Neural Network Estimates of Cloud Water Content Profiles from Passive Satellite Imagery

Accurate estimates of three-dimensional cloud properties are essential for short-term prediction of visibility, assessing the potential for aircraft icing-related hazards, and more generally for our understanding of the earth’s weather and climate processes. Space-borne cloud-profiling instruments onboard the CloudSat and the Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellites have offered such observations in the past. However, their spatial and temporal coverage is extremely sparse relative to passive satellite imagery and insufficient for most near-real-time applications. Low-earth orbiting (LEO) satellites such as the Visible Infrared Imaging Radiometer Suite (VIIRS), and particularly geostationary (GEO) imagers such as the Advanced Baseline Imager (ABI) are much more relevant to time-sensitive applications, but their observations ultimately lack the same sensitivity to vertical variation of cloud properties. This presentation aims to explore the use of the information provided by high-spatial-resolution imagery from LEO and GEO satellites to inform the estimation of cloud water content at several vertical levels. To achieve this, we train neural networks to estimate vertical PDFs of cloud water content observed in collocated VIIRS and CloudSat/CALIPSO observations. Preliminary results illustrate that we are able to significantly improve upon previous methods. Furthermore, we illustrate a potential pathway towards transitioning ML models trained on VIIRS data to five different GEO imagers. This allows for global low- and middle-latitude estimates of vertically-resolved cloud water for single-layer non-precipitating clouds.