Enhancing Subsurface Thermal Structures Reconstruction via a Dual-Branch Framework Integrating Surface Remote Sensing and Model-Predicted Upper-Layer Fields

Driven by satellite remote sensing observations, deep learning models provide a promising solution for the rapid estimation of subsurface thermohaline structures. By learning from large-scale paired surface–subsurface datasets, these models establish nonlinear mappings between surface variables (e.g., sea surface height, temperature, wind, and heat flux) and interior ocean features, enabling data-driven reconstruction of the underwater environment. However, limited attention to subsurface stratification and mesoscale variability often results in reduced accuracy in dynamically active regions, such as the Kuroshio Extension. To overcome these limitations, we propose a dual-branch deep learning architecture that integrates both surface observations and subsurface features to improve subsurface temperature reconstruction. A layerwise progressive reconstruction strategy is incorporated, allowing model-predicted upper-layer fields to inform deeper-layer estimations. The model is evaluated using GLORYS12V1 reanalysis data in the northwestern Pacific, with a focus on the Kuroshio Extension. Results indicate that the proposed framework outperforms conventional surface-driven approaches, particularly within and below the thermocline. It achieves reduced anomalous deviations, improved structural coherence, and better generalization in nearshore regions with sparse training data. Spectral analysis further confirms that the subsurface vertical structure extraction branch effectively suppresses high-frequency noise while preserving mid- and low-frequency energy, supporting the model’s capacity to recover multiscale thermal features.