A Spatially Robust and Interpretable Machine Learning Framework for Rainfall Erosivity Prediction

Rainfall erosivity (R30) is a critical driver of soil erosion processes, yet its spatial and temporal variability poses significant challenges for reliable prediction. This study presents a spatially robust and interpretable machine learning framework for modelling R30 using ensemble and neural network approaches. A comprehensive modelling pipeline was developed incorporating structured pre-processing, logarithmic target transformation, and a station-wise GroupKFold strategy was explicitly designed to eliminate spatial leakage, a common but often overlooked source of over-optimistic performance in erosivity modelling studies. Hyperparameter optimization was conducted using RandomizedSearchCV to ensure model generalization. Among the evaluated algorithms, the optimized Random Forest model demonstrated superior performance compared to a tuned multilayer perceptron (MLP), achieving a mean spatial cross-validated R2 of approximately 0.70 on the original R30 scale. SHAP (Shapley Additive Explanations) analysis revealed rainfall energy–related indices as dominant predictors, confirming physical consistency between model behaviour and erosivity mechanisms. This study jointly combines spatial GroupKFold validation, SHAP-based physical interpretability, and temporal forecasting assessment for R30 modelling at the station scale Spatial residual diagnostics indicated no systematic geographic bias, supporting model robustness across heterogeneous climatic regimes. Temporal generalization was further assessed using a leave-last-year-per-station validation scheme, yielding an R2 of 0.51 under forecasting conditions. Prediction uncertainty was quantified through tree-level ensemble variability, with 90% prediction intervals achieving approximately 88% empirical coverage. Interval widths increased for extreme erosivity events, reflecting heightened stochastic variability. The proposed framework demonstrates that ensemble machine learning methods can provide accurate, interpretable, and uncertainty-aware prediction of rainfall erosivity while maintaining strict spatial validation standards. This is one of the first studies to jointly integrate spatially explicit cross-validation, SHAP-based physical interpretability, and probabilistic uncertainty estimation for station-scale rainfall erosivity modelling