Convolutional Inverted Differential Attention Network for Subsurface Thermohaline Structure Reconstruction

Remote sensing and deep learning have advanced greatly, yet reconstructing global subsurface temperature and salinity structure remains challenging. This study proposes Convolutional Inverted Differential Attention Network(CIDAN), a novel deep learning framework integrating differential attention and multi-scale temporal modeling to enhance 3D thermohaline reconstruction. CIDAN consistently outperforms existing methods in both temperature and salinity reconstruction across all four seasons of 2023. Specifically, the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R$</tex><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> scores for temperature reconstruction reach 0.5545, 0.5305, 0.5452, and 0.5613 in spring, summer, autumn, and winter, respectively. For salinity, the corresponding <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R$</tex><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> scores are 0.3823, 0.3895, 0.3925, and 0.3801. These results demonstrate that CIDAN effectively captures complex vertical variations and dynamic oceanic processes, enabling accurate 3D thermohaline reconstruction from high-resolution satellite and Argo data.