Global open-ocean daily turbulent heat flux dataset (1992–2020) from SSM/I via deep learning

Abstract. Air–sea turbulent heat fluxes – latent heat flux (LHF) and sensible heat flux (SHF) – are fundamental to the Earth’s energy and moisture budgets and to ocean–atmosphere coupling. Global flux estimates via bulk aerodynamic algorithms depend on sea surface temperature (SST), surface wind speed (SSW), near-surface air temperature (T_a), and specific humidity (Q_a), but orbital sampling and cloud contamination leave gaps in satellite inputs that propagate uncertainty to T_a/Q_a and hence to LHF/SHF. Here we present DeepFlux, a global daily 1° × 1° heat-flux dataset for 29 years (January 1992–December 2020). The dataset is produced with a concise completion-then-retrieval workflow: Special Sensor Microwave/Imager (SSM/I) variables (SSW, cloud liquid water, total column water vapor, and rain rate) are first gap-filled using the AI-based Generalized Data Completion Model (GDCM) to yield spatiotemporally continuous inputs; these – together with Optimum Interpolation SST (OISST) – are then used to retrieve T_a and Q_a via the AI-based Matrices-Points Fusion Network (MPFNet). LHF and SHF are then computed using a bulk algorithm. Validation against in-situ buoy observations shows that the dataset closely matches the true measurements, with RMSEs of 0.53 °C (T_a), 0.70 g kg⁻¹ (Q_a), 5.53 W m⁻² (SHF), and 25.28 W m⁻² (LHF). Comparisons with widely used flux products indicate differences among products, reflecting variability in flux estimates from different sources. DeepFlux provides an open, consistent, observation-constrained view of near-surface meteorology and air–sea heat exchange for climate diagnostics, model evaluation, and process studies. DeepFlux v1.0 is openly available under CC BY 4.0 at [repository] (DOI: http://dx.doi.org/10.12157/IOCAS.20250823.001). If you want to download without registering you can visit https://zenodo.org/records/17160579.