Deep Learning for Improved Use of Satellite Observations

We present ongoing research in the application of deep learning to process satellite observations for weather forecasting in the Global Forecasting System (GFS) model. The deployment of new satellites, such as the Geostationary Operational Environmental Satellite (GOES)-16, has increased incoming data volume of several orders of magnitude, which is difficult to process in a timely manner using traditional methods. In collaboration with NOAA, we are working to adapt computer vision techniques to automate data analysis and scientific discovery on NVIDIA GPUs. Convolutional networks (CNNs) for classification and segmentation enable us to track regions of interests such as tropical cyclones, extra-tropical cyclones, cyclogenesis events, and convection initiation. NVIDIA's super slow-motion technique has the potential to smooth and augment the data stream and to fill in dropped or damaged frames. Conditional generative adversarial networks (cGANs) may be used to map observations onto model variables to facilitate model data assimilation. In this session we will discuss our efforts in these areas, describing what works, what doesn't work, and what we have yet to try.