Physics-based data augmentation to enhance deep learning performance in tropical cyclone and storm surge modelling

Accurate modelling of tropical cyclones (TCs) is essential for reliable storm surge simulations, as TCs provide the primary external forcing. While deep learning models have demonstrated effectiveness in TC modelling, a challenge remains regarding the limited diversity of TCs in reanalysis data used for training-specifically, the scarcity of extreme TC events. This study proposes a physics-based data augmentation method that utilizes a numerical weather prediction (NWP) model to physically generate TCs lying beyond the range of reanalysis data. Subsequently, focusing on small-sized yet intense TCs as extreme cases, a model-originally pre-trained solely on reanalysis data-was fine-tuned using this augmented dataset to convert a parametric TC model (PM) field into an NWP-like field. Validation using test data mimicking extreme TCs and a storm surge hindcast of TC Faxai (a compact, intense TC that struck Tokyo Bay in 2019) revealed that the PM failed to simulate storm surges where topographic effects on the wind field are significant, and the pre-trained model underestimated wind speeds and storm surges. In contrast, the fine-tuned model successfully captured the spatiotemporal features of the extreme TCs and the peak storm surges upon TC landfall, achieving the lowest RMSE for storm surges across all TCs and tide gauges. These results suggest that physics-based data augmentation can effectively extend the applicability of deep learning models for TC and storm surge modelling to extreme events.