Performance analysis of physically-based (HEC-RAS, CADDIES) and AI-based (LSTM) flood models for two case studies

Abstract

Megacities in developing countries are commonly affected by flooding events. The use of flood models can contribute to an evidence-based decision-making process. For a good representation, these models require physical data for catchment parameterization, and observed data for calibration and validation, which is often scarce. In this study, we analysed the performance results of physically-based (HEC-RAS, CADDIES) and AI-based (LSTM) flood models for two case studies: the Narmada basin in India and the Aricanduva catchment in Brazil. The models were evaluated for accuracy, interpretability, running time, and complexity.