Generating Synthetic Visible Satellite Data with Machine Learning

The GOES-16 and GOES-18 weather satellites provide imagery in both visible and infrared (IR) spectra, as well as indications of lightning events throughout the Western Hemisphere. While the IR imagery and lightning data are available 24 hours a day, the visible channels only provide useful data during daytime hours due to their reliance on solar reflectance. In recent years, machine learning models, particularly deep neural networks, have been used to enhance the utility of satellite information through improved nowcasting, gap-filling, and estimation of meteorological fields typically estimated from ground-based radar. In this work, we present an application of deep learning that uses GOES IR and lightning information to produce synthetic visible imagery, enabling access to visible-like data 24 hours a day. We train our model using the SEVIR dataset: a publicly available set of severe storm data consisting of one channel of visible data (GOES-16 channel 2), along with two channels of IR data (GOES-16 channels 9 and 13), and lightning flashes from the GOES-16 GLM. This talk will provide an overview of the model, verification results, as well as an investigation on how synthetic visible data can be incorporated into other algorithms to improve nighttime performance.