Oxygen is crucial to Earth's health, directly participating in biogeochemical cycles. However, declining oceanic oxygen levels, primarily driven by climate change, pose significant environmental challenges. Despite technological advancements, historical dissolved oxygen (DO) data remain sparse, limiting comprehensive analysis. Machine learning methods offer the potential to reconstruct DO data, effectively capturing complex relationships. In our study, we employed interpretable machine learning to model DO and related environmental variables, identifying the most influential factors affecting DO predictions across various depths. We integrated dissolved oxygen data from OSD and CTD within WOD, as well as from ARGO, along with 19 environmental factors derived from reanalysis and model data. We developed a high-performing random forest regression model (R²=0.9761). Our findings indicate that temperature is the most critical factor influencing DO predictions at depths of 0-300m, while Dissolved Inorganic Phosphorus (DIP) and Dissolved Inorganic Carbon (DIC) become significant at greater depths. Utilizing this model, combined with spatiotemporal information and environmental variable data, we reconstructed a global DO dataset with 1°×1° resolution from the surface to nearly 6 kilometers deep, covering the period from 1960 to 2023 on a monthly basis. Our study validates the feasibility of using machine learning to reconstruct DO data over extended periods with high temporal and spatial resolution. Further analysis reveals a declining global DO trend across various depths, highlighting the increasing severity of ocean deoxygenation. This study provides a valuable resource for research on ocean deoxygenation and related fields.

dissolve oxygen,Machine learning model,Random Forest