Physics-informed Machine Learning for Compound Flood Mapping

Compound flooding (CF), as a result of oceanic, hydrological, meteorological and anthropogenic drivers, is often studied with hydrodynamic models that combine either successive or concurrent processes to simulate inundation dynamics. In recent years, data methods have emerged as effective alternatives for post-flood mapping and supported current efforts of complex physical and dynamical modeling. Yet, those techniques have not been explored for regional compound flood mapping. Here, we propose to develop a Machine Learning (ML) algorithm for generating CF maps at select freshwater-influenced systems along the Gulf coast of United States. In this project we propose to build on the previous works by PIs to integrate multi-source remotely-sensed data, hydrodynamic modeling, deep learning and data fusion techniques for large-scale compound flood mapping. Our main goal in this project is to address NOAA’s operational gaps in model validation, rapid exposure assessment, and reliable flood mapping in data scarce regions. Our research products enables automated re-evaluation of National Water Model (NWM) and its accompanying ocean circulation models’ performance in compound coastal flood inundation mapping, and thus helps recalibration/parametrization of flood forecasting models after major storms. Given much less computational expenses of the proposed hybrid approach, compared with pure physics-based numerical models, it can be used for rapid and reliable post-storm assessment of impacts. Our proposed approach relies on the ability of deep learning models to learn complex spatial patterns, and therefore improve upon conventional flood inundation mapping techniques with going beyond binary classification and separate purely hurricane-induced and fluvial flooding from periodical flooding in coastal areas (e.g., wetlands, salt marshes and mangroves). Deep learning models with such capabilities can effectively delineate permanent water bodies and non-inundated urban areas, and hence expediting flood hazard assessments. This will provide useful information that facilitates the communication of National Water Center with its stakeholders for quick damage assessment and efficient resource allocation for emergency management. Improved spatial coverage of flood mapping with the help of remotely-sensed data and transfer-learning can further enhance flood forecasting in data scarce regions and help address equity issues in flood risk management.