Image2image Tropical Cyclone wind field diagnosis with Pix2Pix generative adversarial networks (GANs) (Papers Track)
Sarah Ollier (Worldsphere)
Abstract
Recent AI advances enable tropical cyclone (TC) diagnosis from satellite imagery, extracting wind intensity and storm size for operational analyses and forecasts. While complete and sufficiently detailed horizontal wind fields are rarely observed directly, this study presents pix2pix generative adversarial networks (GANs) that predict surface wind fields from real and simulated infrared brightness temperatures using thousands of realistic TC image pairs. Applied to modified real infrared images, the algorithm produces realistic wind fields comparable to observations (6.7 knots root mean square error). Insights from this project will inform future development for live satellite imagery diagnosis to enable higher spatio-temporal resolution TC wind information earlier and faster for meteorologists and emergency managers.