Detailed classification of marine shallow clouds with large cloud fraction using artificial intelligence

Marine shallow clouds are highly abundant over subtropical oceans and play a key role in regulating Earth’s radiative balance, exerting a significant net cooling effect. However, their representation in current climate models remains incomplete, contributing substantially to uncertainties in climate projections and cloud feedback mechanisms. Marine shallow clouds have conventionally been classified into closed cells, open cells, and shallow cumulus clouds. Recent advances in deep learning have expanded this classification by enabling the automatic detection of additional pattern types. Yet, even within these pattern classes lies a spectrum of finer patterns that have not been systematically identified, leaving their underlying dynamics and radiative impacts underexplored.
We present an AI architecture for the classification of marine shallow cloud satellite images into newly defined fine pattern classes. We focus specifically on marine shallow cloud regimes with large cloud fraction values (above 0.9), which we partition into four classes: closed cells, closed cloud streets (rolls), stratiform clouds and unorganized convection. This finer partitioning enables the extraction of richer information from large cloud fraction conditions and a more detailed investigation of physical processes governing their organization. The training dataset was labeled by the Weizmann Cloud Physics Group members. Explainable AI tools are used to analyze the model’s internal representations and learn how it differentiates between the new classes. Applying the trained model to a large number of satellite images enables us to construct a novel, comprehensive and systematically classified database of cloud patterns. The availability of this extensive dataset allows the use of remote sensing cloud properties to characterize the unique features of each regime, and radiative flux data to assess their distinct radiative behaviors, despite their similarly high cloud fraction values. This work provides a clearer understanding of marine shallow cloud patterns and offers insight into the relationships between cloud morphology, underlying dynamics, and radiative effects.