Review of the Paper: "Empirical upscaling of OzFlux eddy covariance for high-resolution monitoring of terrestrial carbon uptake in Australia"

The paper develops high-resolution estimates of GPP, ER, and NEE in Australia using empirical upscaling of flux tower measurements. Comparisons with other products show regional empirical upscaling outperforms global upscaling and process-based models. Rainfall deficits and surpluses drive NEE anomalies, with GPP responding more than ER. The paper introduces "AusEFlux" as a benchmark for high-resolution monitoring of Australia's carbon cycle.

Upon careful evaluation and analysis of the manuscript, it is evident that a major revision is necessary in light of the following critical comments. Addressing these concerns will enhance the overall quality and impact of the paper, ensuring its suitability for publication in our esteemed journal.

• The paper highlights the performance of regional empirical upscaling in improving global upscaling products and outperforming existing LSMs, but could you provide more insight into the specific limitations of Australia's comparatively sparse network of EC towers and their potential impact on the accuracy of the derived estimates?
• How were datasets resampled to monthly resolution and reprojected to 1*1 km in section “2.1.2 Gridded explanatory variables”? what was the raw data specifications?
• Provide additional details or references to explain the 'SOLO' data version used for partitioning NEE into GPP and ER. This will aid readers in understanding the specific data processing steps and methods employed.
• In Section 2.1.3, it is mentioned that the MODIS-GPP and DIFFUSE-GPP products were resampled to a 1 km resolution to match the resolutions of the ML upscaling product. Could you please provide more details regarding the specific method used for resampling these datasets? It would be beneficial to understand the resampling technique employed to ensure compatibility between different resolutions. Additionally, any information regarding potential implications or limitations of the resampling process would be valuable.
• Elaborate on the resampling and reprojection of gridded explanatory variables. Specify the resampling resolution and provide a rationale for selecting a common 1-km x 1-km geographic grid. Discuss potential errors or limitations associated with spatial resampling and its impact on the accuracy or comparability of the datasets.
• In Section 2.1.3.3, it is mentioned that the regional inverse modeling product by Villalobos et al. (2022) provides a spatial resolution of approximately 81 km. Could you please provide details on how the other datasets with different resolutions were processed and plotted to ensure compatibility for comparison? Specifically, how were the ML results, MODIS-GPP and DIFFUSE-GPP products, which were resampled to 1 km resolution, handled in the analysis?
• Specify a specific website or source where readers can access the CO2 flux tower data used in the study. This will facilitate replication and further exploration of the data.
• Provide more details on the implementation of random forest regression and gradient-boosting decision tree algorithms, including parameter settings, and more importantly elaborate on how predictions from the ensemble of random forest and GBDT models are combined or weighted as an ensemble learning.
• Clarify the rationale and details behind the iterative training procedure with randomly selected EC sites for uncertainty estimation. How can you ensure that all sites were removed in the 30 repeats? Provide details on how the randomness is controlled to achieve this objective.
• Provide a detailed description of the data split methodology used in the nested, time-series-split cross-validation approach. Did you consider the aspect of time when splitting and testing the methods? (e.g. did you allocate 5 years for training and 1 or 2 years for testing?)
• Consider incorporating any additional limitations or uncertainties associated with the data sources, processing steps, or comparison datasets. This will provide a more comprehensive understanding of the potential impacts on the study's results and conclusions.
• Why was the model not tested on individual sites after training? It is crucial to determine whether the model can perform effectively at a single location.

I was excited to receive the invitation to review this manuscript, as it is a much needed piece of work that contributes to improved understanding of Australia's terrestrial carbon cycle. It's great to see the use of the extensive OzFlux dataset for validation of the new AusEFlux product, as it is no trivial effort to keep sites running, collate and share eddy covariance data on a regular basis for use in these types of studies. The manuscript itself was very well written and easy to follow, with nicely designed figures and tables (I particularly loved figure 8!) - I applaude the authors on these aspects of their manuscript. 

While I enjoyed reading this manuscript and feel it is an important contribution to the research community, I found the discussion section was very limited. First, I was looking for more critique on how the empirical upscaling approach has improved on previous methods, including a well-articulated argument for why regional empirical upscaling needs to be considered by the modelling community to ensure regions are correctly represented in global estimates. There is a heavy bias of flux sites and model parameterization from the temperate northern hemisphere, with regions such as Australia, South America and Africa lacking on-ground validation sites to adequately verify global models. What I really like about this study is that it has elegantly shown there are regional differences in Australia not being adequately captured by global models. If this is true for Australia, surely it could be true for other regions less represented too. I'd like to see more thought and critique on this point in the discussion. 

There's also a lack of discussion of how the limitations in this study could be overcome. For example, the method itself seems sound, but expanding on the points about more EC data being needed would be really helpful to the research community. Do the authors feel that just longer timeseries from the current network are needed, or are more sites required? If more sites, where are they needed? Figure 8 showed some interesting GPP and ER dynamics in certain areas where there is a distinct lack of EC observation sites, the WA Wheatbelt being one of them. The authors point to these areas in the results section, but do not address how these areas could be better understood by future research efforts. A better discussion around these points, in a nationally focused paper like this, would really help researchers on the ground level to make the case for the need to fill these missing gaps. 

Lastly, there's a lack of discussion around future directions for this work. There's momentum building and wider interest in understanding carbon fluxes from landscapes in real time and in collating annual budgets at national scale more frequently. There's an opportunity for the work presented in this manuscript to be incorporated into a regularly produced national annual estimate of Australia's terrestrial carbon accounting, but there's no mention of this in the discussion. I suggest the authors consider adding text along these lines, perhaps pointing to the vision outlined in Papale 2020 and efforts already underway to deliver national carbon observing infrastructure, such as TERN, NEON (USA) and ICOS (EU), that could be input into approaches like the one presented by the authors to help realise these goals. 

There are a few other specific items I feel need to be addressed before the manuscript is ready for publication. I've identified these as follows and believe that if the authors can address them, their manuscript will be more widely cited as a result: 

- Lines 37-45: This is quite a difference between the two studies, but then at line 55 it's revealed that the Villalobos et al. 2022 study was from years 2015-2019, while the Friedlingstein et al. 2022 study was from 2003-2021. Looking up both studies reveals these time frames to be accurate. While I completely agree that regionally forced studies usually provide more accurate estimates of carbon cycling in Australia, I think it is misleading not to mention the temporal mismatch between these studies. I suspect the temporal mismatch could be the primary cause of the difference of >50 % between studies, as the millennium drought (2001-2009) would be captured in the Friedlingstein et al. anaylsis but not in the Villalobos et al analysis. Please amend the text to take this into consideration. 

- Line 57; Please add spatial resolution for better comparison with OCO-2, i.e. as at lines 54-55. 

- Line 70: I think its important to identify here the unequal representation of EC sites across the globe, as some biomes (i.e. the tropics) contain a limited number of sites compared to the temperate northern hemisphere. This bias is also likely to be affecting ML empirical upscaling approaches. I see the authors allude to this at line 78, but I think it needs to be addressed here too. See Baldocchi et al. 2018 (https://doi.org/10.1016/j.agrformet.2017.05.015) for a good review of inter-annual variability in NEE from sites around the world, and where long-term monitoring sites are lacking.

- Line 77: A better introduction citation for the OzFlux network is Beringer et al. 2022 ( https://doi.org/10.1111/gcb.16141) or Beringer et al. 2016 (https://doi.org/10.5194/bg-13-5895-2016). Isaac et al. 2017 is an excellent publication to cite for how the flux data were processed, which should be in the methods. 

- Line 108: Please add the following text here to clarify how the data were processed "using PyFluxPro vXXX (Isaac et al. 2017),..." The authors may need to check with TERN regarding the PyFluxPro version used. 

- Lines 189-191: Can the authors comment on this more specifically? Are there any biomes or land uses missing that in their opinion would make the analysis more robust? Perhaps this could come in the discussion instead...?

- Lines 321-325: I agree with this statement, but it should appear in the discussion, not results. Please move to discussion, a good place would be the final discussion paragraph. 

- Lines 406-422: This paragraph is mixing results and discussion a bit, i.e. lines 408-410 and lines 415-416, Please consider moving these points to the discussion, which would help beef up the section. 

- Line 438: Remind readers of this study here, it's the Villalobos et al. 2022 study, correct? In fact, it would be useful to include a small table that includes information about each of the models used in this study, who published them, their general characteristics (temporal and spatial resolution), etc... That way the authors can refer the reader here to table X for a refresh and avoid re-citing each study, that would add clutter to the text below. The table could be a brief summary of information presented in section 2.1 and included at the end of that section. 

- Lines 438-442 - this is all one sentence, which is long and rather confusing to follow. Please revise and more clearly articulate to the reader that this study was verified using OzFlux EC sites. 

- Line 460: Table shouldn't appear in the discussion. 

- Lines 467-469: Can the authors expand on this point more?  In an ideal world, how frequently do the authors think a product like this should be updated? Realistically, how frequently is this likely to be? I recommend reading Papale 2020 (https://doi.org/10.5194/bg-17-5587-2020) and publications from the global carbon project to tease this discussion point out further. 

- Lines 477-478: What about the role of fire in consuming biomass in the dry season and how that might affect carbon emissions from savannas? Can the authors expand on this please. Beringer et al. 2015 ( https://doi.org/10.1111/gcb.12686) might be a good place to start. 

- Lines 478-480: Here again, a missed opportunity to critique with site-based studies, such as Cleverly et al. 2013 (10.1002/jgrg.20101) 

- Lines 489-492: This is a rather subjective and negative way to begin a conclusion. One could argue that the OzFlux network already captures a diverse range of Australian ecosystems, and it is certainly the largest network in the underrepresented southern hemisphere. However, one could also argue that there are key systems missing, which can bias any upscaling approaches that use OzFlux data. How many flux towers are needed for a network like OzFlux to have "good" coverage? My point being, given this paper did not assess whether the quantity of sites in OzFlux was adequate for upscaling (in fact it used correlation with OzFlux sites as an indicator that the results were robust), I suggest rethinking the opening sentence of the conclusion to be more focused on the key result/finding and less about the limitations of OzFlux.