Predicting 3D atmospheric structure from geostationary satellites

3D atmospheric wind is a key missing observable in NASA#39;s Earth Observation System (EOS) and presents large initial condition uncertainties in data assimilation systems. Atmospheric motion vectors (AMVs), which track cloud/water vapor movement and assign a height, provide key passively sensed initial condition information for numerical weather prediction. However, data produced by current operational AMV algorithms are sparsely available and contain high uncertainties in the vertical direction. Comparison with radiosonde profiles indicates that AMV height assignment errors comprise 70% of uncertainty. Our work aims to fill EOS gaps by learning a mapping between high-temporal resolution GOES-16/17 geostationary satellites operated by NOAA/NASA and radiosonde observations. We use our WindFlow model that applies deep learning based optical flow techniques to track clouds and water vapor in sequences of geostationary images to produce direction and speed data. WindFlow provides an efficient approach to generate dense and accurate wind vectors. Outputs of WindFlow, paired with thermal infrared imagery, provides information content to directly predict wind speed, humidity, and temperature at varying pressure levels. This proposal aims to generate a 3D atmospheric product from geostationary satellites validated against radiosondes and lidar observations.