Advanced Flood Risk Assessment Using a Novel Hybrid Transformer-SE-ANN Framework

Abstract Flood prediction is a vital yet complex task in disaster management, necessitating precise modeling of multifaceted factors such as weather patterns, topography, and human influences. In this study, we introduce a novel hybrid approach, the Hybrid Transformer SE-ANN, designed to improve flood probability forecasting by combining the Transformer architecture, Squeeze-and-Excitation (SE) attention mechanisms, and Artificial Neural Networks (ANN). The Transformer captures long range dependencies in the data, the SE module emphasizes critical features, and the ANN excels at discerning intricate patterns. Our model demonstrated exceptional performance, achieving a Mean Absolute Error (MAE) of 0.00234, Root Mean Squared Error (RMSE) of 0.00297, Mean Absolute Percentage Error (MAPE) of 0.50%, and an R2 score of 0.99599, reflecting its remarkable accuracy and precision. To ensure these results are robust and free from bias, overfitting, or data leakage, we applied rigorous validation techniques, including SHAP and LIME for interpretability, alongside actual vs. predicted plots and Kernel Density Estimation (KDE) curves to confirm consistency and generalizability. The Hybrid Transformer-SE-ANN offers a reliable and highly effective tool for flood prediction, with promising implications for real-world deployment.