Artificial intelligence (AI)-derived 3D cloud tomography from geostationary 2D satellite data

Abstract

Satellite instruments provide high-temporal-resolution data on a global scale, but extracting 3D information from current instruments remains a challenge. Most observational data are two-dimensional (2D), offering either cloud top information or vertical profiles. We trained a neural network (Res-UNet) to merge high-resolution satellite images from the Meteosat Second Generation (MSG) Spinning Enhanced Visible and InfraRed Imager (SEVIRI) with 2D CloudSat radar reflectivities to generate 3D cloud structures. The Res-UNet extrapolates the 2D reflectivities across the full disk of MSG SEVIRI, enabling a reconstruction of the cloud intensity, height, and shape in three dimensions. The imbalance between cloudy and clear-sky CloudSat profiles results in an overestimation of cloud-free pixels. Our root mean square error (RMSE) accounts for 2.99 dBZ. This corresponds to 6.6 % error on a reflectivity scale between −25 and 20 dBZ. While the model aligns well with CloudSat data, it simplifies multi-level and mesoscale clouds in particular. Despite these limitations, the results can bridge data gaps and support research in climate science such as the analysis of deep convection over time and space.