3D-OSR: Three-dimensional ocean salinity super-resolution reconstruction from remote sensing observations using dual-task learning

Accurately reconstructing high-resolution three-dimensional (3D) ocean subsurface salinity (SS) is crucial for understanding mesoscale processes. However, it is hindered by the sparsity of in situ data and the limitations of existing models. We propose a 3D ocean salinity super-resolution (SR) reconstruction model (3D-OSR), which employs a dual-task learning framework that integrates the strengths of reanalysis and in situ profile data to achieve eddy-resolving (1/12°) daily 3D SS reconstruction. Task 1 is to construct the 3D SR model, which learns to reconstruct spatially continuous fields from the reanalysis data. The 3D SR model effectively fuses signals from different spatial scales by parallel processing of multi-resolution satellite observations, without previous interpolation. Task 2 trains the bias correction model, which corrects errors in the outputs of Task 1: the 3D SR model according to the in situ profiles. The results show that the 3D SR model outperforms convolutional neural network-, Transformer-, and Vision Transformer-based reconstruction models, achieving a root-mean-square error (RMSE) of 0.152 PSU and correlation coefficient ( R ) of 0.883. Wavenumber spectra analysis shows that multi-resolution parallel processing in the 3D SR model prevents spectral discontinuities in PixelShuffle, Deconvolution, and Unpooling upsampling. After processing by the bias correction model in Task 2, the performance of the 3D SR model is further improved (RMSE: 0.127 PSU, R : 0.911), surpassing the reanalysis products CORA2, HYCOM, and ECCO2. The 3D-OSR model provides a cost-effective method for reconstructing accurate ocean 3D SS and offers new insights into ocean dynamics monitoring and research. 重构精确的高时空分辨率海洋三维盐度场对理解海洋中尺度过程至关重要.本研究设计了一种基于双任务学习框架的海洋三维盐度场超分辨率重构模型 (3D-OSR) , 其整合再分析产品与原位观测剖面的优势, 能够重构逐日,涡可分辨率 (1/12°) 的三维盐度场.任务1是构建三维超分辨率模型, 其参照再分析产品来重构空间连续的三维盐度场.三维超分辨率模型能够直接处理不同分辨率的卫星产品, 而无需预先插值.任务2是参照原位剖面来校正三维超分辨率模型的重构结果.结果表明, 三维超分辨率模型的性能优于其他主流重构模型.波数谱分析表明, 三维超分辨率模型避免了传统上采样方法的谱不连续问题.经过校正的三维超分辨率模型的精度超越了再分析产品CORA2,HYCOM和ECCO2.3D-OSR为重构精确的高时空分辨率海洋三维盐度场提供了经济高效的手段, 也为海洋中尺度过程的监测与研究提供了新的视角.