A Machine Learning Approach to Resolving the Internal Structure of Clouds

The ability to resolve the internal structure of clouds is critical to generating accurate weather forecasts and timely decisions in an operational setting. Previously, cloud structure was remotely sensed using the CloudSat and CALIPSO satellites. These were low earth orbiting satellites that collected internal cloud structure data through narrow swaths with observations collected only twice per day for every point location. These satellites observed cloud structure using active sensors but were decommissioned in mid-2023. Other operational satellites, such as GOES, passively sense reflected solar radiation and emitted terrestrial radiation to infer the presence of clouds. Unlike CloudSat and CALIPSO, the collection of operational passive geostationary and polar-orbiting satellites provides global observations several times per day. However, this approach is insufficient for modeling the internal structure of clouds and assimilating initial conditions used in numerical weather prediction models for forecasting. In this study, we account for the missing CloudSat, CALIPSO, and other ancillary data by using machine learning algorithms to train operational GOES satellite data with CALIPSO data. This approach will allow us to use publicly available data to reconstruct the internal structure of clouds. Our experiment is done solely on the AWS Cloud platform to demonstrate the ease and portability of machine learning in a cloud environment. Further, our research demonstrates how to generate governable ML/AI results that focus on repeatability through industry standard MLOps. This research demonstrates the opportunity to replace data missing from CloudSat and CALIPSO, but also increase spatiotemporal coverage of internal cloud structure.