Application of deep learning for high-resolution flood mapping in urban watersheds

• Autor(a) principal: Lago, Cesar Ambrogi Ferreira do
• Data de Publicação: 2023
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/18/18138/tde-18042024-113247/
• topic: aprendizado profundo; artificial neural networks; chuva no grid; deep learning; high-resolution flood mapping; mapeamento de inundações de alta resolução; modelos de inundações rápidos; rain-on-grid; rapid flood models; redes neurais artificiais

Flood events significantly threaten urban environments, causing substantial economic damage and loss of life. Accurate prediction and mapping of these events are crucial for effective mitigation strategies. However, current hydrodynamic models used for flood prediction are expensive to build and often impractical for real-time applications or simulations on large domains due to long computational times. This dissertation explores the utility of Deep Learning (DL) models as a viable alternative for flood prediction and floodplain mapping, addressing the evident gap in current flood modeling practices. The research implements a three-fold methodology across three chapters, focusing on developing and applying ANNs for flood prediction. Chapters 1 and 2 use a conditional generative adversarial network developed for rapid pluvial flood predictions (cGAN-Flood). Chapter 1 demonstrates a novel DL application – improving flood mapping resolution from existing coarse hydrodynamic models using cGAN-Flood. Chapter 2 assesses the performance of cGAN-Flood, in distinct topological settings, specifically catchments in Sao Paulo, compared to its original training in San Antonio, Texas. Lastly, Chapter 3 outlines the creation of a novel model that predicts pluvial flood maps using ANN, requiring only Digital Elevation Models (DEM) and inflow inputs. General results across the chapters show the promising efficacy of ANNs and DL models in flood prediction and floodplain mapping. ANNs demonstrated the ability to emulate hydrodynamic models with high precision, while cGAN-Flood\'s application showed satisfactory predictive capabilities even in geographically distinct and topologically different regions. The newly proposed model in Chapter 3 compared favorably against FEMA floodplain maps, despite the simplicity of its training data. In conclusion, the research demonstrates that DL models, with further enhancements and training, can transform floodplain mapping and prediction, supporting faster simulations and extending applicability to different locations without retraining. This research underscores the potential of these models in bridging the gaps in current flood modeling practices, which is particularly significant for real-time flood prediction and the development of mitigation strategies, especially in developing regions where resources may be scarce or in larger domains.