Kang et al. developed a new machine learning based global GPP dataset (CEDAR-GPP), which incorporated CO2 fertilization effect (CFE). Since the CFE significantly affected the trend of global GPP, the new dataset can answer the question of how the CFE benefits GPP over the last 40 years. This dataset overcomes the previous dataset without consider the direct CFE to GPP (i.e. FLUXCOM, FLUXSAT), so it could have the potential for evaluating the GPP increasing trend caused by the CFE. However, I found serval main issues in the current manuscript and dataset.  

Main point#1 Since there are ten different setups for CEDAR-GPP, so which one is best among them? Or can the authors provide a guidance for the readers use this product to address the specific question? This is a particularly important point for the dataset availability. As the results mostly showing the NT data, why not move all the DT results to the SI?

Main point#2 The authors claimed that they consider the direct CFE to photosynthesis. If my understanding is right, the direct CFE benefits the GPP over the past 4 decades and the CFE-ML based GPP (consider the direct CFE) will have a higher increasing trend compared to the Baseline based GPP. The baseline GPP did not contain the direct CFE, but it has the trend of VI, which can represent for the in-direct CFE trend. The reason I have this question is, as the machine learning model is black box, the trend is based on statistic relationships, so the output for the direct CFE or indirect CFE may just depend on the model training. This means that the Baseline GPP could also capture the indirect CFE to GPP and the indirect CFE can reflect the GPP trend in the real world. Without the site validation, the GPP trend from CEDAR-GPP is hard to distinguish whether they are right or not.

 Main point#3 The validation for the GPP product is inadequate, especially the GPP trend. It should have a specific validation for ten setups at Monthly, Seasonal, Annual trend for different vegetation types or climate zones. Therefore, the readers choose the specific region with high accuracy data they needed. I am not doubting that the global GPP has continuously increased during the past 40 years. However, different regions may have different greening/browning patterns, so what is their contribution to the global GPP trend? The spatial continuously product can answer this, so the author should first provide the GPP trend at sites and proved they are right, and then upscale this result to the global. Therefore, the site validation is very important!

 Main point #4 The annual trend should be evaluated on the long-term reliable observation data. Although the authors used the site GPP with more than 5 years for validation by the method provided by Chen et al. 2022 PNAS. However, the ICOS and the FLUXNET2015 has the longer-term observation data, so the author can evaluate the long-term trend (>10 yr) from these sites. Sometimes, the earlier year GPP data will have more data missing during the year, which underestimated the annual GPP at sites, so the annual trend is not reliable when only using 5 years of GPP observations.

 Main point #5 Cross validation among the GPP products should have more discussions. Although the authors did some comparison between the CEDAR-GPP with existed products, they just showed the results, but did not analyze the results and provide the insight on which product can be improved or what is the inherit reason for the uncertainties. So the discussion section can list some viewpoints for the further GPP product improvement.

Main point #6 The water stress effect should be considered. Line 718 to L720, “Yet the model assumed a fixed ratio of leaf-internal to ambient CO2, and thus did not include any responses to vapor pressure deficit.”. So is the CEDAR-GPP consider VPD effect for GPP? Although CFE is one of the most significant effects to the GPP trend. However, the VPD trend cannot be omitted. Li et al. 2023 reported that the VPD significantly affects the GPP at different vegetation types, so the CEDAR-GPP consider this or some setups reproduce such condition? If the CEDAR-GPP cannot reproduce the VPD effect, so under what condition can I use this product? The soil moisture is also coupled with VPD, so is the CEDAR-GPP consider the soil moisture stress, and can it reproduce the stress from soil moisture? These are needed to be mentioned.

Li, S., Wang, G., Zhu, C., Lu, J., Ullah, W., Hagan, D. F. T., ... & Peng, J. (2023). Vegetation growth due to CO2 fertilization is threatened by increasing vapor pressure deficit. Journal of Hydrology, 619, 129292.

Specific comments:

L14 sugars and starches->carbohydrate

L26 the annual trend validation with statistics should be highlighted since it is the most important outcome for CFE induced GPP trend.

L38 references needed

L90 ‘yet, this important mechanism is still missing in GPP products upscale from in situ eddy

covariance flux measurements.’ This is not accurate, since Zheng et al. also parameterize the model by in situ eddy covariance flux measurements. So this sentence should be revised.

L125 So totally 233 sites are used but not listed, the authors should list them at the SI.

L137 The C3 and C4 map is constant among the investigated years or not? Will the C3/C4 vegetation fraction change?

Table 1 ‘Surface reflectance b1 – b7, Vegetation indices (NIRv, NDVI, kNDVI, EVI, GCI, NDWI), percent snow’. What is GCI?

L 158 vegetation indexes or vegetation indices?

L172 ‘PKU GIMMS LAI4g consisted of AVHRR-based LAI from 1982 to 2003 (generated using machine learning models trained with Landsat-based LAI data and NDVI4g) and MODIS BNU LAI from 2004 onwards (Yuan et al., 2011).’ To my knowledge, PKU GIMMS LAI4g didn’t use the Yuan et al 2011 as input data.

L210 ‘GIMMS LAI4g and NDVI4g data were only filled with mean seasonal cycle due to their low temporal resolution (bimonthly).’ This is unclear, as the CEDAR-GPP with the minimum temporal resolution is at monthly, why the bimonthly data just filled with mean seasonal cycle?

Figure 2 The CO2 should be separate at the left. Besides, I cannot understand why the CFE-ML is omit in the long-term product.

L236 should be section 2.3.2

L238- L241 If the CFE-ML model is well trained by the LAI4g and other input data during 2001-2020, it could also be extent 1982-2000. So why missing the flux data before 2000 hinder the CFE-ML processing? The CFE-ML before 2000 should also be compared to other products.

L249 ’ do not consider the direct effect of CO2 on light use efficiency’, this is inaccurate, the VI sometimes can capture the LUE change from direct CFE.

L257-271 an open question here, I read the EEO theory from Chen et al. 2022 PNAS. The authors may refer to the SI from that paper, you may see the EEO theory cannot reproduce the GPP trend at tropical rain forests (GF-Guy). So whether the CFE-hybrid can capture the actual trend of GPP, more validation is needed.

L325 To my knowledge, the FLUXCOM has removed the annual trend which has been pointed out by the data use guideline, so why used the FLUXCOM data?

Figure3 it should have systematic validation for ten setups but not only validate the CFE-Hybrid_NT. Besides, the GPP anomaly seems to be underestimated compared to the site observation?

Figure5 the trend should be analyzed among PFT and climate zone similar to figure4. More importantly, the sites showed a higher GPP trend than the models. Do you mean the CEDAR-GPP underestimate the GPP trend? Why the trend in CFE-ML-NT is much higher than it in CFE-ML-DT?

Figure7 The standard deviation of GPP at different months should be added to the figure.

Figure8 The FLUXCOM data cannot be applied to evaluate the IAV as the product guideline has pointed out. The baseline and CFE-ML product should be listed here.

Figure11 Can I infer the GPP trend from 2001 to 2018 is double compared to it from 1982 to 2000? Is there any existed GPP product (i.e. machine learning or TRENDY data) can support this result?

Function A2 𝐴 = 𝐴C = 𝐴j? I cannot agree with this! There are just one condition Ac = Aj is the photosynthesis transfer from lighted-limited to nutrient limited.

General comments:

This manuscript presents a new global GPP dataset by using the machine learning method and many various datasets. They mainly considered the direct CO₂ fertilization effect on GPP through both the machine learning and the theoretical approaches. They then modelled the short and long-term global GPP by using different periods of earth observations and climate variables. They also validated their new GPP data by comparing to flux sites GPP data and other products.

This study fits well for ESSD and I appreciate the efforts that the authors made. Several minor concerns from my side are listed below, which I think may be helpful for revising the manuscript.

Specific comments:

Line 20: Maybe not the “first”, since you also said that the revised EC-LUE GPP have considered this.

Line 114: Some flux sites have the data before 2001? Why did include them into the model? And, this means that the information before 2001 from flux sites are not included into your GPP products, right?

Line 214: Both the MODIS and CISF data are used in the model. Will them from the same observations, i.e., the CSIF is also be trained from MODIS data right. Will the multi-collinearity issue exist in this?

Line 252-256: When adding the CO₂ into the machine-learning model, the output of your model is GPP, but LUE. That means, the model could not distinguish the direct CO2 fertilization on LUE and the indirect CO₂ impact on FAPAR? A simple way to check this is to check the CO₂ impacts on GPP, LUE and FAPAR at some FACE sites, and to check that if your model is correct.

Line 261-262: A constant 𝜒 value for global and long-term analysis seems will introduce large uncertainties. Since we know that 𝜒 will change with different climate conditions, and also will change largely with the CO₂ concentration. Although we know that the modelling of global 𝜒 is very hard, but some discussions and limitations on this aspect is at least needed.

Sincerely,

Songhan Wang, PhD

In this manuscript, Kang and colleagues present a novel GPP product that incorporates the CO2 fertilization effect. After several thorough readings over the past two weeks, it has become evident to me that this manuscript is both timely and well-prepared. I have some minor suggestions to enhance its clarity and impact:

1. In the Introduction section, it would be beneficial to provide a more comprehensive explanation of why the authors chose scaling as their approach for global GPP development. Given that there are various other GPP development methods, such as LUE models and SIF retrievals, it's essential to highlight the unique advantages of scaling in comparison to these alternatives.

2. The concept of "direct CO2 fertilization effect" may not be familiar to all readers. It would be helpful if the authors could provide a clearer explanation of this term and include relevant references in the introduction to ensure a better understanding among readers.

3. Figure 1 effectively illustrates the spatial distribution of flux tower data used in this study.  it would be more beneficial to indicate how these flux tower sites represent different biomes. I would suggest adding a Whittaker biome figure within Figure 1 to visually depict the biome types associated with the flux tower locations used in the study.

4. Given the potential wide interest in the various GPP products developed in this work, it would be very useful to include a section that provides guidance on how future researchers and users can effectively utilize these datasets in the future. This could include tips on data access, processing, and interpretation to facilitate broader adoption.

5. To streamline the manuscript, consider moving some supplementary materials, such as Table 1 and Table 3, to the supplementary section. This would help maintain a concise and focused main manuscript while still providing access to important supporting information.