CloudViT: exploring cloud type classification with vision transformers in global satellite data

Abstract. Clouds constitute, through their interactions with incoming solar radiation and outgoing terrestrial radiation, a fundamental element of the Earth's climate system. Different cloud types show a variety in cloud microphysical or optical properties, phase, or vertical extent, and thus disparate radiative effects. Both in observational and model datasets, classifying clouds is important since different cloud types respond differently to current and future anthropogenic climate change. Cloud types have traditionally been defined using a simplified partition of cloud top pressure and optical thickness, but recently using deep learning. In this study, we present a method called CloudViT (Cloud Vision Transformer) building on surface observations and spatial extracts of cloud properties from the MODIS instrument to derive cloud types, leveraging spatial patterns with a vision transformer model. The performance of the model is fair and hampered by the limited number of samples and the challenging matching between data sources arising during the colocation process. The method is then evaluated through the distributions of cloud type properties and global spatial patterns of cloud type occurrences. Potential improvements emerge in the reduction in mismatches between data sources, the extension of the colocated dataset, and the refinement of the classification model. While the application of the method in its current state comes with apparent uncertainties due to limited performance, it raises relevant challenges and limitations, from which the community can benefit from discussing for the development of similar methods. To foster future advancements, the dataset and model are available from Zenodo (Lenhardt et al., 2024b).