability. The SHAP analysis indicated that key factors influencing groundwater NO 3 -N were average elevation, proportion of heavy clay fields and agricultural areas, and urban areas. These results demonstrated the potential of combining geographically weighted structures with machine learning models in groundwater modeling using geospatial data. Consequently, the findings from this study would help to develop targeted management strategies to mitigate groundwater NO 3 -N pollution. • GWRF was applied to predict groundwater NO 3 -N distribution. • Conventional ML models were employed as baseline comparators. • GWRF showed enhanced predictive accuracy. • Especially, GWRF addressed spatial autocorrelation.