To reconstruct a long-term, global ocean DO dataset, we compiled several complementary in situ data products. These include CLIVAR and the Carbon Hydrographic Data Office (CCHDO), GLODAP, the GEOTRACES Intermediate Data Product (IDP2021), the OceanSITES mooring network, and the internally consistent OSD/CTD and Argo DO dataset of Gouretski et al. (2024). The OSD/CTD and Argo profiles are merged under a single automated QC framework, and Argo DO biases are evaluated and corrected using contemporaneous reference measurements. This procedure improves cross-platform consistency across platforms and helps reduce platform-related systematic differences that can affect later modeling.

A rigorous dual-stage QC protocol ensured observational reliability. The primary stage involved standardizing metadata formats and units across disparate sources, retaining only observations flagged as “good” or “probably good”. Spurious terrestrial signals were omitted via land-masking, and duplicate profiles – defined by coincidence criteria of ≤ 1 km spatial distance and ≤ 24 h temporal difference – were identified across archives. In cases of redundancy, we prioritized profiles with the highest vertical sampling density.

Standardization was further achieved by mapping observations to 176 vertical levels, adopting an expanding grid resolution (10 m intervals above 800 m, 20 to 2000 m, and 100 m thereafter down to 5500 m). This configuration aligns with the vertical frameworks of CORA and ISAS17 (Szekely et al., 2019; Kolodziejczyk et al., 2023). For the original profiles, DO, temperature, and salinity were mapped using a shape-preserving piecewise cubic Hermite interpolator (PCHIP) strictly within the observed pressure range, precluding vertical extrapolation. Profiles from Gouretski et al. (2024) were utilized directly without further interpolation as they were pre-aligned to these standard levels.

To avoid adding values in large vertical gaps where observations are sparse while retaining only levels that are locally supported by sufficient observations, we applied an “observation constraint” filter to the interpolated standard-level data. For each standard level z, we first used a strict window xz: the level is kept if at least one observed pressure exists within ±xz. For levels that fail the strict window, we then applied a relaxed window yz, but required at least two observed pressures within ±yz. The windows vary with depth z, as described in Eq. (1). 1xz=2,z≤505,50<z≤80010,800<z≤200020,z>2000m,yz=5,z≤5015,50<z≤80030,800<z≤2000120,z>2000m, We also recorded an interpolation uncertainty term, σinterp, to represent the local error scale introduced by vertical interpolation, as defined in Eq. (2). Using a first-order approximation, this uncertainty depends on the local vertical DO gradient and the distance to the nearest valid observed pressure. We estimated the local gradient ∂O2/∂pz from adjacent observed points along the profile, defined Δpz as the pressure distance between the standard level and the nearest valid observed point. 2σinterpz=∂O2∂p‾z⋅Δpz A second QC stage was applied to the standard levels. First, we excluded records where DO exceeded the 0–600 µmol kg⁻¹ range. Next, to limit uncertainty from vertical interpolation, data with σinterp>3 µmol kg⁻¹ were removed, based on the 95th percentile of σinterp across all standard-level samples. Finally, following TEOS-10, oxygen saturation percentage (Sat %) was computed using in-situ temperature and salinity data. For standard levels deeper than 200 m, records with Sat % ≥ 120 % were considered erroneous and removed.

After two-stage QC, we assembled a high-quality archive of 930 252 DO profiles, of which 94 % are derived from the OSD, CTD, and Argo platforms. Figure 1a captures the decisive transition in the historical data record: a long-standing reliance on ship-based OSD and CTD profiles was superseded post-2000 by the exponential growth of the Argo network, which now serves as the backbone of global oxygen monitoring. Platform-specific dominance is also depth-dependent, defining the vertical representativeness of the dataset. Above 800 m, the combined prevalence of OSD, CTD, and Argo remains robustly above 90 %. However, the intermediate depths (800–1800 m) mark a transitional zone where Argo's increasing share compensates for the thinning OSD/CTD coverage. Beyond 2000 m, where Argo penetration is limited, the deep-ocean constraint returns to the OSD and CTD platforms, which maintain a prevalence of at least 80 %. This tiered vertical structure ensures that secondary data products remain supplementary to the core observational framework.

The geometric evolution of sampling has shifted from ship-track-oriented linear clusters to a near-planar global distribution. This shift is most evident in the Southern Hemisphere, where coverage increased from extreme sparsity prior to 1980 to approximately 36 % of the global total in the 2000–2020 interval. This globalization of the data stream, driven by the Argo array, has mitigated historical hemispheric imbalances, although systematic gaps persist in marginal seas, the central Pacific, and ice-influenced Arctic regions.

Acknowledging that uneven observational density poses a risk of artificial signal dominance from high-coverage eras, our methodology incorporates specific safeguards to ensure trend stability. By adopting heterogeneity-based partitioned modeling and decadal-scale transferability assessments, we reduce the risk of data-sparse regions being dominated by modern sampling patterns. These efforts are further supported by an inverse-density weighting scheme and the use of grid-representative values, which together enhance the influence of sparse historical records and maintain the physical interpretability of multi-decadal oxygen trends.

Accurate reconstruction of ocean deoxygenation needs both long, dense DO observations and drivers that describe physical transport and biogeochemical processes across scales (Oschlies et al., 2018). Consequently, we leverage three-dimensional thermodynamic fields (temperature and salinity) alongside a curated suite of sea surface environmental variables (SSEVs) to establish a process-based predictor space (Table 1).

Three-dimensional temperature and salinity fields are taken from the Coriolis Ocean Dataset for Reanalysis, CORA (Szekely et al., 2025). CORA is a CMEMS objective analysis that uses the ISAS objective mapping system to merge in situ temperature and salinity observations from ships, Argo floats, and other platforms. It also applies delayed mode quality control to support long-term stability and global consistency. From these fields, oxygen saturation (O₂Sat) is derived according to the TEOS-10 standard (IOC et al., 2015). The resulting discrepancy between observed DO and O₂Sat serves as a critical proxy for the integrated influence of microbial respiration and physical ventilation dynamics.

Complementarily, we assembled a multidimensional suite of SSEVs, encompassing thermodynamic, dynamical, bio-optical, and carbon-chemistry features (Shao et al., 2024; Ma et al., 2025). Wind-vector components (U and V) are taken from NASA's Cross-Calibrated Multi-Platform (CCMP) product (Mears et al., 2022). Mixed-layer depth (MLD) is obtained from the CMEMS Multi-Observation Global Ocean 3D product (Guinehut et al., 2012). Dynamical variables include sea surface height (SSH) and eddy kinetic energy (EKE), both derived from AVISO satellite altimetry (Hauser et al., 2020). Bio-optical variables comprise photosynthetically active radiation (PAR) and chlorophyll a (Chl-a) from NASA Level-3/Level-4 ocean-color products (NASA Ocean Biology Processing Group, 2018). Carbon-chemistry variables include dissolved inorganic carbon (DIC), total alkalinity, pH, sea surface partial pressure of CO₂ (pCO₂), and CO₂ flux, all obtained from the CMEMS Surface Ocean Carbon Fields product (Chau et al., 2022, 2024). All feature variables, last accessed in March 2025, were standardized onto a uniform monthly 0.5° grid to maintain spatial consistency across the reconstruction.

To match the spatiotemporal scale of the supervised learning labels with the target reconstruction grid and to reduce sample-weight bias caused by uneven sampling in space and time, all observations are aggregated to a monthly 0.5° by 0.5° grid. Within each grid cell, observations are summarized into a single representative value, defined as the median to limit the influence of outliers and extreme events on the labels. The within-unit dispersion is also computed using the median absolute deviation (MAD), which serves as an empirical proxy for uncertainty.

To quantify representativeness error introduced by finite sampling and sub-grid variability within each cell-month-depth unit, within-unit dispersion is used to estimate σrep. Let a unit contain nobs observations {xi}i=1nobs. The representative value is defined as the robust central statistic, as shown in Eq. (3). We then compute the dispersion and define it as the representativeness error metric, σrep, as described in Eq. (4), where the factor 1.4826 makes σrep comparable to the standard deviation under an approximate normal distribution, facilitating comparison across units. 3x̃=medianxi4MAD=medianxi-x̃,σrep=1.4826×MAD At the same spatiotemporal scale, all environmental variables and DO data are mapped in space and time. To preserve historical information, all QC-passed DO data are retained even when some covariates such as SSEVs are missing. This sample archive provides standardized inputs for later model training.

Our reconstruction framework utilizes a spatiotemporally stratified approach to address the shifting controlling mechanisms of ocean deoxygenation across basins and depths (Ma et al., 2025; Ito et al., 2024a). Following the basin definitions in the World Ocean Atlas 2023 (WOA23; Garcia et al., 2024), we additionally treat the Southern Ocean as a dedicated domain. Accordingly, the horizontal grid is divided into five primary modeling domains (Atlantic, Pacific, Indian, Southern, and Arctic), which are held constant across vertical layers to maintain training coherence (Fig. 3). By avoiding highly intricate province boundaries, this design reduces sensitivity of variability and trend estimates to boundary effects and makes cross-boundary continuity easier to maintain. We further mask the Mediterranean, Red Sea, and other semi-enclosed marginal seas to focus the reconstruction on open-ocean dynamics dominated by large-scale circulation and transport.

The vertical architecture utilizes 176 standard levels, as discussed in Sect. 2.1, similar to the CORA/ISAS17 frameworks (Szekely et al., 2019; Kolodziejczyk et al., 2023). At each standard depth below 3000 m, the framework converges into a single global model, a necessary adaptation to accommodate the physical constraints of the seafloor and data limitations. For each basin and depth layer, separate models are trained based on the partitioned regions.

Within each depth, the self-organizing-map-derived descriptor (SOM_DES) is used as a categorical predictor to encode the large-scale climatological background state. Specifically, we construct a four-dimensional climatological vector [SST, SSS, MLD, DO] from monthly background fields. These vectors are then used to train a self-organizing map (SOM), and each grid cell is assigned to one of 25 discrete classes. The clustering is based on the joint patterns of multiple environmental fields, and, at this step, the O₂ climatology is implicitly assigned a higher weight to obtain a seasonally varying, dynamic partitioning. The SOM is trained directly on these monthly climatological fields, such that the DO climatology implicitly exerts a stronger influence on the resulting state partition and ensures that the partitioning evolves coherently with the seasonal DO cycle. This feature-engineering step leverages the SOM's ability to map multivariate climatological structure onto a discrete set of regimes while preserving topological dependencies, thereby providing clear seasonal and spatial context for month-scale DO reconstruction. The background climatological fields are anchored in WOA23 (upper 1500 m) and the IAP Global Ocean Oxygen gridded product (IAP Oxygen; Cheng and Gouretski, 2024), ensuring vertically and horizontally comprehensive baseline states. Overall, SOM_DES captures the large-scale climatological structure of DO and supplies background-state information that supports refined month-scale DO modeling, improving diagnostic consistency across heterogeneous regimes and reinforcing the physical interpretability of the global DO product.

We train independent sub-models for each ocean basin partition across 176 standard depth levels, employing the CatBoost gradient boosting framework to resolve the non-linear mapping between sparse DO observations and diverse environmental predictors. CatBoost's implementation of oblivious decision trees and ordered boosting is strategically utilized to mitigate variance while precluding target leakage – a critical requirement for stable multi-decadal reconstruction. This framework is algorithmically predisposed to this task due to its innate capacity to process missing predictors without imputation, its support for cost-sensitive sample weighting, and its robust regularization suite (L2 leaf regularization and early stopping) that safeguards against overfitting in data-sparse regimes.

During the period before the satellite era (for example, from the 1960s to the 1980s), this model would automatically identify and properly handle the missing SSEVs. This non-imputational strategy allows the reconstruction to hinge predominantly upon thermodynamic variables (Temperature, Salinity, O₂Sat) and spatiotemporal coordinates without introducing the systematic biases inherent in statistical filling. We tested longitude as a candidate predictor; however, it induced spurious banded (stripe-like) artifacts in data-sparse regions and was therefore excluded from the final predictor set. The CatBoost model uses 20 predictor variables as inputs: Year, month_sin, month_cos, Latitude, Temperature, Salinity, O₂Sat, SOM_DES, U, V, SSH, EKE, MLD, PAR, Chl-a, DIC, pCO₂, pH, Alkalinity, and CO₂ flux. The month_sin and month_cos encodings are defined in Eqs. (5)–(6). 5month_sin=sin2π(m-1)126month_cos=cos2π(m-1)12m∈{1,2,…,12}, To improve interpretability and generalization, a two-stage feature selection procedure is applied to each region–depth submodel. Only training data are used, and cross-validation follows the same decadal-scale design used for model evaluation. Tree-based models often benefit from compact feature sets because redundant predictors can reduce predictive skill (Garabaghi et al., 2023). First, we estimated permutation importance under five-fold cross-validation grouped by decade and retained an initial subset of features using an adaptive rule K=max⁡(10,2p), where p=20 is the number of candidate features, discarding predictors with negligible contribution. Second, we perform recursive feature elimination with cross-validation (RFECV), iteratively removing the least important feature and selecting the combination that minimizes validation Root Mean Square Error (RMSE). This process facilitates a physical evolution of the feature space: surface models prioritize high-frequency biogeochemical forcing, mid-depth models emphasize water-mass transition markers, and deep-ocean models converge on conservative thermodynamic tracers.

To mitigate biases arising from heterogeneous spatiotemporal sampling (Fig. 1), we applied inverse-density weighting within a fixed binning scheme. The sample domain is partitioned on a 5° × 5° latitude–longitude grid and non-overlapping 10-year time windows in each cell. Sample weights were computed as the inverse of the observational density within each spatiotemporal bin and standardized at each depth level, thereby reducing the influence of over-sampled regions and periods during model training.

Let bi denote the spatiotemporal bin containing sample i, and let nbi be the number of samples in that bin. The initial per-sample weight is the inverse square root of this count, as defined in Eq. (7). 7w̃i=1nbi, Specifically, each sample weight is set proportional to the inverse square root of the local sample density. To preserve the aggregate information content, we then normalize the weights to unit mean, as defined in Eq. (8). 8wi=w̃i1N∑k=1Nw̃k, This strategy increases the influence of observations from sparse regions and earlier periods without altering the aggregate sample distribution.

Independent hyperparameter optimization for each basin-depth unit is performed using Bayesian inference (Optuna), targeting the objective of validation RMSE minimization (Table 2). This automated search is integrated with a decadal-block five-fold cross-validation scheme to address the challenges of non-stationary ocean signals. By grouping observations into multi-year blocks, we decouple validation results from the short-term temporal dependencies that often inflate predictive skill in traditional random-split CV (Salazar et al., 2022). This structural separation ensures that the model learns large-scale climatic drivers rather than localized temporal artifacts. In each cross-validation fold, early stopping is activated if validation RMSE fails to improve for 50 consecutive iterations, after which the optimal iteration count is recorded. To secure a robust final model, we set the terminal iteration count to the median of these recorded values across all folds. This iteration-locked retraining on the complete calibration set – with early stopping disabled – prevents overfitting and ensures a stable convergence state.

For an unbiased final assessment, a strictly independent global test set is constructed via decade-stratified random sampling (1960–2024), where one full calendar year per decade is entirely excluded from feature selection and hyperparameter tuning. The resultant test years (1961, 1970, 1984, 1993, 2003, 2012, and 2020) provide a temporally representative benchmark. This withheld-year test set is reserved exclusively for final performance assessment and inter-product benchmarking; all model selection and calibration are conducted via decade-block cross-validation within the remaining training data. Integrated predictions from regional sub-models are then evaluated against this set using RMSE, mean absolute error (MAE), and the coefficient of determination (R2).

The relationship between DO and its environmental drivers undergoes a transformation from the surface to the deep ocean (Fig. 4). In the near-surface layers, thermodynamic solubility constraints dominate, manifesting as a robust negative correlation with temperature. As depth increases, this direct solubility control attenuates, allowing salinity to emerge as the primary integrative proxy for water-mass properties and ventilation history, particularly tracking the signatures of deep-water formation and transport. These depth-dependent patterns provide the empirical foundation for resolving complex non-linear interactions within the water column (Ping et al., 2024; Cao et al., 2024).

Correlations between sea surface environmental variables (SSEVs) and subsurface DO represent teleconnected pathways rather than local drivers. Surface dynamical forcing and upper-ocean mixing modulate water-mass formation and reventilation, thereby constraining the DO budget at depth. Specifically, sea surface height (SSH) exhibits persistent vertical coherence across much of the water column, while wind stress provides a dynamical context for advection and upwelling (Hollitzer et al., 2024). In contrast, the Chl-a signal remains confined to the epipelagic zone, reflecting its role as a productivity indicator. Beyond the euphotic zone, carbonate system variables, notably alkalinity, maintain high correlations with DO, delineating the remineralization background linked to deep-water aging.

Feature-importance diagnostics are reported to describe model dependence and do not imply causality. The adaptive feature-selection results show that the model's reliance on predictors varies strongly with depth and region. In the upper 10 m, temperature and O₂ saturation are consistently among the most informative predictors, whereas at intermediate depths (1000–2000 m) salinity tends to contribute more strongly in certain high-latitude regimes, including the Arctic and Southern Oceans (Fig. 5). This depth-dependent pattern motivates the use of depth-specific predictor sets to better represent distinct hydrographic contexts. In addition, several regionally relevant covariates (e.g., SSH in the Indian Ocean and DIC/alkalinity in low-oxygen environments) are retained more frequently by the selection procedure, indicating that they provide useful contextual information for prediction under specific regimes (Franco et al., 2014).

Although latitude serves as a dominant proxy for broad-scale gradients (Milà et al., 2024), specialized SSEVs are vital for refining local accuracy. Our regionalized architecture prioritizes these idiosyncratic dynamics, utilizing variables like SOM_DES to resolve high-frequency seasonal variations in the Southern Ocean. This adaptive approach mitigates the biases inherent in spatially stationary parameterizations, ensuring the reconstruction respects the intrinsic heterogeneity of the global oxygen cycle.

Vertical stratification of error profiles reveals a coherent link between RMSE and the intensity of the oxycline (Fig. 6). In most basins, uncertainty is concentrated in the upper 600 m, where biological consumption and physical mixing generate high spatiotemporal variance. The Indian Ocean (IND) presents the most significant upper-ocean RMSE peak, likely due to its unique biogeochemical configuration, whereas the Arctic Ocean (ARC) shows fluctuations in RMSE at intermediate depths, but its R2 remains relatively high. This vertical decoupling of error – where water-mass stability increases with depth before decreasing – represents a common feature across the Atlantic, Pacific, IND, and Southern Oceans, further confirming the appropriateness of the regional partitioning approach.

Except for the sparsely sampled Arctic, all ocean basins maintain high R2 values (typically > 0.8) across the sampled water column. While absolute errors are inherently higher in the upper ocean where DO variability is maximized, the high R2 values indicate that the underlying spatial and temporal patterns are accurately recovered. The analysis demonstrates that uncertainty is largely a function of the vertical DO structure, with the deep-ocean regime providing a stable and highly predictable anchor for long-term deoxygenation trends.

The reconstructed DO fields demonstrate high accuracy throughout the global ocean, with scatter plots confirming that estimates follow the 1 : 1 line across the entire concentration range (Fig. 7). At representative depths, the global RMSE/MAE on the independent test set is on the order of 12.8/8.0 µmol kg⁻¹ at 10 m, 16.8/11.4 µmol kg⁻¹ at 200 m, and 7.5/5.2 µmol kg⁻¹ at 1000 m. Most R2 values are above 0.9, and the R2 for deep-layer DO reconstruction reaches 0.99. A slight overestimation in regions with relatively low values reflects the inherent smoothing characteristic of regularized decision trees, which tend to pull extreme values toward local means. However, the symmetric distribution of residuals around zero indicates that these effects are localized and do not introduce a systematic large-scale bias. This performance confirms that the modeling architecture is well-suited for resolving the non-linearities inherent in ocean oxygen dynamics.

The trained models leverage CORA ocean analysis from CMEMS to generate spatially and temporally consistent monthly dissolved oxygen (DO) reconstructions. By using gridded temperature and salinity products, the framework obviates interpolation from sparse in situ profiles and minimizes associated uncertainties. These predictors undergo delayed-mode objective mapping and rigorous quality control to ensure long-term consistency and full traceability. Monthly DO predictions are generated at 176 standard depth levels on a 0.5° × 0.5° grid extending to 5500 m.

To mitigate discontinuity artifacts at basin province boundaries – described as the “step-effect” – a boundary fusion protocol is implemented within transition zones (Wagstaff and Bean, 2022). For a prediction location x, the fused estimate is derived from the regional model prediction y^ix and the minimum great-circle distance di to the provincial boundary, as defined in Eq. (9). With a smoothing bandwidth of S=300 km, predictions from adjacent basins are smoothly blended in proportion to their distance from the edge. This approach preserves continuity across boundaries without excessive smoothing. Far from boundaries (d>S), only the local provincial model contributes to the estimate. 9y^′x=∑iwdi∣Sy^ix∑iwdi∣S,wdi∣S=S-diS2,di≤S0,di>S, Finally, we obtained GEOXYGEN, a global DO dataset that provides monthly fields on a 0.5° × 0.5° latitude–longitude grid at 176 standard depth levels from 1 to 5500 m, spanning 1960–2024. The latitude grid is not strictly uniform, with higher density in the high-latitude regions. Distributed via CF-compliant NetCDF files, each monthly volume (GEOXYGEN_DO_YYYYMM_0p5deg_v1.nc) includes the four-dimensional DO variable (time × depth × lat × lon) and associated coordinates. Time represents monthly means encoded as days since 1 January 1950 00:00:00 UTC (Gregorian calendar), with missing data identified by a large sentinel value. A separate NetCDF file provides the basin provinces and valid-ocean masks. The complete dataset is available at 10.12157/IOCAS.20260223.002 (Wang et al., 2026b).

To evaluate the predictive fidelity of GEOXYGEN against four existing dissolved oxygen (DO) products (Table 3), we utilize the independent test dataset stratified across discrete temporal windows: the early period (1960–1980) and the recent epoch (2000–2020; Table 4). The intercomparison is strictly constrained to spatiotemporally co-located samples, ensuring that only grid-month-depth intersections with concurrent availability across all products are included in the evaluation. This rigorous filtering preserves statistical equity and minimizes biases arising from disparate product masks.

The comparative analysis reveals three dominant patterns. First, all products show larger errors during the early period, a trend predominantly attributable to the historical sampling architecture. Pre-Argo observations were characterized by a disproportionate concentration of coastal and shelf cruises, whereas the modern era is defined by the proliferation of autonomous Argo floats providing expansive open-ocean coverage. Because coastal environments are governed by pronounced small-scale heterogeneity and high-frequency nonstationarity, their reconstruction poses a greater challenge, inherently raising the error floor for the 1960–1980 interval.

Within this data-sparse window (1960–1980), GEOXYGEN manifests superior predictive stability across the upper and intermediate layers (10–500 m). It consistently achieves lower RMSE and MAE than both IAP and ML4O2 across nearly all depth levels while maintaining robust R2 values. These gains demonstrate that the stratified, basin-specific modeling framework facilitates the recovery of mesoscale spatial heterogeneities and vertical gradients even under sparse observational constraints. By leveraging high-resolution target fields, this architecture enhances the representation of regional variability, ensuring that GEOXYGEN maintains competitive skill during the pre-satellite and pre-Argo eras.

In the data-rich modern period (2000–2020), GEOXYGEN exhibits even more pronounced performance gains. Within the critical mesopelagic range (100, 200 and 500 m), the product shows a significant accuracy improvement relative to all reference datasets, while ensuring that residual biases are smaller and more stable. In contrast, several existing products exhibit persistent systematic biases in this depth range, potentially compromising the representation of OMZs and steep vertical gradients. This suite of high-resolution, broad-coverage DO fields provides a reliable foundation for the integrated assessment of global deoxygenation trends and their underlying physical-biogeochemical drivers.

To evaluate the large-scale credibility and multi-year stability of our reconstructed dissolved oxygen (DO) fields, we compare our product's annual-mean DO climatology with WOA23 and examine vertical structure using representative profiles from three major basins (Fig. 8).

In the upper ocean (0–300 m, depth-averaged), both products capture consistent basin-scale spatial patterns. The subtropical gyres exhibit relatively high DO concentrations, while the equatorial region and eastern boundary upwelling systems form distinct oxygen-deficient belts. The spatial extent and location of these structures are in close agreement between the two climatologies. GEOXYGEN reproduces these banded structures continuously, with cross-frontal gradients in transition zones closely matching those in WOA23.

In the vertical, our product accurately depicts the largest hypoxic zone in climatology, which is mainly located in the intermediate depth, ranging from 300 to 1300 m. Meridional sections reveal that the model successfully reproduces the core depths and spatial extents of major hypoxic zones across the three primary ocean basins. The precise alignment of oxygen isopleths and the accurate representation of the transition layer steepness indicate that the reconstruction results are consistent with the climatological constraints of DO.

Unlike the standard 1° grid used in most global products, GEOXYGEN employs a higher 0.5° horizontal resolution to capture more detailed features. By training on localized relationships between DO and hydrographic analyses within discrete basin provinces, the model resolves the morphometry of OMZs with enhanced clarity. Although the reconstruction fidelity remains conditioned upon the quality of input observations, benchmarking against 1° products reveals that the 0.5° configuration yields systematically lower errors in high-gradient regimes. This selection constitutes an optimized operating resolution that maximizes information recovery from the underlying biogeochemical constraints without over-interpreting sparsely sampled intervals.

To quantify the credibility of the GEOXYGEN product, we decompose total uncertainty into two components: observation-related uncertainty and mapping uncertainty. Observation-related uncertainty summarizes measurement error and the representativeness error introduced when observations are aggregated to the cell–month–depth scale. Mapping uncertainty describes prediction error from the machine-learning mapping between environmental predictors and DO. This decomposition separates label-side and model-side contributions. It also supports diagnosing elevated uncertainty in coastal regions and potential bias in earlier, data-sparse periods, and it improves traceability and clarity for product use.

Measurement error depends on the measurement technique and instrument. For Winkler titration, bottle measurements can include random errors on the order of 1 µmol kg⁻¹ or smaller (Carpenter, 1965). Since observing platforms differ in sensor type, calibration strategy, and QC level, we assign a platform-specific measurement error scale σmeas to each data source to represent differences in precision among platforms. Bottle-based sources including CCHDO Bottle, GLODAP, and GEOTRACES IDP are assigned 1 µmol kg⁻¹. Since Argo data are bias corrected, Argo, OSD and CTD, and CCHDO CTD are assigned 1.5 µmol kg⁻¹. OceanSITES is assigned 2.0 µmol kg⁻¹.

When forming supervised-learning labels at the cell–month–depth scale, observation-related uncertainty is defined as the combined contribution of measurement error, vertical mapping uncertainty, and representativeness error, as defined in Eq. (10). 10Uobs=σmeas2+σinterp2+σrep2 Here, σinterp is the vertical mapping uncertainty introduced when a profile is mapped to standard levels, and σrep is the representativeness error estimated from within unit dispersion. Their definitions and computation have been described earlier.

Mapping uncertainty characterizes error generated during the mapping from environmental fields to DO. It reflects the combined effects of model structure, representativeness of training samples, and nonstationarity across regions and depth levels. On the independent test set, we define the residual following Eq. (11): 11r=y^-y where y^ is the model prediction and y is the observed label for the corresponding test sample. Following Ito et al. (2024a), we estimate the mapping uncertainty at each grid cell from the test residuals by taking their second central moment, as defined in Eq. (12). 12σmap=r‾2-r‾2, Here, the overbar denotes the mean over all test samples within the same grid cell. This definition estimates the residual variance at the grid cell scale. It separates the systematic component r‾ from the random error component captured by σmap.

Based on the two components above, total uncertainty (Utotal) is defined in Eq. (13). 13Utotal=Uobs2+σmap2

As shown in Fig. 9, the spatial distribution of uncertainty in GEOXYGEN reveals a distinct bifurcation between observation-related uncertainty and mapping residuals. Observational uncertainty manifests pronounced spatial stationarity across pelagic domains. Conversely, mapping uncertainty exhibits a structured regional geometry, with elevated error bands aligning with high-gradient kinetic regimes such as western boundary currents and upwelling systems. The highest uncertainty is observed in the Pacific (7.385 µmol kg⁻¹), followed by the Atlantic (6.148 µmol kg⁻¹), Arctic (4.439 µmol kg⁻¹), Indian (4.084 µmol kg⁻¹), and Southern Ocean (3.652 µmol kg⁻¹). The higher uncertainty in the Pacific and Atlantic Oceans is primarily due to the structural complexity and dynamic intensity of their oceanographic systems, as well as the coastal distribution of early observational data in these basins. The global-mean total uncertainty of 6.054 µmol kg⁻¹ conceals a pronounced shallow-water divergence: in nearshore and shelf regions where bathymetric depth is shallower than ∼ 200 m, total uncertainty rises to 12.917 µmol kg⁻¹ – more than double the open-ocean baseline (5.970 µmol kg⁻¹). Consistent with the bathymetry-binned diagnostic, uncertainty increases monotonically with decreasing bathymetric depth, indicating progressively reduced predictability toward the shallow, dynamically heterogeneous coastal ocean (Fig. 9d). This intensification is predominantly driven by localized, high-frequency processes – including phytoplankton pulses, riverine influx, and tidal oscillation – which generate non-linear spatial gradients that challenge the transferability of pelagic-trained feature relationships (Gilbert et al., 2010; Regier et al., 2023; Giomi et al., 2023; Liu et al., 2024). These localized dynamics induce steep spatial gradients and temporal non-stationarity, which subsequently reduce the regional transferability of learned DO-environment associations (Valera et al., 2020). For nearshore and semi-enclosed bay environments, we recommend using GEOXYGEN with caution and interpreting results at larger spatial aggregation to reduce sensitivity to local high-frequency variability.

The vertical stratification of uncertainty reflects the underlying hydrographic stability and process coupling within the water column. As depicted in Fig. 9e, the total uncertainty profile reaches its maximum in the epipelagic layer before undergoing monotonic attenuation toward the abyssal depths. This vertical variation is consistent with the change in model accuracy across depth layers. In contrast, the intermediate and deep layers provide stronger water-mass constraints and a more coherent oxygen field, facilitating a convergence of mapping uncertainty as the predictive relationship stabilizes.

Temporal trends in uncertainty serve as a diagnostic of the evolving global observing system. The progressive reduction in total uncertainty across the withheld test years (Fig. 9f) coincides with the expansion of the Argo float network, which transitioned ocean sensing from route-based ship surveys to a spatially distributed autonomous paradigm. This transition significantly improved the observational constraints in the Southern Hemisphere and remote ocean basins, effectively lowering the residual variance in model validation. While this decline highlights the structural evolution of sampling coverage, the resulting time series reflects the collective stability of the multi-decadal reconstruction rather than a year-specific local error estimate.

It should be noted that GEOXYGEN has higher uncertainty in coastal and shelf regions. Therefore, for coastal applications, we recommend using multi-grid-cell averages and focusing on monthly to seasonal or longer timescales. Use of the native 0.5° grid near the coast should therefore be approached with caution and, wherever feasible, cross-validated against local observations or higher-resolution regional products.

A two-configuration sensitivity experiment was designed to test whether satellite-era sea-surface information alters the reconstructed dissolved oxygen (DO) signal. The full predictor configuration used the complete predictor suite, while the reduced predictor configuration excluded the sea-surface predictor group (U, V, SSH, EKE, MLD, PAR, Chl-a, DIC, pCO₂, pH, alkalinity, and CO₂ flux). All other parts of the reconstruction pipeline were kept the same, so differences between the two products isolate the effect of including this predictor group on the same grid and over the same period.

The comparison focuses on deseasonalized DO anomalies averaged over the 1–100 m depth range and further smoothed using a centered 12-month moving average. This metric targets the upper ocean, where sea-surface information is most likely to influence variability, while suppressing grid-scale noise that can obscure low-frequency signals. Predictor impact is quantified by the configuration-induced component, defined as Full - Reduced, which isolates the incremental effect of the excluded predictor group in anomaly space.

The full and reduced reconstructions exhibit strong agreement over 1960–2022 (Fig. 10), yielding a consistent depiction of upper-ocean (1–100 m) low-frequency variability: a sustained positive-anomaly regime through the 1970s–1980s followed by a marked decline after ∼ 1990 and a transition into a persistently negative-anomaly regime by ∼ 2000 (relative to the monthly climatology). The Full–Reduced difference remains close to zero prior to the mid-1980s and stays small compared with the total anomaly amplitude thereafter, indicating that the global-mean, decadal-scale signal is only weakly sensitive to the inclusion of satellite-era surface predictors. As an external benchmark, ML4O2 reproduces comparable variability during ∼ 1965–2010 but shows more negative anomalies in the most recent decade, suggesting a stronger upper-ocean deoxygenation signal in anomaly space (relative to its monthly climatology) than GEOXYGEN over the same period; however, part of this divergence may arise from inter-product differences in baseline climatology, sampling, and reconstruction methodology. Overall, the close concordance between configurations supports the robustness of GEOXYGEN for decadal-scale assessments, with configuration-dependent deviations being minor relative to the dominant multi-decadal evolution.

To quantify the dependency of the early reconstruction on modern autonomous sensing, a ship-only sensitivity experiment was conducted to assess potential retrospective signal contamination from the data-dense Argo era. Within the Southern Ocean (south of 45° S), we benchmarked an “Argo-excluded” configuration – utilizing non-Argo historical records – against an “Argo-included” configuration across the 1960–2000 period. This region serves as a critical diagnostic domain due to its radical transition from sparse historical sampling to comprehensive Argo coverage over the last two decades. By maintaining identical external physical analysis constraints, the experiment isolates the impact of Argo observations on the reconstructed climatological mean and vertical structure (Fig. 11).

The results manifest pronounced morphological invariance in the upper-ocean (1–100 m) DO climatology across both configurations. The annular high-oxygen band and its relative position within the circumpolar system remain effectively stationary, implying that Argo inclusion does not induce systematic structural rearrangement. Notably, the difference map reveals localized positive ΔDO hotspots concentrated around the Antarctic margin, with enhanced amplitudes in several coastal/shelf sectors, whereas departures in the open-ocean circumpolar interior are minimal and barely discernible. Consistent with this pattern, the area-weighted mean vertical profiles are nearly overlapping across most depths (panel d), with only a subtle upper-ocean positive offset in the Argo-included configuration that remains below ∼ 0.5 µmol kg⁻¹, indicating modest amplitude refinement rather than depth-dependent reorganization. Collectively, these features suggest that the reconstructed vertical architecture is primarily constrained by consistent physical structure, while the denser modern sampling acts mainly to fine-tune regional magnitudes, thereby supporting the structural integrity of the historical reconstruction.

While basin-based modeling represents regional environmental heterogeneity, it often induces “step-effect” discontinuities at basin boundaries, resulting in unrealistic shifts in long-term trends and monthly variability at adjacent grid points. To address these boundary artifacts, we implemented a fusion method to smooth inter-basin transitions. We validated this approach by constructing boundary diagnostic samples on a 0.5° by 0.5° grid, specifically selecting adjacent points across basin interfaces.

The diagnostic experiment targeted the 100 m depth layer, a region characterized by high spatial complexity and hydrographic sensitivity. Using a global non-partitioned model as a continuous baseline, we quantified discontinuity magnitudes via the differential trend (Δtrend) and standard deviation (ΔSD) at adjacent cells. Contrasting fused and unfused basin-specific outputs against this reference isolated the consistency gains.

Results reveal that the boundary fusion protocol mitigates abrupt statistical shifts, decreasing local biases in the partitioned map (Fig. 12). Improvements in Δtrend and ΔSD demonstrate that the smoothing operator suppresses interface artifacts. Nevertheless, small discrepancies relative to the global non-partitioned baseline persist, indicating that independently trained regional submodels may retain slightly different statistical mappings that are only partially reduced by simple fusion.

Boundary fusion thus serves as an effective strategy for mitigating partition-induced “step-effect” artifacts, yielding a more coherent global spatial structure. While localized gradients remain in high-contrast hydrographic regions, their reduced magnitude indicates a meaningful alleviation of boundary inconsistencies. Future enhancements could involve adaptive fusion schemes or multi-task learning to explicitly share data across regional submodels while preserving specialization.