C3DIR: Deep learning model retrieves 3D cloud properties from satellite imagers

A team led by Charles H. White from NOAA and Colorado State University has introduced C3DIR (Cloud 3-Dimensional Imager Retrieval), a deep learning framework that extracts 3D cloud properties from passive satellite imagers. The model is trained to match retrievals from the Earth Cloud Aerosol and Radiation Explorer (EarthCARE) ACM-CAP product. Unlike traditional 2D retrievals, C3DIR predicts the occurrence and water content of ice, cloud liquid, and rain along the imager line-of-sight, using a novel voxel-level collocation method to align the differing viewing geometries of passive imagers and active profiling instruments. This allows vertical profiling across multiple imager pixels, constructed from voxels shared by those pixels. Qualitative evaluations show C3DIR can identify multiple distinct overlapping cloud layers, though with some smoothing. Quantitatively, the model excels at hydrometeor detection (especially for higher water content), but struggles with liquid clouds due to their small geometric thickness and tendency to be hidden within ice clouds. Water content estimation is most accurate for ice clouds, with higher uncertainties for liquid and rain. Column-integrated water paths align closely with EarthCARE. Comparisons with current NOAA operational products indicate several areas where C3DIR could offer improvements.

The research, published on arXiv (2607.16929), positions C3DIR as a step toward operational AI/ML 3D cloud retrieval. By providing vertically resolved cloud structure from multiple satellite imagers—not just active sensors like lidar or radar—the method broadens the utility of existing satellite data. Potential applications include aviation hazard detection, numerical weather prediction, and climate studies. The team highlights that while challenges remain, especially with liquid clouds and fine-scale structure, C3DIR demonstrates the feasibility of deep learning for flexible, 3D cloud property retrieval from passive instruments. This could lead to more continuous and global cloud observations, complementing limited active sensor coverage.