Integrating Classification and Regression for Rainfall Forecasting: A DeepLearning and Ensemble-Based Approach

Introduction: Accurate rainfall forecasting is found to be very important for agricultural planning, disaster mitigation, and achieving climate resilience. This research paper proposes a hybrid architecture combining Random Forest (RF) and deep learning-based Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) to enhance rainfall classification and regression within the weatherAUS dataset Methods: The study addresses class imbalance using SMOTE and optimizes performance through systematic hyperparameter tuning. The RF model is applied for both classification and regression tasks, while the CNN-LSTM model with transformer blocks captures temporal dependencies. SHAP analysis is integrated for interpretability. Results: The RF Regressor demonstrated superior performance with low Mean Squared Error (MSE = 0.8185), Root Mean Square Error (RMSE = 0.9047), and a high R-squared of 98.99%. The CNN-LSTM model achieved 90% classification accuracy after hyperparameter tuning and SMOTE application. SHAP analysis identified features 3, 5, and 11 as most critical for rainfall prediction. Discussion: The integration of ensemble learning, temporal modelling, and explainability tools enhances rainfall prediction systems. Hyperparameter tuning reduced MAE by 3.4% and MSE by 4.2%, while SMOTE improved minority class recall from 51% to 91% for RF. Conclusion: The hybrid RF-CNN-LSTM-Transformer architecture effectively captures complex spatiotemporal weather patterns, providing a transparent, high-precision approach for predicting both the occurrence and intensity of rainfall.