Gaussian Mixture Model-Based Cloud- Phase Estimation From GEO- KOMPSAT-2A Observations

Cloud-phase algorithms based on satellite infra- red (IR) observations typically use empirical thresholds and additional cloud properties to predict cloud phases, but these settings may not be consistently accurate under various conditions. This study introduces an unsupervised machine learning (ML) approach based on a Gaussian mixture model (GMM) to avoid the empirical determination of thresholds and auxiliary cloud products for cloud-phase estimation. The GMM-based cloud-phase algorithm proposed in the present study consists of three GMMs that distinguish cluster types representing water, ice, and undetermined phases using the brightness temperature (TB) at
$11.2 \mu \text{m}$
and the difference in TB between 8.6 and
$11.2 \mu \text{m}$
from Geostationary-Korean Multi-Purpose Satellite-2A (GEO-KOMPSAT-2A, GK2A) satellite. The first GMM initially classifies four clusters, while the second and third GMMs regroup the initial clusters to distinguish supercooled water, mixed-phase, and optically thin cirrus clouds. The GMM-derived estimates are compared with operational cloud-phase products from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Cloud-Aerosol Lidar with Infrared Pathfinder Satellite Observations (CALIOP). Results show that the water and ice phases estimated using the GMM-based algorithm are in good agreement with both the MODIS and CALIOP products. The GMM-based algorithm also significantly reduces the misidentified area for undetermined phases observed in the GK2A operational product. Water and ice phases are also effectively estimated in warm regions, resulting in distributions similar to those derived from MODIS and CALIOP products. Unlike most IR cloud-phase algorithms that utilize thresholds and other cloud parameters, the GMM-based cloud-phase algorithm has the advantage of using only TB, thus avoiding auxiliary cloud properties.