Enhancing Retrievals of Air-Sea Heat Fluxes From AMSR2 Microwave Observations Based on Deep Learning

Air-sea heat fluxes play a crucial role in understanding global climate variability. Using the bulk aerodynamic algorithm, we can derive sensible heat flux (SHF) and latent heat flux (LHF) from the satellite sea surface temperature ( ${T}_{s}$ ) and wind speed (WS), as well as the air temperature ( ${T}_{a}$ ) and specific humidity ( ${Q}_{a}$ ). However, traditional retrievals of ${T}_{a}$ and ${Q}_{a}$ tend to be unreliable. To address this issue, we introduced the Matrices-Points Fusion Network (MPFNet), a deep learning (DL) model designed to integrate spatial and point information. This model employs the Fourier Neural Operator (FNO) and Residual Network (ResNet) techniques to enhance retrieval capabilities. The model was pretrained with ECMWF ERA5 reanalysis-based data, followed by transfer learning (TL) with satellite and in situ matchup data for fine-tuning. Validation against independent in situ data showed significant improvements compared to the mainstream products in the community, with root mean square errors (RMSEs) for ${T}_{a}$ and ${Q}_{a}$ reduced to $0.59~^{\circ }$ C and 0.87 g/kg, representing 27%–41% and 16%–33% improvements, respectively. SHF and LHF RMSE values were 6.54 and 29.32 W/m2, reflecting improvements of 32%–36% and 17%–31%, respectively. Using this fine-tuned MPFNet model with satellite data as input, a global daily gridded dataset of ${T}_{a}$ , ${Q}_{a}$ , SHF, and LHF over 11.5 years was generated at a 0.25° resolution. The analysis showed that previous products tended to underestimate ${T}_{a}$ and ${Q}_{a}$ while overestimating SHF and LHF.