A Physically Driven Interpretable Machine Learning Framework for Early Forecasting of Summer Hypoxia in the Semi-Enclosed Bohai Sea Using Remote Sensing Data

Hypoxia has become increasingly frequent in the semi-enclosed Bohai Sea since the early 2000s, posing significant risks to marine ecosystems. To address the limitations of existing dissolved oxygen models—particularly their poor predictive ability and lack of interpretability—we developed a two-month lead probabilistic forecasting framework for summer hypoxia using only multi-source remote sensing and reanalysis data, supplemented by in situ observations for validation. Environmental conditions in June were used to predict hypoxia probability in August via machine learning; among the seven algorithms tested, the optimized Random Forest model achieved the best performance (F1 = 0.76 and AUC = 0.92 on the independent test set). The model successfully reproduced observed hypoxia patterns in 2019 (validated against numerical simulations) and 2022 (validated against field measurements), capturing an increase in hypoxic area from 8229 km2 to 13,866 km2, which is consistent with intensifying thermal stratification under climate warming. SHAP-based interpretability analysis identified reduced wind speed and enhanced thermal stratification as the dominant physical drivers, highlighting the critical role of suppressed vertical mixing in limiting bottom-water oxygen supply. This study demonstrates that a physics-informed, interpretable machine learning approach based solely on satellite and reanalysis data can deliver reliable, early, and physically consistent hypoxia forecasts, offering a scalable solution for environmental monitoring of data-limited coastal seas.