Turbulence Nowcasting through Machine Learning with Geostationary Satellite and NWP

Aviation turbulence remains a confounding nowcasting challenge that requires all available sources of atmospheric retrieval as well as unconventional methods of modeling. We present the latest version of a deep learning model (a fully convolutional CNN) that brings together geostationary imagery and Global Forecasting System model fields to produce a 3D probability field of moderate-or-greater turbulence. Independent testing statistics show that the model is competitive with state-of-the-art systems such as GTG and is applicable globally. This method of nowcasting also reveals the special challenges encountered by deep learning to convert the sparse dataset of aircraft turbulence observations into a generalizable model: orographic influences, rare severe events, image resolution, and others. Our machine learning model will be made operational at NOAA NESDIS in the near future, and we also report on this progress.