General Comments

Daniel Ford and coauthors use a feed-forward neural network (FNN) to estimate surface ocean partial pressure of CO₂ (pCO_2(sw)) in the South Atlantic Ocean. The authors test satellite chlorophyll a (Chl a), satellite-derived net primary production (NPP), and satellite-derived net community production (NCP) as biological predictors in the neural network to determine which produces the most accurate pCO_2(sw). They find that using satellite-derived NCP as a predictor in the FNN scheme produces the most reliable pCO_2(sw) reconstructions for the Amazon River plume and upwelling regions. They also show that, among the neural networks examined, the NCP-based FNN (SA-FNN_NCP) has the highest capacity for improved performance under scenarios of reduced uncertainty. For these reasons, the authors suggest that using satellite-derived NCP as a proxy for biological effects in surface reconstructions of pCO_2(sw) may be desirable going forward. Finally, Ford et al. find that SA-FNN_NCP indicates that the South Atlantic Ocean is a source of CO₂ to the atmosphere, whereas the FNNs with Chl a or NPP as a biological proxy or with no biological proxy all indicate that the South Atlantic Ocean is a CO₂ sink.

This manuscript fits well within the scope of Biogeosciences: it explores the implications of choosing different biological predictor variables in estimation schemes for sea surface pCO₂ and demonstrates the consequences of those choices for carbon cycling calculations. It is based on the very logical assumption that NCP, which captures all biological processes that modulate CO₂ concentrations in the surface ocean, should serve as a better biological predictor than Chl a for estimates of pCO_2(sw). This work has the potential to shift the way in which studies of this nature are typically performed. That shift could result in better representations of sea surface pCO₂ in regions that are highly influenced by biological processes, regions that may contribute a disproportionately large fraction of global CO₂ flux across the air–sea interface.

In general, the manuscript is well-written and the figures and tables are effective in communicating the results. The manuscript addresses an important aspect of the global carbon cycle and is forward-thinking in its assessment of algorithm performance in response to reduced uncertainties. A couple concerns of mine, however, include the lack of quantitative or graphical support for the conclusion that SA-FNN_NCP produces the best representations of pCO_2(sw) compared to the other FNNs (section 4.2) and the shortage of further investigation into one of the manuscript’s major conclusions: that SA-FNN_NCP flips the South Atlantic from a CO₂ sink to a CO₂ source. More discussion of these concerns as well as some minor comments can be found in the following sections.

Specific Comments

Performance of SA-FNN_NCP:

I mainly would like to see some quantitative or graphical evidence supporting the assertion that SA-FNN_NCP outperforms the other FNNs in the Amazon River plume and upwelling regions. Figure 3 shows differences between the mean climatologies given by different FNNs at different stations and Figure 4 shows that some of these differences are statistically significant (in comparison to SA-FNN_NCP), but neither says anything about the performance of any one FNN. That is left to the more qualitative discussion in section 4.2 that compares general patterns in pCO_2(sw) from previous studies to those indicated by the FNNs.

The points made in that qualitative discussion are compelling and certainly do appear to indicate superior performance of SA-FNN_NCP in the Amazon River plume, Benguela upwelling system, and equatorial regions. However, following along with the discussion takes some effort from the reader, and a lot of flipping back and forth between the text and Figure 3. A new figure comparing FNN results to some pCO_2(sw) observations or a brief presentation of some relevant statistics would be more compelling. In particular, for example, pCO_2(sw) data from the moorings at 6° S 10° W and/or 8° N 38° W could be plotted along with the SA-FNNs results to demonstrate the superior performance of SA-FNN_NCP.

I think this area is especially important to improve upon given that the bulk error statistics (Figures 2, A1, and A2) indicate SA-FNN_NCP to be the least accurate of the three FNNs that have biological predictors.

Sink to source transition:

The change in the cumulative regional sink from -7 Tg C yr^-1 with the NPP-based FNN (SA-FNN_NPP) to +14 Tg C yr^-1with SA-FNN_NCP seems rather drastic, and I’m curious to know more about why such a significant change occurs. The reason is not obvious from Figure 5 alone. If indeed the transition occurs because high outgassing events in biologically-controlled regions with relatively limited geographic extent are captured by SA-FNN_NCP but not the other FNNs, as is suggested in lines 399–412, that point should be demonstrated and emphasized more explicitly.

This could perhaps be explored by breaking down the annual fluxes into different sub-regions (e.g., the biogeochemical provinces from Figure 1) and/or into average monthly fluxes to clearly show the spatial and/or temporal differences that lead to the significant discrepancy between SA-FNN_NPP and SA-FNN_NCP. This information could be presented in a table, figure, or even just in the body of the manuscript (like the geographic comparison in lines 419–420 between SA-FNN_NCP and the Watson et al. [2020] product).

Callbacks in Discussion section:

In general, because the work is presented in separate Results and Discussion sections, I’d make sure to refer specifically to figures, tables, and statements from the Results section when commenting on them in the Discussion section. This will make it easier for the reader to follow what exactly is being discussed without having to determine for themselves where those results are presented. I mention a couple specific instances of this in the following section.

Minor Comments and Technical Corrections

Line 17: There shouldn’t be a comma after pCO_2(sw).

Lines 45–48: I’d split this sentence into two; there’s a lot of information here and it’s a bit difficult to follow as written.

Figure 1: Do the colors and/or symbols represent different things in this figure? If so, I would mention it in the caption. If not, they could all be the same color/symbol since they’re labeled with letters anyway. Additionally, can you add the AMT stations or the transect lines to this figure?

Table 1: What is the reference associated with the estimated uncertainty of 1 uatm in atmospheric pCO₂? Also, is it correct that the other uncertainty estimates all come from Ford et al. (2021)?

Line 189: I’d rephrase this as “A non-parametric Kruskal-Wallis was used to test for…”

Line 232: Should the accuracy here for SA-FNN_NCP be 21.68 matm, like in line 235 and Figure 2?

Line 238: Based on Table 2, should these numbers be 36%, 36%, and 20%?

Lines 259–260 (and elsewhere): Should be “minimum” instead of “minima” and “maximum” instead of “maxima”, or remove the article “a”. Minimum/maximum are singular whereas minima/maxima are plural.

Line 283: Add the accuracy of the SA-FNN_NCP after you mention it here, so it can be easily compared with Landschützer et al. (2013) and Landschützer et al. (2014).

Lines 285–286: You’ve already mentioned in the previous sentence that the SA-FNN_NCP approach has a similar accuracy to other approaches in the literature, so I’d say here that training the SA-FNN with Chl a or NPP gave comparable broad-scale accuracy to training it with NCP.

Line 295: “Satellite NCP is reliant on NPP as input” This has already been implied in line 288, so I’d remove the statement here or move it to the previous paragraph. The point is well made, it’s just that the writing is a bit repetitive here.

Line 298–299: “This showed that reducing in situ NCP uncertainties provided the greatest reduction in pCO_2(sw) RMSD, which was three times the reduction achievable using Chl a.” I’d make sure to refer the reader to Tables 2 and 3 so they’re not searching for this result.

Lines 328–329: “The stations (Fig. 1) represent locations of previous studies into in situ pCO_2(sw) variability in the South Atlantic Ocean and allow comparisons with literature values.” This point should be made earlier, in the Methods section, perhaps around line 182.

Lines 338–339: “Valerio et al. (2021) indicated pCO_2(sw) varied above and below pCO_2(atm) at 4° N 50° W consistent with the SA-FNN_NCP.” It looks to me in Figure 3 like pCO_2(sw) from SA-FNN_NCP remains almost exclusively below pCO_2(atm) at this site?

Line 396: Should be “may be” instead of “maybe”

This manuscript explored the importance of biological factors in modeling the seawater pCO2 via the inclusion of different biological proxies (chl-a, NPP and NCP) in the model construction. The results reveal a similar model performance with the inclusion of different biological components but in some cases, projected pco2 with the inclusion of NCP as a predictor works better. Overall, the paper emphasizes the significance of biological activity in controlling the magnitude and temporospatial pattern of sweater pCO2 and underscores the urgency to further improve the accuracy in NCP prediction. However, I have some major concerns regarding the model test and data display.



Major comments:



1) Lack of comprehensive evaluation of multiple NPP and NCP algorithms. The key goal of your study is to elucidate the role of biological proxies in modeling seawater. Also, the author concluded the accuracy of NCP simulation is critical.  If so, it is quite important to test the with the inclusion of commonly-used NPP and NCP algorithms (i.e, CbPM and VGPM-based NPP, NCP models by Li & Cassar, 2016; carbon export model by Laws et al., 2011 or (Henson et al., 2011)), rather than a single model output. Given that the focus is on the surface pCO2, I would say NCP model developed by Li and Cassar seems a better fit than model by Tilstone (2015) as the former provides mixed-layer integration of NCP rates and later is corresponding to euphotic-zone integration. At the very least, the author needs to clarify the reason why they chose the present NPP and NCP models (i.e., they are tuned by the Atlantic dataset and therefore they are supposed to do the best job in your study region). But I am still looking forward to seeing if other NPP and NCP products can further improve pco2 simulations.  



2) Unfair comparison between FNNbio and FNNNO_bio: The author used the model output without the inclusion of biological predictors (FNNNO_bio) as a reference to evaluate the improvement of FNNbio. However, FNNNO_bio was trained with inclusion of more physical parameters such as MLD, SST and salinity whereas FNN_bio just included SST.  To assure a fair comparison, it should keep other predictors consistency except for the inclusion or exclusion of biological parameters  



3) The overstatement of the performance of FNNbio_NCP: In many places, the author argued that FNNbio_NC did the best among the models (i.e., Line 235, 360 and 365). However, it is hard to distinguish the visible improvement from the Figure. For example, I don’t think the difference between 34% and 36% is significant enough to say NCP is much better than NPP given that uncertainty.







Minor comments:

1.p in pCO2 should be italic.



Line 15: it is hard to understand the method description regarding the “reducing uncertainty”.  Please clarify or say more specifically,



Line35 : clarify the full name of chl-a when it appears in the main text for the first time



Line 60: do you have a specific reason why you focus on the Southern Atlantic instead of the entire Atlantic basin or larger scale. You need some statement herein and also add some brief introduction about the setting of your study region.



Figure 1. Please add the data distribution (i.e., SOCAT and AMT cruise) you used for construction and validation of the model. It should be useful to add the name of the sub-regions in the map so the reader can easily navigate what subregions you are talking about in the following part. Probably you need two figures to get two issues done. Add the citation for the region division on the legend.



Line 80: I would suggest changing pco2(atm) to pco2_air, which is more straightforward.



Table 1: does log10 means log10 (chl-a). It is very confusing. The error of 45 mmol O2 m-2 d-1

 is very scaring because the typical range of NCP is from -50 and 50 mmol O2 m-2 d-1



Line 115: quite confused about what ropt means. Does it mean the optimal data numbers for the training dataset?



Line 125: The common practice for the data processing before training a neural network is to normalize the data to reduce the dynamics. Did you apply for it?



Line 126: please provide more details about the structure of FNN. How many layers and nodes do you set?



Line 145: Has pCO2 measurement in AMT cruise been included in the SOCAT dataset already?



Table 2: you probably can bin Table 1 and Table 2 to make it easier for reading.



Figure 2: the most useful information to evaluate the model performance is to look at the comparison with validation or independent dataset. I would suggest removing the figures of the training dataset and moving the validation results for models with inclusion of chla and NPP into the main texts.



Figure 2: what does the color stand for? I don’t think the blue line is useful. 1:1 line is enough. .



Line 2165: I have a very hard time understanding how you derived so-called maximum reduction. To achieve this goal, as I understand, you need to set a wide range of noise in the predictors (i.e., from -500% to 500%) and then find out the largest errors in pCO2 compared to the white control (zero noise in predictors)?



Figure 4: how did you calculate the seasonality? Do you mean the amplitude of pco2? You can consider using Taylor diagrams to demonstrate the inter-model comparison, which provide more information.



Line 235: I don’t think the difference between 34% and 36% is significant enough to say that the NCP is much better than NPP.



Line 360 and Line 365: I had a hard time detecting significant improvement in SA-FNNNCP compared to the other simulations.

Line 380: Can you find more straightforward evidence by examining the seasonality of chl-a, NPP and NCP products to support your argument regarding the disconnection?



Line 430: cite the corresponding Table and Figure when you talked about the result.



Line 440: it should be useful to plot a bar chart to display the annual co2 flux among the regions