Explainable Artificial Intelligence and Machine Learning for Flood Susceptibility Modelling in a Tropical River Basin.

Abstract

The Manimala River Basin (Kerala, India) experienced extensive inundation during the 2018 Kerala flood, underscoring the need for robust, interpretable, and spatially explicit susceptibility assessments to support risk-informed planning. This study develops a flood susceptibility model for the basin using a Random Forest (RF) classifier and explains its behaviour using an explainable AI framework. A set of hydro-morphometric and land-surface conditioning factors was compiled, including rainfall, vertical distance to channel network (VCDN), slope angle, soil texture, distance to streams, land use/land cover (LULC), topographic wetness index (TWI), and terrain curvature metrics (upslope and downslope curvature). The RF model was trained for binary classification of flooded and non-flooded locations, and predictive skill was evaluated using both discrimination and classification metrics. The model achieved strong performance, with an area under the receiver operating characteristic curve (AUROC) of 0.90, overall accuracy of 0.82, sensitivity of 0.83, specificity of 0.81, and an F1-score of 0.83, indicating reliable detection of flood-prone locations while maintaining balanced error rates. The susceptibility map was reclassified into three levels to facilitate interpretation and application. The areal distribution shows that 12.62% of the basin falls within the high-susceptibility class, 15.43% within the moderate class, and 71.95% within the low class, providing a basin-scale overview of priority zones for mitigation and preparedness. Model interpretability was addressed using SHapley Additive exPlanations (SHAP). The SHAP summary and mean absolute contribution rankings indicate that rainfall and VCDN exert the strongest influence on RF outputs, followed by slope angle and soil texture, whereas streams, LULC, TWI, and curvature variables contribute comparatively less. These results emphasize the dominant role of hydro-climatic forcing and drainage-related controls, modulated by terrain and substrate characteristics, in shaping flood susceptibility within the Manimala basin. Overall, the proposed RF–SHAP workflow delivers a high-performing and transparent susceptibility product that can support targeted management actions and communication of drivers underlying predicted flood-prone areas.