Flood Subsidence Susceptibility Mapping using Elastic-net Classifier: New Approach

In light of recent improvements in flood susceptibility mapping using machine learning models, there remains a lack of research focusing on employing ensemble algorithms like Light Gradient Boosting on Elastic-net Predictions (Light GBM) and Elastic-net Classifier (L2/Binomial Deviance) for mapping flood susceptibility in Qaa'Jahran, Yemen. This study aims to bridge this knowledge gap through the development and comparative performance of these models. This approach created the flood inventory map using satellite images and field observations. A geodatabase was used to create flood predictors such as aspect, altitude, distance to rivers, topographic wetness index (TWI), flow accumulation, lithology, distance to road, land use, profile curvature, plan curvature, slope, rainfall, soil type, Topographic Position Index (TPI), and Terrain Ruggedness Index (TRI). The developed models were trained using 80% of the data and evaluated using the remaining 20% to create a flood susceptibility map. The receiver operating characteristic (ROC) curve and area under the curve (AUC) were used to evaluate the map's accuracy. The results of this study indicated that the traditional (Elastic-net Classifier) model possessed high accuracy (AUC = 0.9457, F1 = 0.8916, Sensitivity = 0.9024, and Precision = 0.881) than the ensemble algorithm (Light Gradient Boosting on Elastic-net Predictions) (AUC = 0.9629, F1 = 0.9538, Sensitivity = 0.9688, and Precision = 0.9394). Based on the results of this study it can be concluded that these algorithms has a strong potential to offer a practical and affordable method for geospatial modeling of flood vulnerability. This information can be used to assess the flood emergency, early warning system and provide insights for planning and response purposes.