A deep learning-based global tropical cyclogenesis prediction model and its interpretability analysis

Abstract

Tropical cloud clusters (TCCs) can potentially develop into tropical cyclones (TCs), leading to significant casualties and economic losses. Accurate prediction of tropical cyclogenesis (TCG) is crucial for early warnings. Most traditional deep learning methods applied to TCG prediction rely on predictors from a single time point, neglect the ocean-atmosphere interactions, and exhibit low model interpretability. This study proposes the Tropical Cyclogenesis Prediction-Net (TCGP-Net) based on the Swin Transformer, which leverages convolutional operations and attention mechanisms to encode spatiotemporal features and capture the temporal evolution of predictors. This model incorporates the coupled ocean-atmosphere interactions, including multiple variables such as sea surface temperature. Additionally, causal inference and integrated gradients are employed to validate the effectiveness of the predictors and provide an interpretability analysis of the model’s decision-making process. The model is trained using GridSat satellite data and ERA5 reanalysis datasets. Experimental results demonstrate that TCGP-Net achieves high accuracy and stability, with a detection rate of 97.9% and a false alarm rate of 2.2% for predicting TCG 24 hours in advance, significantly outperforming existing models. This indicates that TCGP-Net is a reliable tool for tropical cyclogenesis prediction.