A Machine Learning Based Cloud Mask and Thermodynamic Phase Classification Method using Suomi-NPP VIIRS Spectral Observations

Clouds play critical roles in the Earth’s energy budget due to their large coverage and strong radiative effect. Among various important cloud properties, cloud thermodynamic phase is a critical midway to link cloud microphysical properties with cloud optical and radiative properties. In this study, we trained two machine learning (Random Forest) based models for cloud detection and thermodynamic phase classification for Suomi-NPP VIIRS. Specifically, a daytime model that uses 5 visible/near-infrared (VNIR) and three infrared (IR) bands is trained for daytime conditions and an all-day model that only uses the 3 IR bands is trained for both daytime and nighttime. The two Random Forest models are trained using 4-year collocated VIIRS/CALIOP data from 2013 to 2016 for 7 different surface types, namely, ocean/water, forest, shrub land, grass land, cropland, snow/ice, and barren. The CALIOP level-2 Layer products are used as reference. The two models are evaluated using 1-year collocated VIIRS/CALIOP data in 2017. It shows that the two models work well for all surface types and better than current MODIS/VIIRS operational cloud mask and phase algorithms.