Abstract
Flooding is a major natural hazard worldwide, causing catastrophic damage to communities and infrastructure. Due to climate change exacerbating extreme weather events robust flood hazard modeling is crucial to support disaster resilience and adaptation. This study uses multi-sourced geospatial datasets to develop an advanced machine learning framework for flood hazard assessment in the Arambag region of West Bengal, India. The flood inventory was constructed through Sentinel-1 SAR analysis and global flood databases. Fifteen flood conditioning factors related to topography, land cover, soil, rainfall, proximity, and demographics were incorporated. Rigorous training and testing of diverse machine learning models, including RF, AdaBoost, rFerns, XGB, DeepBoost, GBM, SDA, BAM, monmlp, and MARS algorithms, were undertaken for categorical flood hazard mapping. Model optimization was achieved through statistical feature selection techniques. Accuracy metrics and advanced model interpretability methods like SHAP and Boruta were implemented to evaluate predictive performance. According to the area under the receiver operating characteristic curve (AUC), the prediction accuracy of the models performed was around > 80%. RF achieves an AUC of 0.847 at resampling factor 5, indicating strong discriminative performance. AdaBoost also consistently exhibits good discriminative ability, with AUC values of 0.839 at resampling factor 10. Boruta and SHAP analysis indicated precipitation and elevation as factors most significantly contributing to flood hazard assessment in the study area. Most of the machine learning models pointed out southern portions of the study area as highly susceptible areas. On average, from 17.2 to 18.6% of the study area is highly susceptible to flood hazards. In the feature selection analysis, various nature-inspired algorithms identified the selected input parameters for flood hazard assessment, i.e., elevation, precipitation, distance to rivers, TWI, geomorphology, lithology, TRI, slope, soil type, curvature, NDVI, distance to roads, and gMIS. As per the Boruta and SHAP analyses, it was found that elevation, precipitation, and distance to rivers play the most crucial roles in the decision-making process for flood hazard assessment. The results indicated that the majority of the building footprints (15.27%) are at high and very high risk, followed by those at very low risk (43.80%), low risk (24.30%), and moderate risk (16.63%). Similarly, the cropland area affected by flooding in this region is categorized into five risk classes: very high (16.85%), high (17.28%), moderate (16.07%), low (16.51%), and very low (33.29%). However, this interdisciplinary study contributes significantly towards hydraulic and hydrological modeling for flood hazard management.