Feature-Enhanced Rainfall Prediction in Bangladesh Using Statistical and Machine Learning Models Integrated with Explainable AI

Rainfall prediction is significant in the fields of agriculture, water resource management, and even the prediction of natural disasters. However, due to the intricacies associated with predicting rainfall, traditional forecasting models may face some challenges. The current research seeks to enhance rainfall predictions through the application of Machine Learning (ML) and Explainable Artificial Intelligence (XAI) by performing feature engineering through the addition of new features such as Rainfall_Lag1, Rainfall_Lag2, Rainfall_MA3, Month_Sin, and Month_Cos to the existing dataset. These elements aim to reflect historical trends and seasonal variations. For evaluating performance, this study includes a statistical model, Seasonal Autoregressive Integrated Moving Average (SARIMA), and ML models such as decision tree regression (DTR), random forest regression (RFR), extreme gradient boosting regression (XG-BOOSTR) and support vector regression (SVR). After applying 10-fold cross-validation, SVR emerges as the best performing model by achieving a Mean Squared Error (MSE) of 518.3105, a Mean Absolute Error (MAE) of 3.0030, a Root Mean Squared Error (RMSE) of 20.6821, an R-squared (R2) of 0.9920 and a Kling-Gupta efficiency(KGE) of 0.9770. The study outcomes are evaluated using Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) in AI, reflecting the need for interpreting those features. The results show that a combination of feature engineering, ML models, and XAI provides a robust model and improves the transparency of rainfall forecasting, which ultimately will contribute to real-world applications.