Discovering the Importance of Mesoscale Cloud Organization Through Unsupervised Classification

The representation of shallow trade wind convective clouds in climate models dominates the uncertainty in climate sensitivity estimates. In particular the radiative impact of cloud spatial organization is poorly understood. This work presents the first unsupervised neural network model which autonomously discovers cloud organization regimes in satellite images. Trained on 10,000 GOES‐16 satellite images (tropical Atlantic and boreal winter) the regimes found are shown to exist in a hierarchy of organizational scales, with sub‐clusters having distinct radiative properties. The model requires no time‐consuming and subjective hand‐labeled data based on predefined structures allowing for objective study of very large data sets. The model enables the study of environmental conditions in different organizational regimes and in transitions between regimes and objective comparisons of model behavior with observations through cloud structures emerging in both. These abilities enable the discovery of previously unknown physical relationships in cloud processes, enabling better representation of clouds in weather and climate simulations. Key Points A neural network can learn without supervision to discover similar cloud structures in satellite imagery The groups of cloud structures produced by the neural network have distinct radiative properties and thus distinct climatic impacts The neural network learns to represent similarity between cloud structures allowing for the study of the temporal evolution between regimes