This study produced a global GPP product (VODCA2GPP) using satellite-based vegetation optical depth (VOD) datasets from 1988-2020. Basically, the authors used the method proposed by Teubner et al. (2019), and evaluated the product accuracy based on the measurements of multiple eddy covariance towers globally. In addition, the authors compared the spatial and temporal differences of VODCA2GPP with MODIS and FLUXCOM GPP products. There are several main concerns in my mind to make me suggest rejecting the manuscript because of the important and inherent flaws of VOD data.

First, VOD indicates the vegetation biomass, a large carbon pool, but the GPP indicates one of the carbon fluxes, much smaller compared to biomass. Teubner et al. (2019) and this study used the changes of VOD to simulate GPP, which requires the high performance of VOD to indicate biomass changes, a huge challenge. GPP is sensitive to leaf area index or leaf biomass, not wood and root biomass. However, leaf biomass is quite small fraction compared to other two components. It is obviously using VOD to estimate GPP has much large uncertainty. From the validation results, we found low model performance of VODCA2GPP shown as fig. 3a. VODCA2GPP only reproduce spatial variations of GPP about 25% (pearson r = 0.515), and I can’t trust the information provided by VODCA2GPP as such low performance. In addition, limited by VOD products, the spatial resolution of VODCA2GPP is only 0.25°, which is not far enough for scientific applications, as MODIS Landsat GPP has reached to 30 meter spatial resolution.

Second, I am confused to model algorithms proposed by this study. The algorithms of this study look different a little bit with those of Teubner et al. (2019). Equ. 3.2 represents the model algorithms, the authors added the third term the temporal median of VOD (mdn(VOD)) derived from the complete time series which serves as a proxy for the landcover. Unfortunately, the authors did not explain how mdn(VOD) can indicate landcover. And I also concern why the authors need this term for simulating GPP. Besides, the authors separated GPP into Ra and NPP components, so the model can provide the simulations of Ra (as least maintenance respiration) and NPP respectively, which may help us to judge if the structure of model is reliable. However, the manuscript also missed this kind important information. Another important model flaw, if the changes of biomass is equal to GPP. Obviously, the answer is NO! When the biomass lost partly at the 0.25 degree pixel, the GPP should be positive assuming the rest vegetation in this pixel still live, but lost biomass may larger than actual GPP contributed by the rest vegetation. Then, VODCA2GPP product may produce negative GPP.

Third, as the dataset description paper, I majorly pay much more attention to validation of dataset. The authors actually conducted very weak model validations, only validate the performance for reproducing spatial variations of GPP (low performance as the above comment). I would like to see if the product can reproduce the seasonal and interannual variations of GPP against the measurements of eddy covariance towers. In addition, the authors seem to hide the magnitude of global VODCA2GPP GPP. From fig. 6, VODCA2GPP estimates for global mean GPP intensity is almost twice than that of MODIS. If I am correct, the global GPP derived from MODIS is 108 Pg C yr-1, thus the global magnitude derived from VODCA2GPP may reach to 200 Pg C or larger? This is a new and the maximum estimates of global GPP that I ever knew. The authors need provide enough robust evidences to prove this number is correct, else I can’t trust the VODCA2GPP, which will lead us to a wrong way.

Forth, I felt disappointed and the writing is very poor. There are some low-level mistakes in writing and format. All sections of the manuscript look very rough. Especially, the introduction and discussion are meaningless mostly, and the authors kept some sentences to repeat rather than discussing important scientific questions. Several paragraphs are very short consisting of only 2-3 sentences, which failed to provide useful information. Feeling so bad when I was reading.

Some minor comments.

line 49-54: this paragraph is not necessary.

it seems the subtitle of “4. Results” missed.

There are no GPP values over the desert area in fig. 1c, but there are uncertainties of GPP in fig. 2. These two maps should keep constant for non-vegetated area.

Fig. 3 shows the poor performance of VODCA2GPP compared to fluxcom and modis products. So why do we need VODCA2GPP product?

4.2 section. The authors only compared temporal variability of VODCA2GPP with modis and fluxcom datasets. How about VODCA2GPP to reproduce the interannual variability of GPP compared to fluxnet observations. Actually, the comparison with fluxnet observations is more important for this data description paper to let the readers and potential users to know if VODCA2GPP can reproduce the long-term or interannual variability of GPP.

4.3 section. Why did the authors add the trendy simulations here, not compared previously?

4.4 section. Totally confused. Sometimes, the authors made comparison from 2002 to 2016, and otherwise 2003-2015? Can not understand the purpose of the authors.

Line 396-399: can’t accept this explanation for poor model performance. The sparse observations in arid and semi-arid regions will reduce the performance of modis and fluxcom, but VODCA2GPP showed the obvious bias.

5.2 section. What is the objective of this section? I did not get any meaning points. The manuscript seems to be rough at the initial stage, and the authors need much more work to improve the writing.

5.5 section. I can’t agree to this point. There are many factors may enhance the global gpp, how the author can claim the increasing trend imply the CO2 fertilization? There are many conclusions that the authors reached without robust evidences. By the way, this section is basically repeat of results.

5.6 section. Again, a meaningless paragraph and section. It looks like a conclusion or summary.

General comments:

This study introduced a new global GPP dataset using satellite-derived vegetation optical depth (VOD) dataset spanning 1988-2020 in spatial resolution of 0.25^o. The method for relating GPP to VOD was proposed in Teubner et al (2019, 2021). This study also presented validation against FLUXNET observation and other global GPP datasets (MODIS, FLUXCOM, TRENDY models). The difference of spatial and temporal GPP dynamics from other GPP reference datasets were shown. However, given the lower VOD-GPP model performance and extremely high global amount of VODCA2GPP, I do have concern about the quality of this dataset and its advantage over other datasets. The writing also needs a lot of improvements since there are both phrasing mistakes and technical mistakes.

Specific comments:

1 The global total amount of VODCA2GPP datasets is not given.

2 line 19, line31, it should be ‘at global scale’.

3 line 32, it should be ‘different’.

4 line 32-33, it should be ‘similar significant increase’. Please check the usage of two ‘for’.

5 line 43-44, please reformulate it.

6 line 46-47, please reformulate it.

7 line 51, please use clause.

8 line 53, what did you mean by ‘measurement bias’? Is it affected by the distribution of flux towers or just measurement techniques?

9 line 236, please reformulate it.

10 line 245, for site-level validation, you should also use cross validation method.

11 line 264, the subtitle for section 4 is missing.

12 In figure 2, the uncertainty is very high in desert area (more than ten time of the GPP itself, why?), for example, northern Africa, which is blank in figure 1. Please use single-hue color series and check to ensure the consistency with figure 1.

13 In figure 3, please check the unit in both x and y axis. Please use dashed line for 1:1 line. Here, you presented the site-level evaluation for mean annual GPP. What about the annual, seasonal (or monthly) evaluation? In fact, Pearson value of around 0.5 seems to be relatively lower for GPP estimation. Did you try 10-fold cross validation for model performance?

14 line 308, when you mention ‘good temporal agreement’, you need to specify the region, since the correlation is even negative in tropical America.

15 line 308-309, it should be ‘the highest’ and ‘the lowest’.

16 line 311, please rephrase it.

17 line 313, why monthly GPP anomalies are compared? Why you did not use monthly GPP instead?

18 line 334, here you used 2002-2016. Why did you use 2003-2015 in next paragraph?

19 line 335, there are only three sentences in this paragraph. It would be better to merge it with next paragraph.

20 In figure 6, at global scale, the VODCA2GPP is two times of the MODIS and FLUXCOM. So the global GPP can be more than 200 PgC yr^-1, which is the highest estimation as far as I know. You should give reasonable evidence for such value, or it won’t be convincing.

21 In figure 7, it should be 2002-2016 or 2003-2015?

22 In section 5.2, TRENDY ensemble is indeed independent from VODCA2GPP. But I couldn’t get the meaning of this paragraph. The process-based estimation without accounting for nutrient constraints is also reported to overestimate GPP.  

23 In section 5.6, in terms of the VODCA2GPP datasets, more similarities with TRENDY than that of other observational products were mentioned. Is it really a strong point for this GPP dataset?  

24 In figure A1, please check the unit in both x and y axis. You can give annual, seasonal (or monthly) evaluation as well.

25 The code for producing the data and figures should also be publicly available in a repository for transparency reason.