introduced the OceanBench framework, which provides standardised preprocessing and evaluation procedures for sea surface height (SSH) interpolation tasks. OceanBench aggregates curated datasets derived from high-resolution NATL60 simulations together with simulated and real nadir altimetry observations, and includes benchmarking pipelines for method comparison. The framework has contributed substantially to improving reproducibility in regional SSH reconstruction studies, particularly in the Gulf Stream region. Extending this effort, provide a more comprehensive benchmark for evaluating global data-driven ocean forecasting across sea surface height, temperature, salinity, and currents. The framework introduces curated datasets from GLORYS12, GLO12, and ECMWF Integrated Forecasting System (IFS) into three standardised evaluation tracks to ensure physical consistency and reproducibility in deep learning-based ocean modelling .

While these frameworks represent important steps toward standardised analyses-ready data collections and benchmarking, they are primarily designed for predefined challenge configurations or focus on specific geographic regions. In contrast, OceanTACO is conceived as a global, multi-source data collection that harmonises observational, reanalysis, and auxiliary surface variables within a unified and queryable specification. Rather than prescribing fixed evaluation tracks, OceanTACO provides flexible spatiotemporal and sensor-level subsetting capabilities, enabling researchers to configure region-specific studies, cross-mission validation experiments, and forecasting setups within a consistent data framework. To our knowledge, no existing framework harmonises SWOT-era L3 observations, L4 gridded products, reanalysis fields, and in situ profiles from diverse ocean missions within a single, uniformly organised specification that preserves per-product provenance while supporting the same query workflow across all modalities.

When multi-sensor ocean analyses are combined from independently distributed archives, each dataset introduces its own unique processing requirements. These choices must be resolved at the preprocessing stage and, when undocumented, introduce sources of variation between studies. Because such decisions are rarely reported in full in published methods sections, nominally comparable analyses may differ in ways that are not apparent from the description alone . The resulting methodological variability limits the comparability of results across studies and complicates efforts to reproduce or extend existing analyses.

Format-level standards such as Zarr and NetCDF4/HDF5 eliminate heterogeneity in binary encoding but do not prescribe how variables, coordinates, or metadata are organised within a file or across a collection. Catalogue standards such as STAC address the related but distinct problem of discovery: they specify where data can be found, not how different products are internally organised or how they relate to one another across sensors and processing levels. As a result, multi-source preprocessing workflows remain data product-specific even when all constituent files conform to a common format or are discoverable through a shared catalogue.

The TACO specification addresses this by enforcing a uniform internal organisation at the sample level: every entry in the catalogue replicates an identical subfolder and file-type layout, regardless of which sensors it contains or at which processing level they were acquired . Because the internal structure is invariant across products, data access procedures developed for one sample apply without modification to any other sample in the collection.

The naming of the specification reflects two complementary design principles. Transparent refers to data access being explicit and auditable: data are referenced through explicit file paths, so every retrieval operation can be traced, replicated, or independently verified. Cloud-optimised refers to the storage layout: NetCDF4/HDF5 files with internal chunking support partial, byte-range reads from remote object storage, so that the same access procedure operates identically on locally held data or on data hosted in cloud object stores, without requiring full-file retrieval before analysis can begin.

Together, these properties mean that analysis workflows developed for one region or sensor transfer to other regions and sensors without modification to the access procedure, improving reproducibility or extensions by other research groups.

OceanTACO integrates ten data sources spanning satellite altimetry, sea surface temperature, sea surface salinity, surface winds, ocean reanalysis, and in situ Argo profiles (Table ). The table consolidates the horizontal resolution, stored temporal cadence, vertical coverage, and coverage across the Core and Extended releases for each source in one place.

OceanTACO is released in two contiguous temporal segments. The core dataset spans 29 March 2023 to 1 August 2025, covering the currently available SWOT period, including the calibration phase. We additionally provide an extended pre-SWOT segment spanning 1 January 2015 to 29 March 2023 for all non-SWOT sources. The 2015 start follows the salinity record. The multi-observation salinity product included here combines SMAP, SMOS, satellite SST, and in situ salinity through multivariate optimal interpolation, and reaches this configuration in 2015 when SMAP becomes available. From 2015 the sensor constellation underlying the collection stays constant, with the L3 SSH missions and, from 2023, SWOT adding sampling without removing any modality. Several constituent products extend further back in time, and earlier years can be appended for backward-compatible sources through the same folder-mode organisation and public processing pipeline without restructuring the dataset. Spatially, OceanTACO provides global coverage between 90° S and 90° N in a WGS 84 (EPSG:4326) projection, with variable-specific native spatial resolutions ranging from approximately 2 km (SWOT) to 0.25° (surface winds). While the WGS 84 projection is not area-preserving, and will lead to significant distortions towards the poles, it is the native projection of both GLORYS and L4 data and was therefore used as the basis for regridding of altimetry track and SWOT data. The extended and core datasets share the boundary date of 29 March 2023, allowing seamless concatenation.

The complete core dataset comprises 857 daily temporal indices, subdivided into eight regional tiles per day. The extended dataset comprises 3010 daily temporal indices. The total storage volume is approximately 311 GB for the core period and 581 GB for the extended period in compressed form, accounting for all stored variables across modalities; Table  reports the storage for the primary variable of each modality as a representative indicator of compression performance. OceanTACO is publicly accessible through Hugging Face for direct cloud access.

Because OceanTACO integrates heterogeneous data products with differing spatial resolutions, sampling geometries, and file formats, we implemented a standardised processing workflow to harmonise the collection while preserving native observation characteristics.

The processing pipeline consists of three primary steps:

Regional tiling. Global daily data files are partitioned into eight equally sized geographic regions in their native WGS 84 projection to facilitate efficient storage and selective retrieval as depicted in Fig. .

Gridded L4 and reanalysis products retain their native grids, L3 nadir and SWOT observations are conservatively binned rather than interpolated, and Argo profiles retain their original sampling coordinates and timestamps. Each product is ingested at its published, quality-controlled processing level, and OceanTACO performs no additional filtering or outlier removal; the data therefore retain the quality of their authoritative source, and application-specific quality decisions are left to the user.

Building reproducible workflows from independently archived ocean datasets demands substantial custom preprocessing. The SpatioTemporal Asset Catalog (STAC) standard is widely used for geospatial dataset discovery and cataloguing; however, it addresses data discovery rather than the internal organisation of samples. As a result, each dataset still requires product-specific loading code and inspection before common data access routines can be written. This is especially challenging when combining heterogeneous data types, including satellite altimetry, sea surface temperature, salinity, winds, and in situ profiles, each of which originates from distinct archives and formats.

OceanTACO implements the TACO (Transparent Access to Cloud-optimised) specification  to resolve this. TACO is fully compliant with STAC, and any TACO dataset can be mapped to a valid STAC catalog without information loss. Additionally, TACO directly follows the FAIR principles proposed by . The contribution of OceanTACO is therefore not a new catalog standard, but the application of a sample-invariant TACO organisation to a heterogeneous ocean collection spanning L3 observations, L4 products, reanalysis fields, and Argo profiles. Beyond cataloging, TACO enforces consistent internal organisation across all samples within a dataset. For example, if one sample contains folders GLORYS/, SWOT/, and ARGO/ with specific file types, every other sample replicates this exact structure. TACO references data through GDAL Virtual File System (VSI) paths rather than HTTP URLs. VSI provides a unified file access abstraction, so the same reading code operates transparently on local filesystems, cloud object stores such as S3 or GCS, and compressed archives, with only the path prefix varying between backends. Consequently, a data loading pipeline developed for a local copy of OceanTACO functions without modification when the dataset is hosted remotely, and partial reads of large files are managed natively by the GDAL driver, eliminating the need for full downloads. TACO supports two storage modes: ZIP mode, which packages the dataset as a single archive for static distribution, and folder mode, which stores data as a directory tree on disk or cloud object storage. OceanTACO utilises folder mode, which allows the dataset to be extended with new satellite passes, reprocessed products, or additional variables by simply writing files into the directory tree and appending rows to the metadata catalog, without modifying existing components.

Access is structured in three phases (Fig. ). The CONNECT phase loads Parquet metadata into DuckDB for lazy evaluation and presents the catalog as relational tables. The QUERY phase filters these tables by geographic extent, observation period, or data source, and materializes results as DataFrames. The FETCH phase retrieves individual files on demand through GDAL VSI handlers, which read from local paths or cloud object storage. Both the core (10.57967/hf/8171, ) and extended (10.57967/hf/8172, ) dataset are available on Hugging Face.

Systematic validation of gridded products against independent observations is essential for quantifying structural biases, effective resolution, and regional uncertainty characteristics. Such analyses require reproducible collocation procedures, consistent temporal matching, and transparent spatial aggregation strategies.

Because all OceanTACO components share a common temporal index and regional tiling scheme, data retrieval operations are reproducible across studies. L3 observations, L4 products, reanalysis fields, and Argo profiles can be queried under identical spatial and temporal constraints, reducing ambiguity in matching procedures and methodological variability between studies.

OceanTACO preserves the native sampling geometry and spatial resolution of each product, enabling controlled comparisons across processing levels without introducing additional harmonisation artifacts. Such comparisons are particularly useful when evaluating differences between L3 observations, which retain native measurement characteristics, and L4 mapped or assimilated products, which introduce spatial smoothing and model-based constraints.

As a representative workflow, OceanTACO enables computation of wavenumber power spectral density (PSD) diagnostics across processing levels. A robust cross-level comparison requires that the spectra be evaluated over identical spatial sampling; otherwise differences in ground-track geometry, coverage, and segment length introduce artifacts that are easily mistaken for resolution differences. OceanTACO makes this matching straightforward: for dynamically active regions such as the Gulf Stream (80–40° W, 25–45° N) or the Kuroshio Current (130–160° E, 25–45° N), users compute the L3 SWOT spectrum from its native 2 km observations and bilinearly sample the L4 DUACS mapped field at the identical SWOT track positions, so both spectra share the same along-track segments. Figure  illustrates this matched comparison produced using OceanTACO from 1 March to 29 May 2025. The comparison is intended as an illustrative diagnostic rather than a comprehensive spectral analysis.

Because the two products are sampled along common track segments, their spectra coincide at large scales (≳100–150 km), where both resolve the dominant geostrophic signal, and diverge only toward shorter wavelengths. There, the optimal-interpolation mapping underlying the L4 product suppresses mesoscale and submesoscale variance, whereas the denoised L3 SWOT observations retain substantially more energy down to a few tens of kilometers, before the spectrum flattens near the effective-resolution limit of the L3 product. The separation over roughly 30 to 150 km isolates the genuine effective-resolution difference between an L3 observation product and an L4 mapped product, rather than a sampling artifact. Independent analyses similarly indicate that SWOT observations resolve ocean variability at spatial scales on the order of tens of kilometers, a substantial improvement over the effective resolution of mapped altimetry products .

This example illustrates how OceanTACO enables reproducible cross-product comparisons under identical spatiotemporal constraints. Such workflows are increasingly relevant in the SWOT era for evaluating scale-dependent variance and interpreting differences between observation-level measurements and mapped products, and apply equally to other eddy-rich regions such as the Antarctic Circumpolar Current in the Southern Ocean.

Beyond comparisons across processing levels, many studies seek to quantify the contribution of individual observing systems to reconstructed ocean fields.

The introduction of new observing systems, such as the Surface Water and Ocean Topography (SWOT) mission, enables new assessments of their incremental contribution relative to established nadir altimetry and assimilated products. Early SWOT analyses emphasise its capability to resolve fine-scale ocean variability and improve characterization of mesoscale and submesoscale dynamics relative to nadir altimetry, while also highlighting the need for careful cross-comparison with existing mapping systems and climate data records .

Observation system experiments (OSEs) and mission impact studies are therefore essential for studying how new sensors improve resolved spatial scales, dynamical consistency and downstream reconstruction performance. These studies typically require selective inclusion or exclusion of specific sensors, consistent regional subsetting, and reproducible definition of validation datasets across multiple processing levels.

OceanTACO preserves mission identifiers and full product provenance within a unified indexing framework. Researchers can configure controlled experiments that isolate nadir-only configurations, incorporate wide-swath SWOT observations, or evaluate L4 fields with and without specific observational inputs. The per-mission temporal coverage that determines which sensors are available for a given experiment is documented in Table , and the matched SWOT-versus-mapped spectral comparison in Sect.  (Fig. ) already illustrates the kind of resolved-scale diagnostic that such configurations support. Because these configurations are defined through explicit spatiotemporal queries rather than bespoke preprocessing pipelines, mission impact assessments can be reproduced consistently across regions and time periods. Because the eight-region partition is a storage-and-retrieval optimisation rather than an analysis unit, experiments are defined on arbitrary user-specified bounding boxes at each product's native resolution, so subregional and fine-scale phenomena remain fully accessible.

In addition to structured workflow categories such as validation or observation system experiments, OceanTACO can facilitate rapid exploratory analyses of individual oceanographic events. The common temporal index means multi-sensor data of any event can be retrieved with a single spatiotemporal query. As an illustrative example, we examine sea surface height variability during Hurricane Milton, a major hurricane that occurred in the Gulf of Mexico during October 2024.

Satellite altimetry has been widely used to investigate extreme sea-level events, including storm surges, by measuring sea surface height (SSH) anomalies along satellite ground tracks. Previous studies have used along-track altimetry to detect and analyse storm surges across large ocean regions . However, the narrow sampling geometry of traditional nadir altimeters limits their ability to resolve the full spatial structure of these events . The wide-swath measurements from the SWOT mission provide two-dimensional SSH fields at high spatial resolution, enabling more detailed observation of the spatial variability associated with extreme sea-level signals .

used SWOT observations to examine SSH anomalies during Hurricane Milton, one of the strongest and most destructive hurricanes recorded in the Gulf of Mexico, which subsequently tracked across central Florida into the Atlantic Ocean . Figure  shows snapshots of L3 Altimetry and SWOT data, as well as L4 wind data along the hurricane track. On 9 October 2024, SWOT directly overpassed the hurricane eye, allowing to analyse the spatial structure of the observed anomaly. Figure  presents a cross-product comparison and correlations between SWOT and conventional altimetry with the L4 DUACS product. For each L3 observation within the domain, the collocated L4 DUACS value was obtained by bilinear interpolation of the gridded field onto the along-track measurement position, and Pearson correlation coefficients and root-mean-square errors were computed over all valid collocation pairs.

While the previous examples focus on observational analysis and product intercomparison, an increasing number of Earth system studies rely on data-driven models that learn relationships between multiple surface variables. These approaches require consistent collocation of heterogeneous datasets across space and time, which can be technically challenging when data originate from independent archives. OceanTACO directly supports such workflows by providing aligned multi-variable samples across observations, gridded products, and reanalysis fields.

Reconstruction of spatially complete ocean surface fields from sparse and irregular observations is a core challenge in Earth system science. For sea surface height (SSH), operational optimal interpolation approaches provide global coverage but may not capture mesoscale variability and reduce effective resolution below that of the observational inputs . In response, a range of learning-based methods has been developed to improve spatiotemporal reconstruction skill: supervised networks that fuse multiple satellite variables into improved gridded SSH , end-to-end learned variational assimilation schemes for nadir and wide-swath altimetry , dynamically constrained joint reconstruction of SSH and temperature from multi-sensor observations , generative assimilation of multi-modal satellite observations , and unsupervised spatio-temporal interpolation of altimetry maps .

Similar developments have emerged for other ocean surface variables. Deep learning approaches have demonstrated improved gap filling and super-resolution for sea surface temperature , while analogous strategies have been proposed for sea surface salinity . These approaches increasingly rely on multi-variable learning, where relationships between ocean surface variables provide additional constraints for reconstruction and forecasting. However, building such datasets requires coordinated access to heterogeneous observations and gridded products that must be consistently collocated in space and time.

OceanTACO directly supports such analyses by enabling consistent retrieval of L3 observations (e.g., nadir altimetry, SWOT, SST, and SSS), L4 gridded products, reanalysis fields, and in situ measurements under identical spatial and temporal constraints. As an illustrative example, Fig.  shows the spatial correlation between L4 SSH and both SST and SSS in two dynamically distinct regions: the Gulf Stream and the North Indian Ocean. The analysis, performed for September 2023, highlights strong spatial variability and region-dependent relationships between surface variables. Such nonlinear and spatially heterogeneous coupling suggests that data-driven approaches are well suited to exploiting correlated information in reconstruction tasks.

Beyond exploratory analysis, OceanTACO also supports the generation of machine learning–ready training and evaluation datasets. Based on the same spatiotemporal queries used in the correlation analysis, the accompanying documentation (https://oceantaco.readthedocs.io/en/latest/tutorials/ml_dataset.html, last access: 24 August 2026) demonstrates how OceanTACO can be used to retrieve aligned multi-variable samples and construct PyTorch dataloaders for model training. Because input-target configurations are defined through the spatiotemporal queries used throughout the collection, training and evaluation schemes of data-driven reconstruction experiments become more reproducible.