A Stacking Ensemble Method for Comprehensive Flood Susceptibility Mapping in Yemen

The necessity of flood risk mapping is critical for effective planning and disaster response, particularly in flood-prone regions like the Qaa'Jahran watersheds in Dhamar, Yemen. This research implements various machine learning methods, including Support Vector Machines (SVM), K-Nearest Neighbors (kNN), Random Forest (RF), Artificial Neural Networks (ANN), and Logistic Regression (LR), with the latter also functioning as the meta-model in our stacking ensemble approach for mapping flood susceptibility. The process began with creating a flood inventory map using SAR images and historical flood records. Our model integrates the individual strengths of each technique and employs a meta-model to synthesize their forecasts. This stacked ensemble approach demonstrated superior performance over each model alone, achieving a remarkable AUC score of 0.9848 compared to the individual scores of SVM, LR, kNN, ANN, and RF. It also surpassed two innovative models, ABRBF and TPOT, in accurately pinpointing all high-risk zones identified in historical flood data. This advancement in flood risk mapping for the Qaa'Jahran watersheds exemplifies the potential of our model in enhancing disaster management and prevention efforts. It offers a significant tool for identifying at-risk areas and guiding mitigation strategies to safeguard communities in Dhamar, Yemen, against the catastrophic impacts of flooding.