Estimation of nighttime cloud properties using machine learning

Measurements of the diurnal cycle of stratocumulus cloud properties are important to understanding the physical processes governing the life-cycle of boundary layer clouds, which play a crucial role in setting the planetary albedo. Retrievals of cloud macro and microphysical parameters using reflected solar radiation are now routine, but nighttime retrievals using emitted photons remain challenging. A key need is identifying cloud-regions that can be monitored as winds advect the boundary layer. This identification has been previously accomplished using Lagrangian particle tracking in wind-fields derived from forecast reanalysis, but the approach is time-consuming and subject to uncertainty. Recent work has revealed the ability of convolutional networks to vastly expand the inventory of identifiable ship tracks, which are long-lived features in stable air masses in eastern ocean basins near large population centers. The use of ship tracks presents the opportunity to identify and follow such scenes, allowing for enhanced tracking and retrievals of cloud properties in advected air throughout a diurnal cycle. We examine ship tracks as a single source proxy for cloud airmass location, and use available daytime retrievals of cloud fraction, liquid water path, droplet size, and cloud top temperatures to characterize the boundary layer. After training a neural network on daytime scenes with available shortwave retrievals and thermal data from the Aqua/Tera MODIS, GOES 17 ABI, and Aqua EMSR, we estimate these cloud properties at night along the Lagrangian trajectory marked by the ship tracks. Testing and validation reveal these cloud properties can be accurately estimated in the absence of available daytime data sources, allowing for improved characterizations of nighttime clouds.