Exploring North Atlantic Marine Cloud Morphology using Self-Supervised Machine Learning

Marine clouds form a variety of morphological structures. These different cloud structures form under different meteorological conditions and have distinct radiative and microphysical properties. A more nuanced understanding of the properties of these different cloud morphologies will likely improve our understanding of the climate system. Several past studies have cataloged and studied specific cloud regimes in satellite imagery using supervised machine learning methods. Typically, these studies have focused on 2-6 specific pre-defined cloud types. In this work, we train a fully self-supervised deep learning model on a year of global 250m resolution MODIS marine cloud imagery. The model learns a vector embedding space that captures the spatial structure of clouds in the satellite images. A clustering algorithm is then applied in the embedding space to identify unique cloud structures that commonly occur in the North Atlantic region. It is shown that a large range of unique cloud morphologies occur in this region, and 10-20 cloud morphology definitions are likely required to capture the full range of diversity. Several cloud regimes identified by the self-supervised algorithm are presented along with their climatological distribution and satellite derived cloud properties.