UNet5 and Other Vision Architectures for Peak Storm Surge, in Global Storm Surge Machine Learning Models

This project contains trained models and processed datasets for predicting storm surge at any location in the globe. The models are trained from a global database of ADCIRC simulations for synthetic tropical cyclones. The machine learning approach is inspired by existing vision models. All features are interpolated to a 2.5x2.5 degree latlon grid with 128x128 cells about the storm landfall location. The input features consist of bathymetry, wind, and pressure fields. Wind and pressure are sampled every three hours starting one day before landfall and ending twelve hours after landfall. This results in thirteen features for pressure, as well as the x- and y- components of wind respectively. In addition to the thirty-nine meteorological features, bathymetry is represented twice: once with raw values, and once with a mask indicating the presence of land. This leads to a total of forty-one input channels for each vision model architecture in this work.
The best performing model is UNet-5, but results from other models are included for comparison purposes. We also include some UNet-5 models trained on regional subsets of the training data. The global model outperforms the regional models, indicating that training data from other ocean basins can improve performance, and that the model is learning the physics of storm surge instead of memorizing location-specific patterns.