Cloud tomography drops from half an hour to 3.5 seconds: HybridCT also cuts extinction error on synthetic shallow clouds
Synopsis
Large-scale observations of the three-dimensional interior of low clouds have been missing because physics-based tomography repeats expensive 3D radiative-transfer calculations, a cost that has kept it from large, realistic cloud scenes. HybridCT splits the problem: a U-Net estimates cloud optical thickness projections from multi-angle passive radiances, and SART then explicitly inverts those projections into a 3D extinction field. On a synthetic low-cloud scene it reconstructs extinction with RMSE 13.3 km⁻¹ in 3.5 s, against 22.7 km⁻¹ in about 30 min for the physics baseline AT3D.
Interpretation
Multi-angle passive observations can now yield a 3D cloud extinction field directly, without iterating in radiance space. The earlier physics-based tomography framework AT3D uses 3D radiative transfer as its forward operator and inverts multi-angle radiances, which requires repeated expensive 3D radiative-transfer calculations. On a held-out synthetic low-cloud scene, HybridCT reconstructs extinction with RMSE 13.3 km⁻¹ versus 22.7 km⁻¹ for AT3D, taking 3.5 s against about 30 min.
Machine learning only converts each view's radiance into cloud optical thickness, while vertical structure is left to an explicit geometric projection constraint. A U-Net is trained separately per viewing angle to map 2D radiance to cloud optical thickness, giving path-integrated projections along each line of sight; SART then iteratively inverts these projections with positivity and a prescribed cloud mask. Training data come from six large-eddy simulation scenes of marine subtropical shallow cumulus, with MYSTIC Monte Carlo radiative transfer simulating MISR-like radiances at nine viewing angles within ±70° at 670 nm; one scene was held out for final evaluation.
Perspective
The demonstrated setting is synthetic multi-angle radiances generated from six large-eddy simulation scenes of marine subtropical shallow cumulus, with one scene held out for final evaluation; the U-Net is trained for MISR-like nine-view geometry and a limited range of environmental conditions.
Cloud optical thickness does not preserve all information in the original radiances; real-data application will require broader training, observation-based 3D cloud masking, and independent validation with active and in situ observations. The static-scene assumption requires near-simultaneous multi-angle acquisition, which next-generation satellite constellations such as CloudCT are expected to provide.