Assessment of tomography-derived three-dimensional cloud information and its suitability for downstream applications

Abstract. Meteorological geostationary satellites carry passive imagers that provide multispectral observations over a large fraction of the Earth disk every 5–15 min. Cloud properties retrieved from these observations are primarily two-dimensional, as they are inferred from top-of-atmosphere radiances, and contain only limited information on vertical cloud structure. Recent SEVIRI-CloudSat tomography studies have shown that it is possible to infer CloudSat-like three-dimensional cloud structure from multispectral passive imagery. The main open question is therefore no longer feasibility of such approach, but the reliability of the reconstructed cloud information for downstream applications. The present study addresses this gap through a systematic performance evaluation of a SEVIRI-CloudSat tomography method on an independent year-long (2019) dataset of 17,110 scenes. The analysis assesses the fidelity of the estimated reflectivity, but also evaluates cloud occurrence, cloud geometry, vertically aggregated summaries derived from the 3D information, and their dependence on different variables. Scene-wise reflectivity RMSE over all CloudSat-valid voxels, including cloudy and clear-sky voxels is consistent with previous studies (2.76 dBZ). However, evaluation restricted to radar-detectable cloudy voxels yields a much larger median RMSE (14.48 dBZ) and bias (-11.16 dBZ), indicating systematic attenuation of moderate and strong echoes. Cloudy-voxel fraction is only weakly biased (-2.0 %), whereas cloudy-column fraction is underestimated by -17.4 %, consistent with missed observed-cloudy columns (POD = 0.63) and few false detections (FAR = 0.031). Cloud top and base heights remain informative when evaluated only where cloud geometry is defined in both observation and prediction (about 60 % of observed-cloudy columns) with median errors of 1.38 and 1.41 km. The scene-level vertical cloud centroid provides a complementary aggregate summary, with 96 % cloudy-scene recall, r = 0.89, and an RMSE of 1.04 km, while 97.4 % of observed multilayer columns collapse to single-layer reconstructions. The results show that SEVIRI-CloudSat tomography is limited in resolving fine-scale vertical structure and multilayer organization, but remains informative for bulk descriptors such as cloud fraction, vertical cloud centroid, and cloud top and base heights.