Ensemble, Kernel-based, and Deep Learning approaches for flood susceptibility mapping: A case study at Lake-Watersheds

Lake Tana and its surrounding regions experience frequent flooding, necessitating improved susceptibility mapping to mitigate risks and enhance resilience. This study applies data-driven machine learning techniques to assess flood susceptibility utilizing data sets commonly used in large-scale river basin studies. A comprehensive flood inventory of approximately 2,080 flooded locations was compiled alongside 14 predictive variables. The predictive features include elevation, slope, distance to Lake, maximum precipitation and topographic wetness index. The models tested include Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Convolutional Neural Network (CNN). Model performance was validated using the Kappa score and the area under the receiver operating characteristic curve (AUC). Results indicate that all models perform exceptionally well, with a minimum AUC of 0.87 for the testing dataset. RF consistently outperformed other models, achieving an AUC of 0.96 for the flood inventory and predictor variables. Elevation and distance to Lake Tana emerged as the most critical influencing factors. This study underscores the effectiveness of machine learning-based flood susceptibility mapping for Lake Tana and its surrounding watersheds, provided that a reliable flood inventory is available. The findings support data-driven approaches in flood risk assessment and offer valuable insights for disaster preparedness. Keywords: Flood susceptibility; Lake Tana Basin; Machine Learning