Cross-scale 3-D thermohaline modeling via dual-residual swin transformer with multisource ocean observations

Integrating multisource Earth observation data and reconstructing subsurface thermohaline structures from remote sensing at global and basin scales will provide a better understanding of oceans. However, previous methods relied on layer-by-layer modeling, which required separate reconstruction of subsurface temperature and salinity fields at different depths, resulting in many models, inefficiency, and weak vertical connections between thermohaline data at different depths. A fast deep neural network-based reconstruction can reduce models and enhance the overall consistency of thermohaline data, which is significant for the reconstruction of ocean environmental variables. This study proposes an improved Swin Transformer approach, i.e. SwinOcean3D, to perform one-shot reconstruction of three-dimensional (3-D) thermohaline structures (upper 1000 m) in different scales of Global (1° × 1°) and Indian Oceans (0.25° × 0.25°) by integrating multisource remote sensing and observation-based ocean products. SwinOcean3D combines the Swin Transformer, U-net, and dual-residual blocks to enhance the representation capability of the global scale, local detailed, and vertical features of ocean thermohaline structures. The significant advantages of SwinOcean3D in the reconstruction of multiscale 3-D thermohaline structures outperform other classical approaches. Furthermore, interpretability experiments suggest that SwinOcean3D can effectively capture the evolution of 3-D thermohaline structures from multisource observations.