OceanDepths: A Global Dataset of Paired Subsurface and Surface Ocean Observations

Simon Donike; Ruben Cartuyvels; Antonino Ian Ferola; Elisa Carli; Diego Fernandez Prieto; Marie-Helene Rio

Source-linked AI summary

OceanDepths: A Global Dataset of Paired Subsurface and Surface Ocean Observations

Simon Donike, Ruben Cartuyvels, Antonino Ian Ferola, Elisa Carli, Diego Fernandez Prieto, Marie-Helene Rio

arXiv:2608.16373v2Aug 17, 2026cs.LGcs.AIcs.CV

TL;DR

Ocean research lacks a standardized, high-resolution observational dataset coupling satellite surface fields with co-located subsurface profiles. OceanDepths provides global weekly paired observations and demonstrates that climatology can outperform simple machine-learning baselines for subsurface reconstruction.

Problem
Existing ocean-learning studies largely rely on reanalysis data, while observational alternatives are limited to specific basins.
Method
OceanDepths co-locates satellite SST, SSS, and SSH with EN4 subsurface profiles on a global 0.1° grid at weekly resolution from 2000–2024.
Results
Climatology often outperforms simple machine-learning baselines for subsurface ocean-state reconstruction.
Takeaways & Limitations
OceanDepths provides a unified resource for ocean-related machine-learning tasks across a challenging, sparse four-dimensional observational setting.
Takeaways & Limitations
OceanDepths inherits Argo sampling biases, with sparser coverage in marginal seas, near coasts, under ice, and below 2,000 m.

Abstract

from arXiv · show

Despite comprising over 70% of its surface, the world's oceans are critically underobserved compared to the land surface or the atmosphere. Understanding the global ocean requires jointly observing its surface and subsurface structure, yet no standardized, high-resolution dataset couples satellite surface fields to co-located in situ depth profiles in an AI-ready format. Existing resources either consist of model-reconstructed gridded products rather than observations, cover only a single variable or basin, or operate at resolutions too coarse for mesoscale dynamics. We introduce OceanDepths, the first open, global, regridded AI-ready dataset that pairs satellite-derived sea surface temperature (SST), sea surface salinity (SSS), and sea surface height (SSH) L4 products with co-located EN4 subsurface temperature and salinity profiles, complemented by matched GLORYS12 ocean reanalysis data to support comparisons or multi-stage learning. The dataset spans 2000-2024 at 0.1 degrees x 0.1 degrees spatial resolution and at weekly temporal resolution, covering the entire globe's sea surface and with over 9.5 million paired profiles interpolated to 50 standardized depth levels. We provide a configurable system to split the globe in equally sized spatial patches. The 4D multivariate structure, high resolution, long temporal extent, and extreme sparsity of subsurface observations (approximately 0.01% per depth level) make OceanDepths a challenging testbed for novel AI methods. We demonstrate subsurface state reconstruction as an example task with simple baseline models, but also envision OceanDepths to support the development of observation-based forecast methods and other related tasks. Available at: https://huggingface.co/datasets/ESA-philab/OceanDepths.

1 Introduction

OceanDepths addresses the mismatch between dense satellite surface observations and sparse subsurface profiles by providing an open, global, standardized dataset of paired observations for AI-based ocean-state reconstruction and related tasks. It combines high-resolution, weekly SST, SSS, and SSH products with co-located EN4 profiles and complementary GLORYS12 data spanning 2000–2024.

Motivation: The ocean’s three-dimensional temperature and salinity structure is essential for monitoring heat content, circulation, extreme events, and climate projections.
Observation gap: Satellite observations provide high-resolution near-global surface coverage, whereas Argo subsurface sampling remains sparse, with Core floats targeting 3°×3° spacing.Satellites observe SST, SSS, and SSH; Argo profiles temperature and salinity primarily in the upper 2000 m.
Dataset contribution: OceanDepths pairs satellite-derived SST, SSS, and SSH L4 products with co-located EN4 subsurface profiles at 0.1°×0.1° and weekly resolution from 2000–2024.GLORYS12 data are included at the same resolution to complement, not replace, the observational data.
AI opportunities: The dataset contains 9.5 million profiles across 1283 weekly time steps and supports standardized evaluation with baseline results for subsurface ocean-state reconstruction.Its four-dimensional latitude–longitude–depth–time structure and near-empty grids create challenges for AI and computer-vision methods.
Open resources: OceanDepths releases the data, code, baselines, preprocessing scripts, configurable patching system, and PyTorch DataLoaders as open-source resources.

2 Related Work

Existing ocean datasets provide subsurface profiles, reanalysis fields, reconstructed products, or forecasting benchmarks, but typically lack raw paired observations at high resolution and broad historical coverage. OceanDepths addresses these gaps with globally paired surface–subsurface observations, matched reanalysis, and an AI-ready common-grid format.

Subsurface profiles: Argo and related measurement programs provide temperature and salinity profiles, while EN4 and WOD compile such observations with bias correction and quality control.Raw profiles are measured at irregular depth levels that vary across floats and individual casts.
Reanalysis and analysis products: GLORYS12 provides spatially dense, dynamically consistent 3D reanalysis fields, but model and assimilation biases can smooth observations and introduce spurious correlations.Reported discrepancies include residual seasonal temperature bias above 100 m and significant deep-ocean heat-content differences against independent observations.
Reconstructed gridded products: Reconstructed gridded products derive subsurface variables from satellite and profile data, but produce gap-filled, smoothed model outputs rather than raw paired observations.Examples include global temperature and salinity products at 23 depth levels or 0.25° monthly resolution.
Ocean forecasting benchmarks: OceanForecastBench and OceanBench support AI ocean prediction but rely heavily on reanalysis, use coarser grids or limited observational evaluation periods, and provide profiles mainly for evaluation.OceanForecastBench uses 1.41° data across 23 depth levels, while OceanBench observational profiles are limited to 2024 and OceanForecastBench to 2022–2023.
OceanDepths contributions: OceanDepths contributes raw paired observations at 0.1° weekly resolution over 25 years, combines SST, SSS, SSH, subsurface temperature and salinity, and includes matched GLORYS12 reanalysis.Its common-grid, configurable patching system supports tensor loading, while matched reanalysis enables curriculum or multi-stage learning and immediate comparison.

3 The OceanDepths Dataset

OceanDepths is presented as a documented dataset with data sources, construction, storage, and patching systems, alongside publicly available processing and PyTorch DataLoader code. It is provided as a static resource for machine-learning training and intercomparable evaluation rather than continuous updating.

The section covers OceanDepths’ data sources, construction pipeline, storage format, and patching system.
Data export, processing scripts, and PyTorch DataLoader code are publicly available under a permissive CC BY 4.0 license.
OceanDepths is provided as a static resource for machine-learning training and intercomparable evaluation, without planned continuous updates.

3.1 Data Sources

OceanDepths integrates subsurface EN4 profiles, satellite-derived surface observations, and GLORYS12V1 reanalysis. These sources provide complementary measurements of ocean temperature, salinity, height, and subsurface structure for observation-based and reanalysis-supported analyses.

Data sources: OceanDepths combines quality-controlled EN4 profiles, gridded satellite surface products, and GLORYS12V1 global ocean reanalysis.EN4 integrates Argo floats with ship-based CTD and other hydrographic measurements, while GLORYS12V1 supplies dense three-dimensional fields.
Subsurface observations: EN4 profiles provide temperature and bias-corrected salinity at instrument-specific corrected depths, with up to 400 samples per profile.Raw measurements occur at irregular depth levels that vary across floats and casts.
Surface observations: Satellite products supply daily SST, SSS, and absolute dynamic topography, with native resolutions of 0.05° for SST and 0.125° for SSS and ADT.The selected products are OSTIA SST, MULTIOBS SSS, and DUACS multisatellite altimetry.
Reanalysis: GLORYS12V1 provides daily three-dimensional temperature and salinity fields at 1/12° (∼8 km) resolution across 50 vertical levels spanning 0.49 m to 5728 m.The reanalysis assimilates in situ profiles, satellite altimetry, SST, and sea-ice concentration through a reduced-order Kalman filter.
Reanalysis applications: GLORYS12 can support reciprocal training and validation, second-stage finetuning on Argo profiles, discrepancy analysis, and dense spatial supplementation of sparse sampling.These uses include training on reanalysis and validating against EN4 observations, or reversing that direction.

3.2 Data Processing

OceanDepths aligns EN4 profiles with weekly satellite and GLORYS12 products on a common 0.1° global grid and 50 fixed depth levels. The pipeline preserves missingness without extrapolation, co-locates surface context, and distributes dense rasters alongside aligned and raw profiles.

Depth alignment: Profiles are linearly interpolated onto 50 fixed GLORYS12 depth levels using a depth-adaptive acceptance criterion, with rejected depths marked missing and no extrapolation.The criterion is more restrictive near the surface and more permissive at depth.
Depth alignment: 9.5M profiles have valid temperature and 7.0M have valid salinity after alignment.Aligned examples show EN4 and GLORYS12 profiles tracking across the water column, with occasional discrepancies.
Spatial and temporal alignment: Weekly SST, ADT, SSS, and GLORYS12 fields use centered 7-day means, and EN4 profiles are assigned to the corresponding measurement week.For each retained profile, co-located surface and GLORYS12 variables are saved with its assigned raster-cell index.
Data formats: OceanDepths is distributed as quantized dense GeoTIFF rasters, aligned-profile Zarr archives with collocated context, and unaltered raw-profile Zarr archives.The aligned archives include temperature, salinity, validity masks, surface observations, GLORYS12 estimates, and parquet indices for spatiotemporal querying.

3.3 Configurable Patching and Data Splits

OceanDepths provides configurable PyTorch DataLoaders and configuration files for overlapping spatial patches, land-fraction filtering, regional exceptions, and date-based data splits. Recommended settings use overlapping 128×128 patches with configurable regional thresholds, while reserving 2018 for evaluation.

Configurable patching: PyTorch DataLoaders and configuration files support patched overlaps, land-pixel cutoffs, relaxed regional rules, and adjustable settings such as patch size.The system is designed for ready-to-use loading and configuration of patch construction and filtering behavior.
Configurable patching: 128×128-pixel patches span 12.8°×12.8° on the 0.1° GLORYS12 land-mask grid, using a 32-pixel stride and 75% overlap.Candidate patches are filtered using land fraction computed from a binary land mask.
Configurable patching: Patches with more than 30% land are discarded, while configured regional boxes retain patches under thresholds of 60% for the Mediterranean, 85% for the Baltic and Red seas, and 95% for Hudson Bay.Regional exceptions preserve enclosed or narrow basins that default filtering would otherwise remove, and the areas and constraints are editable.
Data splits: 2018 is reserved for evaluation and all other years form the training set, preventing spatial leakage between splits.The date-based split contains 343 K training and 15 K evaluation non-overlapping patches at 128×128 crop size.
Data splits: With recommended strided patching, the training and evaluation sets increase to 4.8 M and 205 K patches, respectively.These counts contrast with the 343 K and 15 K non-overlapping-patch counts at the same crop size.

4 Baseline Experiment

The baseline experiment reconstructs the subsurface thermohaline state from weekly surface context, sparse in situ profiles, and a land/ocean mask, evaluating predictions against held-out EN4 observations and GLORYS12. Climatology and IDW remain strong temperature baselines, while U-Net models improve explained variance and show the value of spatial context.

4 Baseline Experiment: Each input is a 128×128 weekly patch with SST or SSS, sparse matching Argo profiles, and a land/ocean mask; the model predicts 50-depth-level 3D fields.Evaluation uses held-out EN4 profiles and the dense GLORYS12 cube.
4 Baseline Experiment: The comparison includes climatology, nearest-profile IDW, point-wise LSTM, point-wise CNN, and spatial U-Net encoder-decoders with 2D/3D convolutions.The climatology aggregates EN4 training samples by patch and interpolates them with inverse-distance weighting over 2000–2024, excluding 2018.
4 Baseline Experiment: Performance is measured with RMSE, MAE, and R2 at individual depths and in depth-integrated form, using 80% input profiles and 20% held-out profiles.Metrics are evaluated exclusively at held-out profile locations and averaged across depths no deeper than 2000 m in Table 3.
4 Baseline Experiment: Climatology and IDW are strong temperature baselines for RMSE and MAE, while learned U-Nets remain competitive and improve explained variance.The results indicate that seasonal background and nearby profiles capture much of the weekly thermal structure, with spatial context benefiting reconstruction.

5 Conclusion

OceanDepths is an open, global, AI-ready dataset pairing satellite surface fields with in situ subsurface observations and matched GLORYS12 data at high spatiotemporal resolution. The authors demonstrate that subsurface reconstruction is challenging and identify sampling and aggregation limitations affecting use.

Contribution: OceanDepths pairs SST, SSS, and ADT with Argo profiles and matched GLORYS12 reanalysis data at 0.1° resolution and weekly intervals from 2000–2024.The dataset contains 9.5 M profiles across 1283 weekly dates and supports a range of ocean-related ML tasks.
Results: Subsurface ocean state reconstruction is challenging because predicting climatology often outperforms simple ML baselines.The authors present reconstruction as one task enabled by OceanDepths and hope it will spur research at the intersection of ML and oceanography.
Limitations: OceanDepths inherits Argo sampling biases, with sparser coverage in marginal seas, near coasts, under ice, and below 2,000 m.Reliance on L4 products, depth-wise interpolation, and weekly aggregation improves usability and provides a shared baseline, but may not be optimal for every user.
Availability: The dataset, code, and reproducible loading examples are publicly available on Hugging Face under a CC BY 4.0 license.The availability passage provides the OceanDepths dataset URL and describes the release as approximately 120 GiB.

A Appendix: Extra figures

The appendix shows that EN4 depth sampling is irregular and profile-dependent, motivating interpolation onto a fixed 50-level GLORYS12 vertical coordinate.

Figure 7: Irregular, profile-dependent EN4 depth sampling motivates interpolation onto the fixed 50-level GLORYS12 coordinate.Figure 7 compares EN4 depth indices with corrected depths in meters using a log-scale heatmap, median curve, and P10/P90 curves.