A convolutional neural network was used to detect occurrences of pockets of open cells (POCs). Trained on a small hand-logged dataset and applied to 13 years of satellite imagery the neural network is able to classify 8,491 POCs. This extensive database allows the first robust analysis of the spatial and temporal prevalence of these phenomena, as well as a detailed analysis of their micro-physical properties. We find a large (30%) increase in cloud effective radius inside POCs as compared to their surroundings and similarly large (20%) decrease in cloud fraction. This also allows their global radiative effect to be determined. Using simple radiative approximations we find that the instantaneous global mean top-of-atmosphere perturbation by all POCs is only 0.01Wm-2.