Karapetyan Areg ; Chow Aaron Chung Hin ; Madanat Samer

Abstract

In light of growing threats posed by climate change in general and sea levelrise (SLR) in particular, the necessity for computationally efficient means toestimate and analyze potential coastal flood hazards has become increasinglypressing. Data-driven supervised learning methods serve as promising candidatesthat can dramatically expedite the process, thereby eliminating thecomputational bottleneck associated with traditional physics-based hydrodynamicsimulators. Yet, the development of accurate and reliable coastal floodprediction models, especially those based on Deep Learning (DL) techniques, hasbeen plagued with two major issues: (1) the scarcity of training data and (2)the high-dimensional output required for detailed inundation mapping. To removethis barrier, we present a systematic framework for training high-fidelity DeepVision-based coastal flood prediction models in low-data settings. We test theproposed workflow on different existing vision models, including a fullytransformer-based architecture and a Convolutional Neural Network (CNN) withadditive attention gates. Additionally, we introduce a deep CNN architecturetailored specifically to the coastal flood prediction problem at hand. Themodel was designed with a particular focus on its compactness so as to cater toresource-constrained scenarios and accessibility aspects. The performance ofthe developed DL models is validated against commonly adopted geostatisticalregression methods and traditional Machine Learning (ML) approaches,demonstrating substantial improvement in prediction quality. Lastly, we roundup the contributions by providing a meticulously curated dataset of syntheticflood inundation maps of Abu Dhabi's coast produced with a physics-basedhydrodynamic simulator, which can serve as a benchmark for evaluating futurecoastal flood prediction models.