Urban Flood Modeling and Forecasting with Deep Neural Operator

A practical tool and community that can support effective, cross-scenario, and downscaled spatiotemporal urban flood forecasting, featuring the Urban Flood Forecasting Benchmark Dataset and machine learning method codes.

[Urban Flood Forecasting Benchmark Dataset] This dataset is introduced for evaluating various neural urban flood modeling and simulation methods. Two urban areas in Berlin, Germany are considered. For both areas, simulations are conducted using 125 distinct design storm rainfall events.

[Codes for DNO and TL-DNO] A deep neural operator (DNO) is designed for fast, accurate, and resolution-invariant urban flood forecasting. The DNO features an enhanced Fourier layer with skip connections for improved memory efficiency, alongside a deep encoder-decoder framework and an urban-embedded residual loss to enhance modeling effectiveness. Additionally, we propose a transfer learning approach, termed transfer learning-based DNO (TL-DNO), to enhance cross-scenario forecasting capabilities.