This paper is well-written, and some interesting conclusions can be drawn from the results. However I had trouble appreciating the paper as the conclusions were presented in what felt to be a very misleading way. Recommendations are given in the abstract and conclusions without emphasizing appropriate caveats making this paper feel biased and unscientific. I recommend major revisions to appropriately implement these caveats. The title and abstract are misleading with the conclusions not put in proper context. I would prefer if you took Southern Ocean out of the title. As it stands the paper tells me much less about the Southern Ocean than I had envisioned it would. Your decisions made in this paper are fine, but they significantly impact the results and cannot be hidden in the details. Indeed the experiment was designed to make nUSV do the best as you explain on line ~657. Here are some more specific comments (in no real particular order):

- The choice to reconstruct one year of a small box (10 degrees of lat by 20 degrees of long) and call that the Southern Ocean is quite a leap and should be presented clearly in the abstract so as to not mislead about results. This isn’t the Southern Ocean and this choice has significant implication on observing system recommendations.

- As you state, Figure 2c shows this model is inconsistent with both reconstructions. Line ~195 needs to be changed. This is significant as it means the pCO2 may be driven by a mechanism different from the real world. This is relevant to the objective of the study (i.e. you can predict largely from knowledge of SST) and you need to discuss this. If the correlation to SST was weaker you would need to know the biological component better and the statistics may be different. This also has significant implication on observing system recommendations.

-Is nUSV sampling realistic? How many platforms are envisioned? It seems very dense – how long would it take one saildrone to reach all those points? Shouldn’t the nUSV sampling look more like the ship track? Moreover, how many nUSVs would it take to sample the full Southern Ocean with that density?

-The goal of the floats is to give the large spatial structure and this aspect of them is not valued in your experiment design. This should be mentioned when giving recommendations.

- If one were to extrapolate this to the full Southern Ocean, which is about 100 times larger than your box, what would the monetary cost to sample with the density of your experiment for each platform? For example, if it would take 100 nUSVs or 200 wavegliders is that practical?

- Even in your box you may give some information about the monetary cost of each sampling effort to give some practical perspective.

- WGs can be “driven”, but you have chosen to give them a mooring-like program. This impacts the results. You finally mention this at the end, but not in the abstract when giving recommendations

- You give the caveat “We emphasize that these experiments are intentionally made to reproduce the sampling resolution of their real-world counterparts, not necessarily their spatial resolution in practice but at least the temporal one.” In the text, but it needs to be featured more prominently (e.g. in the title)

- It appears you train with mixed layer depth (MLD) for all platforms, but surface platforms don’t give MLD. How will you get that information and deal with the uncertainty from this? This wasn’t discussed, or I missed it, and this is a serious caveat.

- Temporal aliasing by the floats is exacerbated by the fact you don’t advect them in the model, so their quasi-Lagrangian nature is not taken advantage of. If you plotted pCO2 in figure 9 following a water parcel (or plotted it from integrating the full material derivative of pCO2) I assume it would be much smoother. What you show in figure 9 is likely dominated by advection and not air-sea fluxes. This discrepancy impacts your findings and should be addressed.

    Related to the figure 9 comment above, it seems odd to conclude the sampling need be at least daily (line 793) and that the floats should sample 1-2 days when the model output used is daily averaged. How much signal is really being aliased? Please do a more thorough time series analysis (e.g. show power spectra density to see how much variance is truly being aliased). You call into question a discrepancy with Bushinsky et al 2019. I suggest you repeat the calculation done in that paper with the model output. Since you have the model timeseries I suggest you be quantitative with your statements.

    My other major criticism is that one expects mapping estimates to never get worse when more data is added. If they are getting worse when more data is added it signifies a flaw in the mapping method and not the data sampling itself. The fact that SHIP(smr) + FLOAT(SAZ+PFZ) has a significantly higher overall RMSE than SHIP + FLOAT(PFZ) is troubling. It points to either a serious issue with your methods or that the uncertainty in the RMSEs you are discussing are very large. You need to explain what is going on here. (This also calls into question your claim of the importance of temporal sampling as it appears something else is going on as well.) Did you carry out WG(SAZ+PFZ), and if so did it also have this troubling behavior?

Minor:

• Model is 1 yr, but Southern Ocean has significant interannual variability and this should be discussed a bit more.
• Line 121 sentence is incomplete
• Supplementary Information text needs cleaning
• In SI you say what hyper-parameters you tune but you don’t give the final values you used. You should give these for reproducibility and to inform the results
• Some typos in figure captions. Check these
• Time, J, is stated as a predictor in some cases and not in others. Please check for consistency.

Summary and overall impression

The paper addresses an important question for the global carbon cycle community: how to reduce the uncertainties and biases of machine-learning based mapping approaches in the Southern Ocean, a data-sparse but globally important region. The authors create synthetic data by subsampling a high-resolution model in a subregion of the Southern Ocean over 1 year. The synthetic data resembles different observational platforms in terms of the typical temporal resolution of the different platforms. These platforms include ship, float, Windglider, and Saildrone data. They then run two different machine learning mapping approaches with these synthetic observations and compare the mapped reconstruction of the seasonal cycle to the actual model field to estimate the biases and uncertainties. They run the method multiple times with different subsets of synthetic data to highlight how sampling in different seasons, as well as with different types of observational platforms affect the uncertainty and bias. They find that the addition of wintertime ship data would greatly reduce the errors in the reconstructions. They also find that Saildrones are an optimal platform to address both the large-scale spatial and high-resolution temporal sampling and have the most effective impact on reducing the uncertainties and biases of the seasonal and annual mean reconstructions of air-sea CO2 fluxes in the Southern Ocean.

This paper addresses a crucial gap in our current knowledge and provides suggestions on how the carbon cycle community can improve current estimates of the Southern Ocean carbon fluxes, through an improved sampling strategy. I very much support the method of using synthetic data to create a sampling strategy, the paper is well-written and has clear figures that support the findings. However, I have some major comments, which I believe should be addressed before publication. In my review, I mostly focus on the Methods section, as requested by the editor.

General comments

• My main concern is that the machine-learning approach used to reconstruct the model fields is very different to the common methods (e.g., by Landschuetzer, and Gregor…) and thus, I am not convinced that the lessons learned from the authors’ approach can be directly translated to these methods. Specifically, the established mapping methods use training data from quite large regions (from a clustering step), which are a lot bigger than the region of this study. Thus, more data flows in, and they might be more robust to be able to reconstruct the seasonal cycle despite the sparsity in winter data. In addition, there are zonal differences and hot spots of in the Southern Ocean, and the subregion might not be representative for the Southern Ocean as a. whole. I do think we can still learn from this current study, but this issue should be discussed thoroughly. A follow-up study could later focus on the Southern Ocean as a whole (or even globally within e.g., the clusters by Landschuezter or Gregor et al.).
• I think it’s great that the authors use an ensemble of two ML-based approaches. However, I would appreciate a short analysis of how the two estimates differ, to understand how robust the findings are.
• The authors say that there is very limited data to allow for both training and testing/validation data. But how do the authors then know that the outcome is not overfitted? As the data is synthetic, could one not add more synthetic data that would then allow for both training and testing data?

Specific and minor comments to the text:

Introduction:

I found the introduction a bit misleading. After reading the introduction, I expected that the paper would include the interannual to decadal variability, but it “only” focuses on the seasonal cycle based on data from one year. Consider rephrasing this to not disappoint the reader.

Similarly, the introduction should mention clearly that this study “only” focuses on a subregion within the Southern Ocean.

Gloege et al. 2021 did an in-depth analysis of the uncertainty of ML-based mapping approaches, using synthetic data at a global scale. I think the introduction should mention that study and be explicit about how this current approach differs and what’s new about this study in comparison.

L.114: It’s mentioned later that how well the model matches the observations does not really matter in this context. However, please consider mentioning here already why using that model works (considering e.g., the Mongwe et al. 2018 study that showed how the CMIP models completely disagree on the phase and magnitude of the seasonal cycle).

L.196: Is this really the case? I would assume that any differences between the model and the observations could matter. I.e., the model might be generally a lot smoother than reality and thus the sampling strategy might be less sensitive in the model than in the real world. I would appreciate a short discussion on that.

L.335: I think the explanation of error and uncertainty is the wrong way round.