A comparative analysis of deep learning models for classifying shallow mesoscale cloud patterns in satellite images

Abstract. Representation of clouds in climate models is challenging, not the least due to their heterogeneous spatial structures and dynamic behavior. In this study, the potential of advanced machine learning (ML) techniques to identify and categorize mesoscale low-level cloud structures in satellite imagery is explored, with particular emphasis on those patterns that are frequently observed over the trade wind regions of the south Atlantic Ocean.

Rectified Level 1.5 satellite images from the spinning enhanced visible and infrared imager (SEVIRI) for the year 2021 are used for the analysis. To assess the potential gains in classification accuracy under limited labeled datasets, several deep learning approaches are evaluated. The analysis considers a custom-built convolutional neural network, a pre-trained 50-layer residual neural network adapted through transfer learning using EuroSat, and a self-supervised vision transformer framework known as DINOv2 (self-distillation with no labels version 2). The embeddings, i.e. the feature representations yielded by DINOv2 are used in two separate approaches, one based on manually-labeled data and the other using the k-means clustering algorithm.

The results show that combining the DINOv2 model with a multilayer perceptron and training on labeled data achieves the highest cloud pattern classification accuracy among the evaluated ML approaches.