Unleashing the power of Geomorphic Descriptors and machine learning to assess flood hazards over large watersheds

Código

I-EBHE0087

Autores

Vaibhav Tripathi, Mohit Prakash Mohanty

Tema

WG 1.05: Comparative understanding of runoff generation processes

Resumo

Quantifying flood hazards using hydraulic hydrodynamic modeling (HHM) requires huge computational efforts and costs, therefore a huge constraint for low and middle-income nations. To overcome this, a blend of Geomorphic Flood Descriptors (GFDs) and Machine Learning (ML) are integrated together to identify the flood hotspots in the severely flood-prone Ganga River Basin (GRB). GFDs have been widely recognized by the scientific community for assessing and mapping flood hazards across large geographical areas, being computationally penurious, and ease of availability makes them ideal for large watersheds. An assortment of single and composite indices (GFDs) generated from a high-resolution CartoDEM (resolution~30m) was forced as input to ML algorithms. A widely used wrapper method for feature selection, namely, Boruta and Recursive Feature Elimination (RFE) methods were employed to get the most influential feature. The current study encompasses two important works: one to delineate flood extent, and second to predict the flood water depth in that extent using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models. An Artificial Bee Colony (ABC) algorithm is incorporated to optimize the ML models parameters. A suite of six performance metrics, namely precision, recall, Cohen?s kappa coefficient, accuracy, F1 score, and ROC curve for classification, and Mean Absolute Relative Error (MARE), Nash Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), Percentage bias (PBIAS), Mean Absolute Error (MAE), and Kling Gupta Efficiency (KGE) for regression were adopted to enumerate the performance of the ML models. An explainable artificial intelligence (XAI) approach, namely SHapley Additive exPlanations (SHAP) is also incorporated to understand the complex nature of these algorithms. The results obtained may be implemented for proactive flood risk management and mitigation strategies in data-scarce and large basins, where HHM is still challenging.