Image-to-Image Tropical Cyclone Wind Field Diagnosis

Learn how to apply generative adversarial networks (GANs) for tropical cyclone wind field prediction from satellite imagery in this 53-minute conference presentation. Discover two innovative Pix2Pix GAN models that transform single-channel infrared satellite observations into complete 10-meter wind field reconstructions over 666 km × 666 km domains at high spatial resolution. Explore the wrf2wrf model trained on Weather Research and Forecasting Model simulations and the mir2para model using real satellite observations paired with parametric wind fields from extended best-track datasets. Examine cross-validation results showing realistic wind field generation with mean pixel-wise RMSE of 6.7-7.3 knots and learn how these models capture storm asymmetry, rainbands, and eyewall structures. Understand the operational advantages of this approach, including higher temporal resolution capabilities that convert new tropical cyclone infrared satellite images every 10-15 minutes into detailed wind field analyses with minimal computational cost, offering significant improvements over existing AI and operational methodologies for tropical cyclone analysis and forecasting.