Improving pluvial flood mapping resolution of large coarse models with deep learning

Abstract

Deep Learning (DL) models are a promising complement to hydrodynamic models. However, the application of DL for detailed predictions on large domains has not yet been tested. We aim to narrow address this gap by improving flood mapping resolution derived from large coarse flood models. We have used cGAN-Flood, a conditional generative adversarial-based model (cGAN), that showed satisfactory generalization. We demonstrate the applicability of cGAN-Flood by coupling it with mesh and raster-based coarse models. A HEC-RAS model (cell size of 32m), which is mesh-based, was created for a 350 km2 watershed. In contrast, a HydroPol2D, raster-based, was created for a 150 km2 with a 15m pixel size. We evaluated our method’s performance against a 3m resolution HEC-RAS model in seven different catchments across San Antonio and Sao Paulo cities. Results indicate that the DL model considerably improved flood map accuracy, illustrating how DL can enhance flood mapping resolution.