Interpretable machine learning for daily precipitation forecasting: spatiotemporal dynamics and extreme event prediction in New Jersey

Accurate daily precipitation forecasting is essential for operational hydrology. This study evaluates six machine learning (ML) models for 1-day-ahead precipitation forecasting across 31 diverse stations in New Jersey (NJ), USA, using an integrated meteorological, remote sensing, and topographic dataset. Model performance was assessed for both general and 95th percentile extreme events. Ensemble models (random forest and XGBoost) consistently outperform other methods, particularly for high-impact extremes. Explainable AI (XAI) techniques, specifically SHAP, were used to interpret the models, revealing antecedent moisture conditions as dominant predictors. Crucially, the analysis demonstrates predictor importance is significantly modulated by land use/land cover (LULC), with urban areas showing heightened sensitivity to recent rainfall, reflecting the altered hydrological response of impervious catchments. Spatiotemporal analyses confirm predictive drivers are geographically structured. This research validates interpretable ML as a tool to improve forecast skill and enhance scientific understanding of heterogeneous hydrological processes.