Advancing tropical flood susceptibility mapping with multi-head attention mechanism and deep learning models

Deep learning (DL) models have demonstrated a significant potential in predicting complex phenomena such as flood susceptibility mapping, particularly with large datasets. However, overfitting remains a challenge, limiting DL models generalizability. This study aims to enhance flood susceptibility mapping by integrating the multi-head attention (MHA) mechanism into DL models, focusing on the Muda River Basin. Six machine learning (ML) models were evaluated: Support Vector Machine (SVM), Bidirectional Gated Recurrent Unit (BGRU), K-Nearest Neighbors (KNN), Bidirectional Long Short-Term Memory (BLSTM), and two advanced models, MHA-BGRU and MHA-BLSTM. Fifteen flood influencing factors were selected, including plan curvature, elevation, stream power index (SPI), slope, topographic position index (TPI), profile curvature, topographic wetness index (TWI), rainfall, convergence index (CI), land use/land cover (LULC), aspect, normalized difference vegetation index (NDVI), topographic ruggedness index (TRI), distance from rivers, and soil type. Pearson's Correlation Coefficient (PCC) and Information Gain Ratio (IGR) methods were applied for optimal feature selection. Results revealed that MHA-based models outperformed the others, with improvements exceeding 11 % in the area under the curve (AUC), 12 % in accuracy, 16.67 % in F1-score, and 33 % in hit rate. Among these, the MHA-BLSTM model achieved superior performance, confirmed by the Wilcoxon signed-rank test and bootstrap uncertainty analysis. Model performance was further assessed using mean annual and total maximum rainfall during flash floods. The MHA-BLSTM model produced more balanced predictions across susceptibility categories. These findings highlight the effectiveness of the MHA mechanism in enhancing flood susceptibility mapping in tropical regions.