Emulating Daytime Cloud Optical Properties at Night with Machine Learning

Estimates of cloud optical properties from passive satellite imagers are often most reliable during the daytime. Typical retrieval algorithms rely on a pair of visible and near-infrared channels to determine cloud optical depth and effective radius from which an estimate of cloud water path can be derived. Unfortunately, IR channels lack spectral sensitivity to optically thick clouds making nighttime retrievals extremely difficult. We investigate whether spatial context – information about local organization of cloud features – can overcome the lack of spectral sensitivity in the IR channels to produce more reliable cloud optical properties at night. We train a U-Net neural network to emulate the GOES-16 ABI NOAA operational Daytime Cloud Optical and Microphysical Properties (DCOMP) algorithm using only IR channels. We evaluate this approach with several intercomparisons between the U-Net, the daytime and nighttime NOAA operational products, and Nighttime Lunar Cloud Optical and Microphysical Properties (NLCOMP) which uses a nighttime low-light visible band present on VIIRS in collocated imagery. These comparisons show that the use of spatial context is critically important in improving estimates of nighttime cloud optical properties. The U-Net can reasonably match DCOMP during the day and significantly reduces artifacts associated with day/night terminator. We also find that the U-Net estimates align more closely with NLCOMP at night compared to the nighttime NOAA operational products for ABI.