An End-to-End Deep Learning Approach for Tropical Cyclone 2-D Surface Wind Analysis

Although tropical cyclone (TC) forecasts fairly capture the TC’s track and primary rainfall distribution, limited skills are found in forecasting TC structural changes and asymmetric gusty winds. A major barrier to further understanding TC structural change is the lack of observations and systematic analysis of TC winds. Therefore, a new method, Deep Learning 2-D Structure Analysis Model for Tropical Cyclones (DSAT-2D), is proposed to produce TC 2-D surface wind analysis at high-spatiotemporal resolutions based on generative adversarial networks (GAN). The model input includes satellite infrared and passive microwave images and environmental flow from the ERA5 reanalysis, while the DSAT-2D model is trained based on the labeled data of modified Scatterometer (ASCAT) winds, in which the underestimated wind speed and overestimated radius of the maximum wind are corrected. Furthermore, we transform all the calculation into polar coordinate to help the DSAT-2D model capture TC structure more efficiently.

We use the satellite synthetic aperture radar (SAR) imagery data during 2016–2020 as independent testing dataset to evaluate the model performance, and composite analysis with respect to different environmental factors are conducted. Results show that the DSAT-2D model has the capability to capture the asymmetric structure of TC caused by vertical wind shear and low-level flow. The interactions of 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 2-D surface winds from satellite images and model analysis.