Flood susceptibility mapping using explainable machine learning models

Abstract

Flooding is one of the most frequently encountered natural disasters globally. Frequent severe flood occurrences in Rathnapura city, Sri Lanka caused damages to both human lives and infrastructures. Data-driven models have been showing their ability of flood susceptibility mapping (FSM) in data-scare regions as an alternative to traditional hydrological models, but they are not widely used by stakeholders due to their black-box nature. This research suggests utilising the shapley additive explanation (SHAP) method to interpret the results generated by the CatBoost machine learning model and to assess the influence of different variables on flood susceptibility mapping. A flood inventory (445 flooded locations) and thirteen flood conditioning factors were used to implement the model and results were validated using the area under curve (AUC) method, which showed a success rate and prediction rate of 93.1% and 92.5%, respectively. SHAP plots indicated that the regions with lower elevations and topographic roughness values, gentler slopes, closer proximity to rivers, and moderate rainfall are more susceptible to flooding. According to the results obtained, we suggest incorporating SHAP-based datadriven models in forthcoming studies on FSM to enhance the interpretations of model outcomes.