Machine learning-based rainfall prediction across temporal scales: model benchmarking and explainability analysis

Abstract Accurate rainfall forecasting is essential for effective water resources management, agricultural planning, and disaster risk reduction, particularly in monsoon-influenced regions characterized by high spatiotemporal variability. While machine learning (ML) approaches have shown strong potential for precipitation prediction, most existing studies focus on a single temporal resolution and primarily emphasize performance comparison, offering limited insight into scale-dependent rainfall dynamics and predictor relevance. Addressing this gap, the present study proposes a scale-aware and interpretable ML framework for daily and weekly rainfall forecasting by explicitly treating each temporal resolution as a distinct learning problem. Using high-resolution IMDAA reanalysis data for Bengaluru, India, and neighbouring regions, multiple ML, deep learning (DL), and ensemble models were systematically evaluated for 1-day-ahead and 1-week-ahead rainfall prediction. To enhance interpretability, SHapley Additive Explanations (SHAP) were employed to examine scale-dependent predictor importance. The results show that model performance varies substantially across temporal scales. At the daily scale, the Multilayer Perceptron (MLP) achieved the highest predictive accuracy, while tree-based ensemble models such as Gradient Boosting (GB) and Light Gradient Boosting Machine (LGBM) demonstrated robust and consistent performance. At the weekly scale, gradient-boosting-based models, particularly XGBoost (XGB) and LGBM, outperformed other approaches, achieving coefficients of determination (R²) exceeding 0.99. SHAP analyses reveal that antecedent precipitation is the dominant driver of forecast skill at both temporal scales; however, the relative importance of synoptic-scale atmospheric variables, especially geopotential height, increases in weekly forecasts. These findings demonstrate that daily and weekly rainfall predictions are governed by different atmospheric controls and should not be treated as simple temporal aggregations. The proposed framework provides both methodological and physical insights into scale-dependent rainfall forecasting and highlights the importance of interpretable, scale-aware modelling strategies for reliable precipitation prediction.