Transformer Aided Construction of a Long-term Tropical Cyclone Concentric Eyewalls Dataset

Concentric eyewalls (CEs) and associated eyewall replacements are key processes in the evolution of tropical cyclone (TC) intensity and structure. Long-term data of CEs are of great significance in TC research. Using long-term (1991-2020) global satellite observations, we developed a Transformer-based algorithm to automatically identify CEs from passive microwave (PMW) imageries and present a 30-year dataset of CEs with global coverage. Through random rotation to alleviate class imbalance and greedy soup method to boost the classification performance, precision and recall for CE detections achieve 92.9% and 92.9%, respectively. To the best of our knowledge, this is the first time that deep learning has been used for CE identification. Aided by the end-to-end algorithm, the 30-year CE dataset is generated, including 377 CE events (611 CE snapshots), far more than any previous study. The publicly available dataset, which incorporates the best-track TC information, provides a comprehensive and in-depth view for understanding the nature of CEs.