GEOXYGEN: a global long-term dissolved oxygen dataset based on biogeochemistry-aware machine learning framework and multi-source observations

Abstract

Dissolved oxygen (DO) serves as an essential indicator of marine ecosystem health. However, sparse and uneven observations have limited our ability to characterize its full spatiotemporal variability, underscoring the continued need for long-term, high-resolution, and physically consistent global DO datasets. Here, we present GEOXYGEN, a global dataset of monthly DO fields at 0.5° × 0.5° resolution spanning 1960–2024 and depths from the surface to 5500 m (Wang et al., 2026a, https://doi.org/10.5281/zenodo.19703198; Wang et al., 2026b, https://doi.org/10.12157/IOCAS.20260223.002). GEOXYGEN is generated with a hierarchical modeling framework that accounts for regional and vertical heterogeneity. By combining physical and biogeochemical predictors with an adaptive feature selection strategy, GEOXYGEN demonstrated high predictive accuracy (R2 > 0.9) in independent temporal tests. The reconstructed spatial patterns align closely with the World Ocean Atlas 2023 climatology, and in subsurface waters, GEOXYGEN demonstrates superior generalization relative to existing data-driven products. Uncertainty analysis shows that the uncertainty in nearshore and shelf regions is approximately twice that in the open ocean, while long-term deoxygenation trends remain stable even without satellite-era sea-surface predictors. Additionally, a ship-only analysis of the Southern Ocean indicates that early reconstructions are robust, unaffected by the inclusion of Argo observations. GEOXYGEN offers a consistent, physically informed baseline for investigating global and regional DO variability, providing an important tool for evaluating the representation of DO in climate and Earth system models.