Abstract: Driven by continuous global warming, large-scale ocean deoxygenation/hypoxia has emerged as one of the key global scientific issues of this century. Conventional monitoring of marine dissolved oxygen (DO) relies on in-situ sampling, which is constrained by high costs, long time consumption, and challenges in achieving large-scale observations. Although there have been studies on the reconstruction of marine dissolved oxygen data through traditional machine learning methods, there remains a significant deficiency in deep learning-based forecasting research targeting the spatio-temporal evolution patterns of dissolved oxygen, as existing models struggle to effectively capture the spatio-temporal dynamic associations and heterogeneity of the marine environment. Focusing on the South China Sea, this study employs the Earthformer architecture, a model specifically designed for geoscience problems. Based on multi-source remote sensing and reanalysis data from 1998 to 2022, a comprehensive dataset was constructed, including sea surface temperature, salinity, chlorophyll-a concentration, currents, and dissolved oxygen. Using the aforementioned dataset, this study systematically trained and fine-tuned the Earthformer model, achieving the objective of forecasting the future (12 months) spatio-temporal distribution of marine dissolved oxygen from historical data. The experimental results indicate that the optimized Earthformer model possesses excellent predictive performance for dissolved oxygen. On the test set, the model demonstrated outstanding performance across multiple evaluation metrics, achieving a structural similarity index measure (SSIM) of approximately 0.78, while the un-normalized root mean square error (RMSE) and mean absolute error (MAE) were both maintained below 3μmol·kg⁻¹. Furthermore, ablation experiments revealed the impact of key hyperparameters on the model's prediction performance. This study, for the first time, validates the applicability and superiority of the Earthformer model in the long-term spatio-temporal prediction of marine dissolved oxygen (a typical complex marine biogeochemical variable), addressing the insufficiency of deep learning model applications in this research area. It provides new technical methods and theoretical support for data-driven marine hypoxia/anoxia environmental forecasting, thus possessing both significant scientific value and practical application potential.

Key words: Deep Learning, Earthformer, Dissolved Oxygen Prediction, Remote Sensing, South China Sea