A large-scale flood modeling using geometry-adaptive physics-informed neural solver and Earth observation data

Abstract

Large-scale hydrodynamic models generally rely on fixed-resolution spatial grids and model parameters as well as incurring a high computational cost. This limits their ability to accurately forecast flood crests and issue time-critical hazard warnings. In this work, we build a fast, stable, accurate, resolution-invariant, and geometry-adaptative flood modeling and forecasting framework that can perform at scales from continental to global. We achieve this by combining continuous flow modeling of a geometry-adaptive physics-informed neural solver (GeoPINS) with the authenticity of Earth observation data. Specifically, the GeoPINS is proposed based on the advantages of no training data in physics-informed neural networks (PINNs), as well as possessing a fast, accurate, and resolution-invariant architecture through the implementation of Fourier neural operators (FNO). In particular, to adapt to complex and irregular geometries that exist in rivers, we reformulate PINNs in a geometry-adaptive space by taking full advantage of coordinate transformations and the efficiency of numerical methods in solving the spatial gradient. We validate our GeoPINS on popular partial differential equations on both regular and irregular domains, demonstrating fast, stable, and accurate performance, as well as resolution-invariant, geometry-adaptive properties. Next, due to a lack of large-scale ground truth data, time-series flood records are generated using freely available Sentinel-1 data and a SAR-based flood mapping algorithm. These flood records are used as boundary conditions and flood inundation extent verification of the proposed hydrodynamic model. Finally, we compare our GeoPINS results with a 30 m resolution, SAR-based flood record, and measured discharge from gauging stations, obtaining good agreement among the three. The experimental results for the Pakistan flood in 2022 indicate that the proposed method can maintain high-precision large-scale flood dynamics solutions at different resolutions and flood hazards can be forecast in real-time with the aid of reliable precipitation data.