Satellite-based reconstruction of high-resolution ocean subsurface temperature using spatiotemporal graph attention networks

• STGAT model is developed to reconstruct high-resolution subsurface Temperature from satellites. • Multi-source satellite data and DSTAG enable high-accuracy reconstruction of ocean subsurface temperatures. • Joint training on reanalysis and EN4 data enhances subsurface temperature fidelity. The distribution and variation of ocean temperature are closely associated with ocean dynamic processes. High-precision ocean subsurface temperature fields are critical for studies on ocean dynamics and climate change. While satellite remote sensing provides extensive data on sea surface temperature, it lacks direct observations of ocean subsurface temperatures. In-situ measurements typically result in sparse and uneven spatiotemporal data coverage. Consequently, obtaining accurate high-resolution ocean subsurface temperature data remains a significant challenge in marine science. This study introduces a novel ocean subsurface temperature reconstruction model based on Spatiotemporal Graph Attention Networks (STGAT). STGAT is capable of inferring subsurface temperature fields from multiple satellite-derived sea surface observations, including Sea Level Anomaly (SLA), Sea Surface Temperature (SST), Sea Surface Salinity (SSS) and Sea Surface Wind (SSW), as well as the derived Depth-Specific Temperature Anomaly Gradient (DSTAG). The Kuroshio Extension region in the Northwest Pacific, characterized by intense ocean dynamic activity, was selected as the experimental area. Ocean temperature fields at 21 depth levels (20–1941 m) for 2022 were reconstructed using satellite sea surface observations from 2017 to 2022. The accuracy of the reconstructed subsurface temperature fields was evaluated using the GLORYS12V1 reanalysis datasets and the EN4 in-situ observation datasets. Comparative analysis with reanalysis data demonstrates that the proposed model effectively captures the spatiotemporal characteristics of ocean temperature across all depth levels, accurately reflecting realistic spatial distributions throughout the water column. The reconstructed subsurface temperature fields achieved mean RMSE and R 2 values of 0.916 °C and 0.866, respectively. Validation against EN4 data further confirms that the model’s ability to reproduce vertical thermal variations, with mean RMSE and R 2 values of 0.898 °C and 0.976, respectively. When compared with other machine learning-derived models, the STGAT model demonstrates superior performance in terms of reconstruction accuracy. In conclusion, the STGAT model effectively addresses the challenges posed by complex and highly turbulent oceanic processes, offering a promising new approach for retrieving high-resolution and high-accuracy ocean subsurface temperature data.