Abstract
Abstract
Cyclones, commonly known as typhoons and hurricanes, are among the most impactful meteorological disasters. Traditionally, cyclones have been identified and measured with scales, such as the Saffir-Simpson scale, generally requiring human effort. Recently, machine learning methods have demonstrated potential in accelerating the assessment of cyclones; however, their accuracy remains somewhat low, particularly for weak cyclone classes. In this study, we present a machine learning framework, called CycTrack, to detect, localize and classify cyclones given global atmospheric weather state. Combining three specialized submodels (for cyclone detection, localization and classification) in a sliding-window approach, CycTrack automatically tracks and classifies cyclones on the Earth. The power of our approach relies on several deep learning techniques, including an ensemble of convolutional neural networks and an Attention U-Net. Through elaborate evaluation on the Global Forecast System (GFS) data, we demonstrate that the performance of our models, especially for weak cyclone classes, ranks high in the literature.