Learning to Generate 3D Atmospheric Winds from Geostationary Imagery with Deep Optical Flow
Satellite-based retrievals of 3D atmospheric structure are of high-importance to improve numerical weather prediction (NWP) and our understanding of the planetary boundary layer. Observations of the vertical structure are sparse and noisy, particularly winds, from radiosondes and low-earth orbit satellites such as NASA’s Cloudsat/Calipso (clouds and humidity), microwave sounders, and ESA’s Aeolus (winds). Atmospheric motion vectors (AMVs) aim to fill the gap in wind observations by making numerical estimates of cloud movement between sequences of multi-spectral Geostationary (GEO) satellite images by tracking clouds and water vapor. Recent imaging hardware and software advancements have enabled the use of numerical optical flow techniques to produce accurate and dense vector fields outperforming traditional methods. We introduce WindFlow as the first machine learning based system for feature tracking AMVs using deep optical flow. Experiments show that WindFlow is a robust method for dense feature tracking and presents more opportunities for machine learning advancements in satellite based Earth observations. We further present an approach to predict the complete 3D atmospheric structure of temperature and humidity from GEO thermal infrared imagery by learning from incomplete radiosonde profiles. Using sequences of 3D humidity and WindFlow, Zeus AI is capable of producing dense atmospheric winds in near-real time. Lastly, we discuss data assimilation efforts and challenges ingesting dense atmospheric winds into Goddard’s Earth Observation System (GEOS).
