Low-level marine clouds evolve rapidly through interactions among boundary-layer dynamics, cloud microphysics, aerosol perturbations, and mesoscale organization. Continuous tracking of their evolution is essential for characterizing cloud life cycles and constraining aerosol–cloud–radiation interactions. However, geostationary visible imagery, which provides the fine-scale cloud textures required for feature-based tracking, is unavailable at night. Thermal infrared observations provide continuous coverage, but their coarser resolution and weaker contrast for low clouds limit their ability to preserve cloud boundaries, ship-track signatures, and mesoscale cloud textures across the day–night transition.

Here we develop a hybrid machine-learning framework to reconstruct nighttime visible-like imagery and apply it to continuous day–night tracking of marine low clouds. Using observations from the Geostationary Operational Environmental Satellite-17 (GOES-17) Advanced Baseline Imager (ABI), a U-Net-based super-resolution convolutional neural network first translates multispectral thermal infrared information into a high-resolution 0.64 μm visible-like reflectance field. A nighttime 3.9 μm residual enhancement is then used to recover localized cloud contrasts that are weakly constrained by longwave infrared channels. Finally, a conditional diffusion model refines the reconstruction by restoring cloud-edge sharpness and high-frequency textures while remaining anchored to the infrared-derived large-scale cloud structure.

Evaluation against withheld daytime ABI visible observations shows that the U-Net reconstruction captures the broad radiometric structure with a coefficient of determination (R²) of 0.918. The diffusion refinement improves texture fidelity, high-frequency detail recovery, and edge preservation relative to the deterministic U-Net output. The reconstructed nighttime imagery is merged with daytime visible observations using a solar-elevation-based switching strategy and then used for optical-flow tracking at 10 min resolution. Case studies over the northeastern Pacific show robust tracking of ship tracks, shallow cumulus, and marine stratocumulus across the full diurnal cycle. Multi-scale structural similarity (MS-SSIM) generally remains above 0.7 at night and recovers to above 0.95 during daytime, whereas trajectories from the Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model frequently diverge from the observed cloud features within a few hours.

These results demonstrate that diffusion-refined nighttime visible reconstruction can bridge the nocturnal gap in geostationary visible observations and provide a consistent image sequence for Lagrangian tracking of marine low-cloud evolution. This framework offers a new basis for linking cloud structural evolution with aerosol loading, meteorological forcing, microphysical changes, and radiative effects across the diurnal cycle.