Deep learning rapid flood risk predictions for climate resilience planning

Abstract

Floods have been causing the world’s costliest weather-related catastrophes and their magnitude and frequency are projected to increase even further due to climate change. Current flood risk quantification procedures include the use of complex and highly uncertain hydrologic-hydraulic models for hazard mapping and computationally-tedious manipulations for vulnerability evaluation—hindering urban centers climate resilience planning. Adopting a novel approach that bypasses such time-consuming procedures, this study presents a deep learning-based rapid and accurate flood risk prediction tool, RAPFLO, to directly relate flood risk characteristics (level, extent, and likelihood) to their main drivers (e.g., climate, topography, and land cover). The approach employed to develop RAPFLO is generic in nature and the associated methodology is not site-dependent. To demonstrate its utility, RAPFLO is deployed on the City of Calgary, Canada, and is used to reproduce the fluvial flood risk across the city between the years 2010 and 2020. RAPFLO efficiently replicated the risk level with an overall accuracy of 80 % and the risk likelihood with a coefficient of determination of 0.96. Subsequently, RAPFLO was employed for predicting future fluvial flood risk from the year 2025 to 2100 under the RCP 8.5 climate scenario. RAPFLO presents a valuable computationally efficient, accurate, and rapid decision support system that empowers city managers and infrastructure operators to devise effective climate resilience strategies considering different climate projections and future what-if scenarios.