Prediction of Tropical Cyclogenesis Using Vision Transformers

The formation of tropical cyclones remains a challenging forecast problem. Recent advances in deep learning-based weather prediction have yielded impressive results in the realm of medium-range weather forecasting. Vision transformers serve as the backbone architecture of some of these purely data-driven models. These models learn relationships between grid cells at a variety of spatial scales to generate a prediction. The success of vision transformers in general weather prediction suggests they may perform well when targeting specific forecasting phenomena such as tropical cyclogenesis.

We use ERA5 reanalysis and GridSAT data to train a neural network with a vision transformer backbone to predict the formation of tropical cyclones. We train different network versions to predict tropical cyclone formation at lead times of hours to days. We use these sets of predictions to generate forecast maps of where tropical cyclogenesis is possible and when it is most likely to occur. We compare model performance across multiple tropical basins and identify regimes of cyclogenesis where our model performs well and where it performs poorly. We also compare our results to those of a UNet-based model. Our results demonstrate a potential application of new deep learning methods in the realm of tropical cyclone forecasting.