Summary

García-Espriu conduct an observing system simulation experiment (OSSE) to evaluate the feasibility of reconstructing ocean interior temperature and salinity from in situ observational data and satellite observational data products. The authors leverage output from the CMEMS Global Ocean Ensemble Reanalysis product to conduct this experiment, and they subsample the product at times and locations where Argo float profiles are available. They then use these subsampled synthetic profiles to train machine learning models, which they apply to satellite products to reconstruct ocean interior properties, and they compare these reconstructions against the reanalysis “truth” to evaluate the skill of the reconstruction methods.

The authors find that the more complex versions of their random forest regression (RFRv2) model and Long-Short Term Memory (LSTMv2) network are able to reproduce ocean temperature with an R² of around 0.85 and salinity with an R² of around 0.95. They validate their models with synthetic profiles withheld from model training and by using a regional subsection of the reanalysis dataset. They report validation statistics spatially and by depth, concluding that the RFRv2 model performed better in terms of the evaluation statistics against the test dataset but the LSTMv2 model was better able to represent the data in terms of variability over time and space. The authors also use SHapley Additive exPlanations (SHAP) to interpret their trained models.

Overall, I support the approach this manuscript takes to question of how in situ and satellite observing systems can be leveraged to reconstruct ocean interior properties. However, it falls short in its execution and interpretation of the analysis. Most importantly, the authors could attempt to remedy or discuss more extensively the shortcomings of the models to predict ocean interior variables from primarily surface data and the results could be better placed into context among similar studies that reconstruct ocean interior properties from observational data.

General suggestions

One aspect that I think is missing from the manuscript is the contextualization of the authors’ results with similar methodologies that have been applied to map salinity and temperature from observations (a few of which are referenced in the introduction). Although not all studies that reconstruct ocean interior properties from observations include a reanalysis-based evaluation of mapping accuracy (as is the focus of this manuscript), many report error statistics of their reconstructions evaluated against independent data. Su et al. (2018), for example, evaluate their reconstructed subsurface temperature anomalies using root mean squared error and R² as metrics, and the results of the OSSE reported here could be evaluated against those results.

In general, I was surprised to see such high disagreement with the test data at depth, when temperature and salinity should be more constant in space and time, and therefore relatively easier to reconstruct than at the surface. Buongiorno Nardelli (2020), for example, retrieve minimum errors for temperature and salinity at depth. This points, in my opinion, to an aspect of the methodology that can be significantly improved. It is not particularly surprising that a model based primarily on surface characteristics would struggle to estimate temperature and salinity at 1000 meters. I suspect a strategy of somehow de-emphasizing the impact of the surface predictor datasets as depth increases might improve these high offsets at depth. In any event, this is another instance where contextualization of the results of this OSSE would be helpful.

Lastly, the authors miss an opportunity to incorporate uncertainties into their experiment, or at least to discuss their implications. OSSEs present an opportunity to mimic real-world conditions; in reality, satellite observations are not perfect, nor are temperature and salinity measurements from profiling floats. Incorporating measurement uncertainty estimates in the analysis would be an important piece for answering the central question of how feasible it is to use satellite and in situ data to reconstruct ocean interior properties.

Line-by-line comments

Abstract: I would suggest defining the simulated in situ measurement platforms as “Argo floats” or “profiling floats” rather than buoys in the abstract.

28: Presumably this should say “subsurface temperature and salinity”?

86: Awkward phrasing in reference to the equatorial region.

97: punctuation issue here

161-166: I’m not sure I understand the training and test split. Are you withholding some percentage of the dataset on a daily frequency (if so, what percentage?) for testing during model training? How does this differ from the ground truth dataset that is being used for evaluation?

173: It would be helpful to specify the metric you are referring to when discussing “accuracies”

239: What is meant by “it does not overlap with the training dataset”? There are no Argo profiles from 2008-2009 in this region?

272: should be “…each of them with their own…”

Review of the paper "On the reconstruction of ocean interior variables: a feasibility data-driven study with simulated surface and water column observations"

by Aina García-Espriu, Cristina González-Haro, and Fernando Aguilar-Gómez

This study investigates the feasibility of reconstructing ocean interior variables, specifically temperature and salinity profiles, using AI-based algorithms applied to simulated satellite surface data and in-situ buoy observations. Leveraging an Observing System Simulation Experiment (OSSE) with outputs from a numerical ocean reanalysis model from the EU Copernicus Marine Service, the authors compare the performance of Random Forest Regressors (RFR) and Long-Short Term Memory (LSTM) networks. The results show that both models reasonably capture the spatial and temporal variability of ocean interior conditions (particularly for salinity), with RFR offering higher accuracy in direct reconstructions and LSTM demonstrating better extrapolation capabilities with ground truth observations. The findings highlight the potential of data-driven approaches to enhance 4D ocean reconstruction and contribute to future digital twin ocean frameworks, while also identifying current challenges in capturing vertical variability and reducing biases. Nevertheless, the study lacks some aspects which should be integrated at least at the discussion level.

Major Points

• I could not fully understand how surface information is synthesized from the Copernicus marine service numerically modelled data. To the best of my understanding, the aim is to provide insights on a potential 4D reconstruction that exploits satellite based surface observations. In particular, the Authors claim the intention to perform reconstructions at the spatial resolution provided by space-based microwave sensors. However, it seems surface observations are directly extracted from modelled surface data. To be consistent, an assessment of the type and effective resolutions of satellite input data should be performed and the synthetic input data should be adjusted accordingly. For example, present-day satellite-based sea surface heights/currents/temperature could differ significantly with respect to the outputs of a hydrodynamic model. A discussion on how this could impact the results of the 4D reconstruction could be beneficial;

• On the same note, I think the paper lacks discussions on the capability of current satellite missions and, more importantly, future missions for Earth observations in the microwave band, how this could impact e.g. sea surface temperature and salinity monitoring and which could be the impact of such missions on the proposed ocean 4D reconstruction. I think this should also be integrated in the discussion section, at least.

Minor Points

• I was wondering if the proposed reconstruction methodology is able to provide un uncertainty estimate to verify if the profiles provided in Figure 10 can be considered significantly different. Could the Authors quickly comment on that?
• Have the Authors tried to inter-compare the feature-resolution of the reconstructed fields versus the ground truth? Are you expecting significant differences?
• Could the Authors also provide a broad overview of which could be the “real in-situ and satellite” data more suitable for their future applications?

Typos

• In general, please always use Copernicus Marine Service instead of CMEMS when referring to data generated within the EU Copernicus Marine Service
• Line 88: earth-> Earth