MSFFNet: A Multi-Scale Spatiotemporal Feature Fusion Network for Ocean Subsurface Temperature and Salinity Reconstruction

Reconstructing subsurface sea temperature and salinity from multi-source surface observations is important for understanding ocean dynamics and environmental variability, but it remains challenging because of strong spatial heterogeneity and complex spatiotemporal coupling. This paper presents MSFFNet, a spatiotemporal reconstruction network for subsurface thermohaline estimation. The proposed model consists of three coordinated components. A spatial feature extraction branch combines depthwise separable convolution with coordinate attention to enhance the representation of local gradients and direction-aware spatial structures. A temporal feature extraction branch adopts multiscale temporal modeling to capture historical variations at different temporal resolutions while preserving sequence-level temporal information. A spatiotemporal fusion block then performs bidirectional interaction between spatial and temporal features, followed by temporal conditioning and attention-based refinement, so that the two types of information can be integrated more effectively before final prediction. Experiments on western Pacific data from 2000 to 2020 demonstrate that MSFFNet consistently outperforms several competing baselines in both temperature-anomaly and salinity-anomaly reconstruction. The results further show that the proposed spatial enhancement and spatiotemporal fusion strategies make substantial contributions to the overall performance. These findings confirm the effectiveness of MSFFNet for subsurface thermohaline reconstruction in complex ocean environments.