ObsGate-SST: observation-gated background–detail learning for global daily sea surface temperature reconstruction

Infrared cloud screening and incomplete sampling limit the process-scale use of daily satellite sea surface temperature (SST), whereas many gap-filled Level-4 analyses attenuate narrow fronts, tropical-cyclone cold wakes and short-lived warm extremes. This study presents ObsGate-SST, an observation-gated background–detail learning framework for global daily SST reconstruction. The model ingests cloud-contaminated ODYSSEA L3S SST, a valid-pixel mask and an OISST v2.1 prior. A low-frequency branch estimates the basin-scale thermal background, a high-frequency branch reconstructs gradient and texture residuals from observed pixels, and a deterministic observation gate copies all valid satellite observations exactly so that inference is restricted to missing open-ocean pixels. Training used 2000–2020 daily fields, validation used 2021–2022 and independent testing used 2023–2025. Against 2023 quality-level-5 iQuam matchups, monthly bias ranged from 0.193 to 0.291 °C, RMSE from 0.637 to 0.849 °C and correlation from 0.993 to 0.996 across 902,665 collocations. A cloud-gap-shaped withheld-observed-pixel audit isolated inferred-pixel skill: ObsGate-SST achieved the lowest pooled RMSE (0.4802 °C), while removing the low-frequency branch, high-frequency branch, gradient/detail loss and physical/prior regularization increased RMSE by 17.0 %, 9.9 %, 3.9 % and 2.0 %, respectively. Relative to DINEOF, U-Net, Swin-Unet and SegNet under common masks, ObsGate-SST better retained spectral energy, frontal continuity, marine-heatwave structure, typhoon cold wakes and equatorial Pacific gradients. These results support observation-gated multiscale learning as a physically interpretable strategy for reconstructing incomplete satellite SST without sacrificing observed-pixel fidelity.