DEMARCATING FLOOD RISK ZONES BY CONSIDERING GEOMORPHIC CLASSIFIERS THROUGH AN ADVANCED MACHINE LEARNING-CUMSHAP FRAMEWORK

Abstract

Rising global occurrences of catastrophic flooding demand robust strategies for community protection and resilience building. This research endeavours to pioneer a comprehensive flood susceptibility mapping (FSM) framework tailored for the Kosi River Basin (KRB), often known as "The Sorrow of Bihar." Employing an integrated approach blending geomorphic classifiers with cutting-edge machine learning (ML) techniques, this study develops a stacking ensemble model from four leading ML algorithms: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Categorical Boosting (CatBoost), and Long Short-Term Memory (LSTM). This study examines a wide range of flood conditioning factors (FCFs) spanning from the topographic, hydrologic, morphologic, and stream-related domains, rigorous feature selection processes involving multicollinearity analysis and information gain ratio (IGR) lead to the identification of fifteen essential FCFs for modeling. Subsequently, six performance metrics, namely, accuracy, precision, recall, F1 score, Cohen’s kappa coefficient, and area under the receiver operating characteristic (ROC) curve (AUC) are utilized to evaluate the performance of the ML models. The ensemble model demonstrated its superiority by achieving the highest AUC value of 0.981, outperforming individual models such as XGBoost (0.977), CatBoost (0.973), RF (0.970), and LSTM (0.900). Moreover, it excelled in differentiating between "flooded" and "non-flooded" areas with a Cohen’s kappa coefficient of 0.868, reflecting an exceptional accuracy rate of 93.4%. In comparison, XGBoost (92.3%), CatBoost (91.6%), RF (91.1%), and LSTM (81.2%) followed closely. Additionally, the ensemble model achieved the highest F1 score of 92.4%, surpassing XGBoost (92.3%), CatBoost (91.6%), RF (91%), and LSTM (81.1%). These findings underscore the stacking ensemble model's effectiveness in accurately predicting flood susceptibility within the KRB. This study also incorporates SHapley Additive exPlanations (SHAP), an explainable artificial intelligence (XAI) technique, to improve the interpretability of the ML models. The analysis reveals that elevation, rainfall, CN, and DD are consistently identified as the most crucial features across various ML classifiers. This insight provides relevant authorities, including regional river basin organizations and disaster management divisions, with an opportunity to devise and implement more effective flood control measures based on a better understanding of the influential factors.