For the past decades, NDVI has been one of the most frequently used and verified remote sensing indices to monitor vegetation dynamics at regional and global scales, with its unique advantages including the simple form, a low sensor requirement, robustness to changes in plant canopies, insensitiveness to atmospheric effects and the sun-sensor geometry, etc. The developments of global long-term NDVI products have thus been critical. The manuscript presented a new version of GIMMS NDVI products (denominated as PKU GIMMS NDVI in the manuscript), (I) to address the issues of orbital drift and sensor degradation effects in AVHRR data using massive Landsat samples and (II) to extend the temporal coverage to recent using MODIS NDVI. All parts of the manuscript were well organized and written. The methodology is feasible and reliable to me, and the results provide relatively comprehensive evidence on the spatial and temporal accuracies of the products. The product could potentially benefit many future vegetation studies. I have an overall positive attitude on the product and basically recommend the manuscript for publication, as long as the authors address the following concerns.

My major concern is on the data consolidation with MODIS NDVI. I am happy that a pixel-scale linear fusion method was adopted, as the relationship between AVHRR NDVI and MODIS NDVI could be spatially different. However, the relationship might also be driven by some other factors such as the plant function type, phenological cycle, and sensor design, which I believe are beyond the capacity of the simple linear function used in the study (Mao et al., 2012). This could be confirmed by the failure of the method in some EBF (Figure 10). Also, the phrases ‘before consolidation’ and ‘after consolidation’ have frequently confused me when reading. To me, the data before consolidation are an intermediate product that certainly needs a discussion but is improper and unnecessary to compare with other NDVI products, for example, in trend analysis. It is not an individual product.

The second issue is related to NDVI products used for comparison. I have noticed that only GIMMS NDVI3g was involved; meanwhile there exists some other global long-term NDVI products such as the LTDR4 and VIP3 (as the authors also mentioned in Introduction). So why were they excluded in the data comparison? Please clarify.

Last, the data published in ZENODO (https://zenodo.org/record/7441559#.Y7J7y3ZByCo) have not been well structured and the data size is way larger than the analogs. I suggest the authors, if possible, compress the data and stack the monthly NDVI to preferably every 10 years before uploading.

Specific comments

line 17-19

Better specify how were R2, MAE, and MAPE calculated.

Line 44-46

This is particularly severe when estimating long-term changes (e.g., Shen et al 2022).

Shen et al 2022, Plant phenology changes and drivers on the Qinghai–Tibetan Plateau. Nature Reviews Earth & Environment, 3, 633–651.

Line 126-127

Could the authors provide more details here?

Line 149

How was the spatial aggregation used? By highest frequency?

Line 196-197

would 9 30m-resolution landsat pixels represent 8km GIMMS pixel?

and was this done separately for each time step? How was the data temporally matched?

Or did I miss something here?

Line 201 section 3.2.2

was there any procedure performed to ensure the data quality of the two NDVI datasets?

Line 203-204

were all the currently available landsat data included here? or just for the sample locations?

Line 207-208

Please consider if it is necessary to build more than one models for a given biome due to large spatial variability within the same biome (not mandatory).

Line 219

2016 or 2015?

Line 220-222

But some studies have shown that modis VI also has quality issues due to, such as, sensor degradation.

Line 226

Which is the dependent variable here? MODIS NDVI or GIMMS NDVI?

Line 228-229

May be a little more detail about the spatial and temporal matching between modis and pku gimms ndvi would help.

And was there any procedure performed to ensure the data quality of the two NDVI datasets?

Line 229

2004-2015 or 2001-2015?

Line 252

does landsat have an orbit shift problem?

Figure 5

Please delete ‘unitless’ in the figure.

This study proposes PKU GIMMS NDVI, a new global long-term NDVI time series data that covers 1982 to 2020 based on AVHRR and MODIS sensors onboard satellite platforms. The PKU GIMMS NDVI extends the GIMMS NDVI3g data and has better data quality. It has better agreement with Landsat NDVI compared to GIMMS NDVI3g, alleviating the orbital drift problem in the AVHRR sensors. The method proposed in this study could be used to generate consistent global NDVI data in the future, which would help study global terrestrial biosphere dynamics.

Detailed comments:

Line 14: “global-wide” may be simplified as “global”

Line 26: When introducing NDVI, please cite the original NDVI paper:

Rouse, J.W., Haas, R.H., Schell, J.A. and Deering, D.W., 1974. Monitoring vegetation systems in the Great Plains with ERTS. NASA Spec. Publ, 351(1), p.309.

Line 97: Please provide literature references to the data sources.

Line 126: How is the time-weighted aggregation performed? Please explain in detail.

Line 127: Maybe the authors want to say “upscale” instead of “downscale.”

Line 139: How were the Landsat NDVI samples aggregated to 1/12˚? Please explain in detail.

Line 144: “temporal” should be “temporally.”

Line 226: could the authors elaborate more on how they spliced the PKU GIMMS NDVI and MODIS NDVI?

Line 254: How can seasonal fluctuations in the time series of NDVI bias be removed via the multi-year averaging method? Please explain.

Line 257: Maybe it should be “… was evaluated at 1,000 random points …”

Figure 2: maybe the authors could also show the regression line and equation in each panel?

Line 301: The section title could be “Validation of PKU GIMMS NDVI and GIMMS NDVI3g”

Figure 6: please explain in the figure caption how the R2 was computed in detail.

Line 337: The section title could be “Comparison with MODIS NDVI”？

Overall Comments:

A relaiable long-term vegetation time series is critical to understand the dynamic of vegeataion and its feedback to the climate. This study by Li et al reconstructs a spatiotemporally consistent global NDVI dataset for 1982-2020 integrating Back Propagation Neural Network and a total of 3.6 million Landsat NDVI samples that well spread across the globe as input. This product, along with its predecessor (the GIMMS NDVI3g dataset), has been evaluated with the Landsat NDVI samples, showing substantial improvement. This study has originality and significance in uniqueness and usefulness.

Below are some comments that may help to further improve the manuscript. First, it seems that the golden truth of NDVI is Landsat NDVI samples. I recommend providing details in section 3.2.1 of Landsat NDVI samples to illustraluate: 1) Why Landsat NDVI is more accurate than other products? 2) Does Landsat NDVI have any limitations (e.g. the influence of clouds)? 3) Adding a plot to show the distribution of these samples (time and space).

Second, this dataset extends the time-span of 1982-2015 for its predecessor to 1982-2020, but there is no figure or analysis of this extension. It would be great to show more details of this extension (e.g. a long time series spanning from 1982 to 2015). In addition, it would be great to explain the results of consolidating the PKU dataset with MODIS NDVI data for the years after 2016. A comparison between MODIS and PKU NDVI datasets from 2017 to 2020 would be helpful (e.g. Figure 12).

Some minor comments: The Landsat NDVI samples were used both to train and validate the PKU GIMMS NDVI dataset. How the Landsat NDVI samples were seperated into two groups?

The figure caption for Figure 6 is confusing: comparison of R2 between the GIMMS NDVI4g and PKU GIMMS NDVI products. Is it R2 between the GIMMS NDVI4g and PKU GIMMS NDVI products? Or R2 between the GIMMS NDVI4g and Landsat NDVI, and that between PKU GIMMS NDVI and Landsat NDVI?

Why MOD13C1, not MOD13Q1, MOD13A3 or MOD13C2, is used in this study?

Figure 11: In the PKU dataset, most tropical regions show greening during 2004-2015, which may be unreasonable.

Figure 4: S5 shows no significant improvement compared to S4.

Figure 8: A color bar is missing.