Analyzing Tropical Cyclone Surface Wind Asymmetry with a Polar-Coordinate Periodic-Domain Deep Learning Model

Abstract Although tropical cyclone (TC) forecasts capture the track and primary rainfall distribution of a TC, their ability to forecast TC structural changes and the asymmetry of wind fields is limited. A major barrier to understanding TC structural changes is the rarity of observations and systematic analyses of TC winds. Therefore, a new method, the deep learning 2D structural analysis model for tropical cyclones (DSAT-2D), is proposed to perform 2D surface wind analysis at high spatiotemporal resolutions on the basis of generative adversarial networks (GANs). The model input includes satellite infrared and synthetic passive microwave images and environmental flow from numerical model analyses. The DSAT-2D model is trained on labeled data from modified scatterometer winds, in which underestimated wind speeds higher than 12 m s −1 are corrected. Furthermore, DSAT-2D is calculated in polar coordinates to help the model capture the rotational nature of TCs. This study also evaluates how the asymmetry of TC surface winds is affected by the interaction between a TC and the environment. Composite analyses with respect to different environmental factors are conducted. The DSAT-2D model can capture the asymmetric structure of TCs due to vertical shear and low-level flow. The interactions between these factors and TC wind asymmetry are discussed. Overall, the DSAT-2D model provides the possibility of studying TC wind asymmetry and improving TC forecasts by generating 2D surface winds with a generative artificial intelligence approach.