Explainable dual-model framework for rainfall anomaly detection using SHAP-enhanced machine learning

Extremes in rainfall caused by climate present increasingly serious threats to agricultural economies and water resource systems, but traditional techniques of identifying anomalies remain black and opaque and do not offer an interpretative basis on which policy changes can be made in response to climate. The study introduces a novel two-model design that combines Isolation Forest and Autoencoder models, utilising Shapley Additive exPlanations (SHAP) to provide transparency at the feature level and identify hydro-climatic anomalies in 40 years of rainfall data at Rajshahi. In comparison to current methods that focus on predictive accuracy without regard to interpretability, our hybrid consensus model achieves 92% precision, 95% recall, and 94% F1-score, while also demonstrating the effect of a seasonal relationship and time lag, which leads to the development of outliers. The interpretability perspective by SHAP indicates that 78 per cent of the identified anomalies occur during monsoon months, which has a direct correlation with documented floods and sensitive agricultural seasons. This framework addresses the black-box limitations of deep learning climatic models by integrating the strengths of unsupervised learning with clear mechanistic insights, offering actionable intelligence on early warning systems and adaptive hydrological planning in monsoonal basins with limited data.