Can geomorphic flood descriptors coupled with machine learning models enhance in quantifying flood risks over data-scarce catchments? Development of a hybrid framework for Ganga basin (India).

Abstract

Quantifying flood risks through a cascade of hydraulic-cum-hydrodynamic modelling is data-intensive and computationally demanding- a major constraint for economically struggling and data-scarce low and middle-income nations. Under such circumstances, geomorphic flood descriptors (GFDs), that encompass the hidden characteristics of flood propensity may assist in developing a nuanced understanding of flood risk management. In line with this, the present study proposes a novel framework for estimating flood hazard and population exposure by leveraging GFDs and Machine Learning (ML) models over severely flood-prone Ganga basin. The study incorporates SHapley Additive exPlanations (SHAP) values in flood hazard modeling to justify the degree of influence of each GFD on the simulated floodplain maps. A set of 15 relevant GFDs derived from high-resolution CartoDEM are forced to five state-of-the-art ML models; AdaBoost, Random Forest, GBDT, XGBoost, and CatBoost, for predicting flood extents and depths. To enumerate the performance of ML models, a set of twelve statistical metrics are considered. Our result indicates a superior performance of XGBoost (κ = 0.72 and KGE = 82%) over other ML models in flood extent and flood depth prediction, resulting in about 47% of the population exposure to high-flood risks. The SHAP summary plots reveal a pre-dominance of Height Above Nearest Drainage during flood depth prediction. The study contributes significantly in comprehending our understanding of catchment characteristics and its influence in the process of sustainable disaster risk reduction. The results obtained from the study provide valuable recommendations for efficient flood management and mitigation strategies, especially over global data-scarce flood-prone basins.