Retrieval of Inner-Core Tropical Cyclone Surface Winds from Satellite-Based Microwave Imagers and Sounders with Regression-Based UNet Models

Satellite-based microwave sounders and imagers are critical for assessing the state of a tropical cyclone. The scattering and attenuation signals are commonly used both subjectively and empirically to estimate eye wall properties, banding and inner core heating, which then indicate maximum sustained winds and central pressure. This project uses deep learning through regression-based UNet models to take the conventional process a step further and infer the full 2D surface windspeed of the inner core from the microwave channels of AMSR, TMI, SSMI/S, AMSU/MHS or ATMS. The model is trained on direct aircraft wind measurements and extrapolated to the surface; there is also a second independent confirmation with SAR windspeed observations. We find that this model can be competitive with other state of the art models and reveals a uniquely detailed wind distribution. It is able to show multiple eye walls, eyes smaller than the instrument resolution, and sharp drop-offs in windspeed with radius. Another unique trait is the ability to accurately depict wind profiles in transition and thus not captured by steady-state models. We also compare the relative impact of various satellite sensors on the model to show the importance of sensor resolution in windspeed retrieval skill.