Deep Learning-Based 4D Reconstruction of Arctic Ocean Hydrography and derived Geostrophic Currents from Satellite and In Situ Observations

Werner-Pelletier, Nicolas; Crespin, Júlia; Rosquete, Aleida; Sánchez-Urrea, Maria; Hoareau, Nina; Martin, Mario; Umbert, Marta

doi:10.5194/essd-2026-519

Preprints

https://doi.org/10.5194/essd-2026-519

Preprints

21 Jul 2026

| 21 Jul 2026

Status: this preprint is currently under review for the journal ESSD.

Deep Learning-Based 4D Reconstruction of Arctic Ocean Hydrography and derived Geostrophic Currents from Satellite and In Situ Observations

Nicolas Werner-Pelletier, Júlia Crespin, Aleida Rosquete, Maria Sánchez-Urrea, Nina Hoareau, Mario Martin, and Marta Umbert

Abstract. The Arctic Ocean plays a critical role in the global climate system through its influence on freshwater storage, ocean circulation, sea-ice dynamics, and air–sea heat exchanges. However, subsurface observations remain sparse in space and time, limiting the characterization of Arctic hydrographic variability and circulation changes. This study presents an observation-constrained Arctic Ocean reconstruction framework based on a Long Short-Term Memory neural network (LSTM), combining satellite-derived surface variables and in situ hydrographic profiles to infer subsurface temperature and salinity profiles.

The framework is applied over the 2011–2021 period to reconstruct four-dimensional temperature and salinity fields across the Arctic Ocean at 3-day temporal resolution and on 102 WOA standard depth levels. A pan-Arctic reconstruction is provided on a 25 km EASE grid, together with four higher-resolution regional reconstructions on 6.25 km EASE grids covering the main Arctic gateways: Bering Strait, Davis Strait, Fram Strait, and the Barents Sea Opening. The reconstructed temperature and salinity fields are also used to derive steric height, absolute dynamic height, and geostrophic currents.

Independent evaluation against withheld in situ observations shows that the reconstruction improves the representation of Arctic hydrography relative to the baseline reanalysis (GLORYS) across most regions and depth ranges, with the largest error reductions in the upper ocean and in ice-covered regions. The reconstructed fields reproduce the main large-scale hydrographic and dynamical structures of the Arctic Ocean, including the Beaufort Gyre, Arctic boundary currents, and the major gateways. They also enhance the representation of seasonal variability, freshwater accumulation, upper-ocean stratification, and Atlantic Water pathways entering through Fram Strait and the Barents Sea Opening.

The resulting five spatially and temporally continuous datasets provide observation-constrained estimates of Arctic hydrography and geostrophic circulation. They are intended to support studies of freshwater variability, Arctic circulation change, climate model evaluation, and data-driven Arctic Ocean research. The datasets are available at World Data Center for Climate (WDCC) via https://doi.org/10.26050/wdcc/dlrec-ao_v1.

Received: 03 Jul 2026 – Discussion started: 21 Jul 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Nicolas Werner-Pelletier, Júlia Crespin, Aleida Rosquete, Maria Sánchez-Urrea, Nina Hoareau, Mario Martin, and Marta Umbert

Status: final response (author comments only)

RC1:
'Comment on essd-2026-519', Anonymous Referee #1, 10 Oct 2026
Werner-Pelletier et al. use an LSTM framework integrating satellite observations, in situ hydrographic profiles, and GLORYS reanalysis to reconstruct four-dimensional Arctic Ocean temperature and salinity fields for 2011–2021 and derive geostrophic currents. The resulting datasets could provide a useful resource for investigating Arctic freshwater storage, stratification, and circulation variability. However, important questions remain regarding the independence of the validation data, the sources of reconstruction skill, and the accuracy of the derived currents and temporal variability. These issues require additional analyses and clarification before the reliability and intended applications of the products can be adequately assessed. I therefore recommend major revision, with detailed comments provided below.
Major comments
Please supplement the current evaluation with validation that withholds selected years or geographic regions. Report the sample coverage, bias, and improvements relative to GLORYS under these alternative partitions. Section 2.3 states that GLORYS assimilates in situ temperature and salinity observations. Please clarify the potential overlap between the test profiles used in this study and the observations assimilated into GLORYS. If this overlap cannot be determined, address this limitation in the Discussion. Observations withheld from LSTM training should not automatically be considered fully independent of the entire reconstruction framework.

Comparison with the original GLORYS fields demonstrates the benefit of the correction, but does not establish whether the improvement primarily comes from time-varying satellite information or from learning regional and seasonal biases in GLORYS. Please conduct ablation experiments that remove variables such as sea surface salinity (SSS), sea surface temperature (SST), or absolute dynamic topography (ADT) individually to quantify the contribution of each input. If feasible, comparison with an appropriate simpler model would also be useful. All experiments should use identical data partitions to establish when the framework adds information from satellite observations and when it mainly provides a statistical correction to GLORYS.

Geostrophic currents are a component of both the title and the delivered products, but their validation is substantially less developed than that of temperature and salinity. Please validate the currents against suitable mooring, acoustic Doppler current profiler (ADCP), or other independent velocity observations, accounting for differences between observed total currents and the diagnosed geostrophic component. If adequate validation cannot be provided, the claims concerning current accuracy and transport applications should be narrowed accordingly.

The current evaluation mainly consists of pointwise temperature and salinity root mean square error (RMSE), mean spatial structures, and seasonal mean corrections. Given the proposed applications to freshwater storage, stratification, and climate change research, please provide time series of ocean heat content, upper-ocean salinity, and liquid freshwater content for representative regions, depth ranges, or the whole domain. Compare the reconstruction with findings from previous studies, GLORYS, and available independent observational estimates, using consistent sampling, integration domains, and reference definitions. Where feasible, further illustrate the evolution of interannual anomalies.

Specific comments
Please consolidate or extend Tables 2, 3, and 5, drawing on the information in Section 3.5, to provide an overview of all input datasets, variables, and delivered products. For the inputs, include the source and version, temporal coverage, native spatial and temporal resolution, gap-filling treatment, and role in the framework. For the outputs, include variable definitions, units, dimensions, and grid spacing, distinguishing directly predicted variables, reconstructed fields, retained input fields, and diagnostically derived variables. Separate but complementary input and output tables may be clearer than one excessively large table.

Figure 3 mainly illustrates the network architecture. Although the processing steps are described in the text, an overview of the full reconstruction workflow would make it easier to identify which datasets enter at each stage and what roles they serve. Please redesign Figure 3 or add a workflow panel integrating the steps described in the text: data preprocessing and collocation, construction of input features and target anomalies, training/validation/testing, gridded inference, addition of predicted corrections to GLORYS, and calculation of dynamic height and geostrophic currents. Clearly distinguish training labels from inference inputs, and emphasize that the LSTM sequence progresses along depth rather than time.

“Global domain evaluation” in Section 4.1.1 could be interpreted as an evaluation over the global ocean. Please consider “Full-domain evaluation” or “Evaluation over the full study domain.” “Pan-Arctic evaluation” would also be appropriate if it accurately describes the geographic extent of the evaluated samples.

In Figure 6, according to the panel titles and color bars, panels c and d show RMSE improvement, whereas panels e and f show model uncertainty. These descriptions are reversed in the caption. Please correct the caption and check the corresponding references in the text.

The color-bar ranges in Figures 6c and 6d are asymmetric around zero, particularly for salinity. Please clarify the color normalization and ensure that improvements and deteriorations of comparable magnitude receive comparable visual emphasis. A zero-centered diverging color scale with symmetric limits would facilitate this comparison. If asymmetric limits are retained, please explain the choice and make the mapping around zero explicit.

Although the caption of Figure 15 defines the anomalies relative to GLORYS, “mean anomaly” may still be confused with a departure from a climatological mean. Please consider “mean difference relative to GLORYS” or other words in the panel labels, and use consistent terminology in the related figures and discussion.

Please add reconstructed-versus-observed plots for temperature and salinity with a 1:1 reference line. This would help reveal systematic departures that may be obscured in pooled comparisons.
Citation: https://doi.org/10.5194/essd-2026-519-RC1
RC2: 'Comment on essd-2026-519', Anonymous Referee #2, 04 Nov 2026

This manuscript presents a four-dimensional reconstruction framework for the Arctic Ocean based on an LSTM neural network, using satellite and in situ observations to estimate temperature and salinity profiles between 2011 and 2021. Using GLORYS reanalysis data as a baseline, the model overcomes observational limitations mostly in ice-covered regions, enhancing the representation of hydrographic variability at the surface and depth, as well as dynamic heights and geostrophic currents. The authors therefore generate consistent datasets for the entire Arctic Basin, as well as for the main Arctic gateways. The manuscript is worthy of publication, subject to a few minor corrections listed below.
Since I am not an expert in hydrography and Arctic circulation, my comments focus primarily on methodological aspects.
The data processing pipeline is clear and rigorous. To ensure full reproducibility, the authors made their open-source code available on GitHub and Zenodo, alongside the final published datasets on the World Data Center for Climate (WDCC). Note that the Zenodo link (https://www.doi.org/2010.5281/zenodo.20744925) is currently broken or not resolving.
Overall, the text and the figures are clear and precise. Regarding the figures, I would simply suggest that the authors include a map showing the locations of the cross-sections through the Fram Strait and the Barents Sea opening, presented in Figures 13 through 17, to help readers unfamiliar with the region’s geography.
Lines 270–272 and Table 4: The authors briefly note that "several hyperparameter configurations were tested, looking for improvements in validation loss, as well as different LSTM architectures in terms of the number of hidden units for LSTM1 and LSTM2". They do not specify the search algorithm used (e.g., whether it was manual trial-and-error, grid search, random search, or Bayesian optimization), nor do they disclose the full search space or validation scores for the discarded configurations. Adding a sentence or paragraph explaining the procedure used would make this section much clearer.
Lines 273–276. The authors should provide a more thorough justification for their use of Monte Carlo Dropout (MCDO). They could emphasize the advantages of applying MCDO to a single optimized hyperparameter configuration rather than building a multi-configuration ensemble. Specifically, highlight how single-model MCDO provides a mathematically grounded approximation of Bayesian uncertainty while drastically reducing memory/computational overhead and avoiding the spatial smoothing of ocean gradients that can occur when averaging across various model configurations.
Section 4.2.2 (Lines 532–537) Seasonal vs. Interannual Variability. The authors note that interannual variability is learned implicitly by the LSTM through the multi-year variations of surface satellite predictors (SST, SSS, ADT). However, because the reconstructed 4D fields are generated by adding the LSTM’s predicted anomalies back to the GLORYS reanalysis baseline, interannual signals are also explicitly inherited from GLORYS. Could the authors elaborate on the respective roles and relative contributions of these two distinct sources? Specifically, it would be valuable to clarify to what extent the year-to-year variability in the upper 100 meters is driven by the surface-forced anomaly corrections inferred by the LSTM versus the underlying physical baseline of GLORYS.
Conclusions (lines 631-624): Could the authors elaborate on this important point: e.g. ‘while MSE minimizes numerical discrepancies against observational profiles, purely data-driven predictions do not guarantee thermodynamic consistency. Independent temperature and salinity outputs can generate spurious density inversions. Incorporating a composite loss function – combining a statistical error term with physical constraints based on the thermodynamic equation of state and static stability penalties – would enforce physical consistency across the water column.’
References (line 764) The reference National Snow and Ice Data Center: A Guide to EASE Grids is listed in the bibliography without being explicitly cited in the text. Please insert the NSIDC reference at the first introduction of the EASE grid in the methodology.

Citation: https://doi.org/10.5194/essd-2026-519-RC2

Nicolas Werner-Pelletier, Júlia Crespin, Aleida Rosquete, Maria Sánchez-Urrea, Nina Hoareau, Mario Martin, and Marta Umbert

Data sets

Deep Learning-Based Four-Dimensional Reconstruction of Temperature, Salinity and Geostrophic Currents in the Pan-Arctic Ocean with 3-Day Resolution on EASE-Grid 2.0 at 25 km Nicolas Werner-Pelletier et al. https://www.wdc-climate.de/ui/entry?acronym=DLRec-AO_v1

Deep Learning-Based Four-Dimensional Reconstruction of Temperature, Salinity and Geostrophic Currents in the Bering Strait with 3-Day Resolution on EASE-Grid 2.0 at 6.25 km Nicolas Werner-Pelletier et al. https://www.wdc-climate.de/ui/entry?acronym=DLRec-AO_v1

Deep Learning-Based Four-Dimensional Reconstruction of Temperature, Salinity and Geostrophic Currents in the Davis Strait with 3-Day Resolution on EASE-Grid 2.0 at 6.25 km Nicolas Werner-Pelletier et al. https://www.wdc-climate.de/ui/entry?acronym=DLRec-AO_v1

Deep Learning-Based Four-Dimensional Reconstruction of Temperature, Salinity and Geostrophic Currents in the Fram Strait with 3-Day Resolution on EASE-Grid 2.0 at 6.25 km Nicolas Werner-Pelletier et al. https://www.wdc-climate.de/ui/entry?acronym=DLRec-AO_v1

Deep Learning-Based Four-Dimensional Reconstruction of Temperature, Salinity and Geostrophic Currents in the Barents Sea Opening with 3-Day Resolution on EASE-Grid 2.0 at 6.25 km Nicolas Werner-Pelletier et al. https://www.wdc-climate.de/ui/entry?acronym=DLRec-AO_v1

Model code and software

Ocean LSTM Profile Reconstruction Nicolas Werner-Pelletier https://doi.org/10.5281/zenodo.20744925

Nicolas Werner-Pelletier, Júlia Crespin, Aleida Rosquete, Maria Sánchez-Urrea, Nina Hoareau, Mario Martin, and Marta Umbert

Viewed

Total article views: 723 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
382	180	161	723	192	139

HTML: 382
PDF: 180
XML: 161
Total: 723
BibTeX: 192
EndNote: 139

Views and downloads (calculated since 21 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	130	39	6	175
Aug 2026	44	38	10	92
Sep 2026	71	35	103	209
Oct 2026	108	54	36	198
Nov 2026	29	14	6	49

Cumulative views and downloads (calculated since 21 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	130	39	6	175
Aug 2026	44	38	10	92
Sep 2026	71	35	103	209
Oct 2026	108	54	36	198
Nov 2026	29	14	6	49

Viewed (geographical distribution)

Total article views: 689 (including HTML, PDF, and XML) Thereof 689 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 11 Nov 2026

Download

Preprint (40993 KB)
Metadata XML

Short summary

The Arctic Ocean is changing rapidly, but many areas remain hard to observe. This work combines satellite measurements, in-water observations, and deep learning to create a new eleven-year enhanced picture of Arctic temperature, saltiness, and ocean currents. It can help researchers study freshwater change, circulation, sea ice links, and how polar changes may affect the global climate.


Total:	0
HTML:	0
PDF:	0
XML:	0