General comments:

Bias in MODIS VCF is a known issue. The study by Adzhar et al. adds new interpretation to the issue by comparing the latest version (Collection 6) of the satellite-based dataset with field data collected at ecological sites (24 forest and 24 savanna sites). The authors developed a simulation technique to address the scale discrepancy between the 100m x 100m sites and the 250m x 250m MODIS pixels, and extrapolated the derived relationship across the tropics. The study went on and analyzed the bias patterns of MODIS VCF per land-cover class using the MODIS land cover product. A major finding of the study is that MODIS VCF underestimates tree cover in tropical savannas and woody savannas, which has important implications for carbon cycle and forest restoration potentials. The paper is well written and easy to follow.

Specific comments:

Multiple definitions of “tree cover” exist in the literature. The manuscript provides a clear definition of Canopy Area Index (CAI, the fraction of ground covered by tree crowns). A primary semantic difference between CAI, and percent tree cover of MODIS VCF (the portion of skylight orthogonal to the surface that is intercepted by trees), is whether within-crown gaps are taken into account (CAI) or not (VCF). Within-crown gaps can be non-linear and species specific. For example, a recent analysis by Tang et al. (2019) shows that within-crown gap in a 30m pixel increases as tree cover increases and reaches a maximum of ~10% for conifer trees. The current study addressed the definitional discrepancy by dividing MODIS VCF by 0.8 based on Hansen et al. (2002). A discussion on the uncertainties associated with applying this uniform relationship will benefit the manuscript. Better yet, quantitatively analyze the uncertainty with the collected in situ data, if possible.

The manuscript can also benefit by adding a clear description of how Canopy Area Index (CAI) is measured or calculated. The reference of Torello-Raventos et al. (2013) is provided, but the details are not sufficient for understanding how CAI is derived and what’s the uncertainties associated with CAI at the site level, especially for those who are not familiar with this paper. From Torello-Raventos et al. (2013), it appears that CAI was not directly measured at sites, but was calculated using allometric equations. Then, what is the uncertainty level associated with CAI estimations at the 100 m x 100 m sites, attributable to the allometric equations or the reference data underlying the allometric equations? Site-level uncertainty could be associated with specific direct measurement techniques. Moreover, tree cover definitions and measurement techniques could be connected, as hemispherical photography, terrestrial laser scanner, airborne waveform and discrete-return lidar all tend to measure different types of tree cover (e.g. Fiala et al. 2006, Korhonen et al. 2006, Rautiainen et al. 2005, Tang et al. 2019).

The inherent scale discrepancy between 100 m x 100m sites and 250m x 250m pixels is nicely addressed by simulations. Comparison results between the four types of simulations are also interesting. The authors could consider including the Landsat VCF data (Sexton et al. 2013) in the analysis, which is a satellite-based product most close to MODIS VCF. With a 30m x 30m spatial resolution, Landsat VCF can be averaged to close to the site scale, and a circa-2005 Landsat VCF product is available. This might generate additional insights, and might help resolve the difference with Brandt et al. 2020 in the Sahel region.

The striking difference between open shrub and savannas (Figure 3) is puzzling, even with the discussion provided. Much like fractional land cover (e.g. MODIS VCF), the accuracy of discrete land cover classification such as the MODIS IGBP land cover product is also generally lower over open canopy ecosystems, and misclassifications often occur among those classes. Could the in situ data provide some estimates on the accuracy of MODIS land cover product as an additional analysis? In addition, if open shrubland and grassland do have higher accuracy, for completeness, this might be better pointed out in the abstract.

Technical corrections:

Figure 1. Quite a few savanna sites have TROBIT tree cover > 60%, which falls in the definition of forest in the text (lines 22-23). Are those sites better considered as savanna or treated as forest sites?

Line 313, change “classed” to “classified”

References cited:

Brandt, M., Tucker, C.J., Kariryaa, A., Rasmussen, K., Abel, C., Small, J., Chave, J., Rasmussen, L.V., Hiernaux, P., Diouf, A.A., Kergoat, L., Mertz, O., Igel, C., Gieseke, F., Schoning, J., Li, S., Melocik, K., Meyer, J., Sinno, S., Romero, E., Glennie, E., Montagu, A., Dendoncker, M., Fensholt, R. 2020. An unexpectedly large count of trees in the West African Sahara and Sahel. Nature, 587, 78–82

Fiala, A.C.S., Garman, S.L., Gray, A.N., 2006. Comparison of five canopy cover estimation techniques in the western Oregon Cascades. For. Ecol. Manage. 232, 188–197.

Hansen, M.C., DeFries, R.S., Townshend, J.R.G., Marufu, L., Sohlberg, R. 2002. Development of a MODIS tree cover validation data set for Western Province, Zambia. Remote Sensing of Environment, 83, 320-335

Korhonen, L., Korhonen, K.T., Rautiainen, M., Stenberg, P., 2006. Estimation of forest canopy cover: a comparison of field measurement techniques. Silva Fenn. 40, 577–588.

Rautiainen, M., Stenberg, P., Nilson, T., 2005. Estimating canopy cover in Scots pine stands. Silva Fenn. 39, 137–142.

Sexton, J.O., Song, X.-P., Feng, M., Noojipady, P., Anand, A., Huang, C., Kim, D.-H., Collins, K.M., Channan, S., DiMiceli, C., Townshend, J.R. 2013. Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. International Journal of Digital Earth, 6, 427-448

Tang, H., Song, X.-P., Zhao, F.A., Strahler, A.H., Schaaf, C.L., Goetz, S., Huang, C., Hansen, M.C., Dubayah, R., 2019. Definition and measurement of tree cover: A comparative analysis of field-, lidar- and Landsat-based tree cover estimations in the Sierra national forests, USA. Agricultural and Forest Meteorology, 268, 258-268

This manuscript undertakes an anlysis of the accuracy of the MODIS Vegetation Continuous Fields product with particular reference to sparsely wooded ecossystems using tropical forest and savanna field inventory data.

The manuscript is well written and presented. The figures are of high quality and the analysis is clearly described. As the authors related the VCF has been subject to considerable analysis and discussion over some period of time. It could reasonable be postulated that all products from moderate resolution sensors struggle with accurate discrimination along the ecotone between forest and grassland due to non-linearilties in the reflectance and VIs applied to derive them, and due to the enormous variation in the morphology, architecture, density and clumping, and phenology of the overstory.  

The authors have to work with a limited inventory data set - limited in geographical coverage, and limited in sampling the above mentioned variation. They quite reasonably focus on the issue of clumping of the tree fraction within pixels and undertake a nice analysis based on this. I do wonder why they did not explore the actual representativeness of the field inventory sites versus the VCF pixel resolution in the manner of Roman et al., 2009 and subsequent publications.

[M.O. Román, C.B. Schaaf, C.E. Woodcock, A.H. Strahler, X. Yang, R.H. Braswell, P.S. Curtis, K.J. Davis, D. Dragoni, M.L. GouldenThe MODIS (collection V005) BRDF/albedo product: assessment of spatial representativeness over forested landscapesRemote Sens. Environ., 113 (2009), pp. 2476-2498, 10.1016/j.rse.2009.07.009] 

This is not in any way disqualifying since the analysis here is coherent and valid in itself.

However, I think that there are several issues that the authors need to address more fully especially in their Results and Discussion.

1. The manuscript proposes a correction to VCF for savannas and forests and combined. The authors state in the Results (lines 224-229) that VCF estimates forest tree cover well and greatly under-estimates savanna tree cover. However, based on Figure 1. the performance of VCF at the forest sites looks pretty terrible based on the sites being in a pretty clear blob with very wide variation between Trobit and VCF.  This makes me wonder about the representativeness issue for these Trobit sites and VCF pixels and whether clumping issues are the only thing happening here. In any case, Figure 1 does not suggest that VCF is doing well in forest but is this a sampling issue?

2. I am concerned about proposing an overall correction to VCF based on such a limited sampling of the ecotone between grassland and forest. I wonder if it would be better to more clearly identify in the written text, the kind of savanna that is sampled by the Trobit. Figure A1 shows the distribution of sites. Although the sites appear to sample the gradient between the amazon and the cerrado reasonably well (however there is enormous variation within cerrado from wooded to very open short sparse shrubland), they do not sample the variation in savanna structure in Africa and Australia very well. The African sampling is confined to the tree cover gradient in West Africa passing from the Guinean Savana to the Sahel, whilst the sampling in Australia is confined to areas around the small tropical rainforest area and surrounding Einasleigh Uplands. in northern Queensland. There is enormous variation in structure, morphology, and phenology across African and Australian savannas. I would find the study more compelling if it: 1) paid more attention to the actual species, structure , phenology and composition of the sites; 2) constrained the narrative to an analysis of VCF for these particular systems; and 3) suggested a correction approach that is relevant to these kinds of systems and maintains the focus on the issue of sparsly wooded systems. I believe that it is a stretch to propose an overal correction to VCF from this study simply becuse of the available inventory data are limited in coverage of geographical and vegetaion diversity. 

3. A key issue with VCF and this analysis is the thorny one of the definition of tree cover. Really with VCF and any other remote sensing products it should be about detection of woody cover (of any sort) from the canopy reflectance which is distinguished from the background soil and understory wherein the sensitivity is constrained by pixel resolution and the discriminatory capacity of available reflectance bands. I think that the manuscript gets a bit bogged down in this thorny issue since VCF has a certain definition which is apparently based on height. This is problematice in many ways since although related, height does not have a one to one relationship with canopy extent. However this MODIS product has been around for a long time with clear definition.  This rather means that analysis using these Trobit sites needs to provide information on the "tree height" distributions at these sites. This once again returns to representativeness, and maybe explains why the forest data are so scattered for Trobit versis VCF in Figure 1 . So again I want more information about what the vegetation is at these Trobit sites. 

4. As a result, the Discussion is a bit convoluted. The section between lines 301 and 319 therefore is rather confusing and muddled to read. A whole lot of speculation about where trees are > 5 m or not is included. To make statements about this one needs evidence of the tree height distribution at the Trobit sites and a matching analysis. I treat VCF as attempting to estimate woody cover. I don't care if it is getting at trees > 5 m or less. So this is kind of a furphy. The results in Figure 1 for savannas are pretty convincing about underestimation of woody cover which is probably unsurprising given the VCF method, but readily corrected based on your analysis at least for these particular savanna systems. If the Trobit data are limited then a whole lot of speculation about tree height really is not relevant and just clouds the Discussion. It damages the message which is best from the savanna analysis. Hence I think that clearer and more simplified findings and discussion points are required.

In summary, the study is well written, the methods are good, and the presentation is excellent. It is an interesting and useful study. However the authors need to address the following major issues before it can be considered for publication.

1. Limit the scope to developing corrrection to VCF that can be applied in the systems where Trobit inventory is available and representative.

2. Address the issue of representativeness of the Trobit sites, describe them much more fully, provide information of tree heights or tree height distributions or information from the literature on typical structure for these types of savannas and forests.

3. Focus the correction of the savannas for which your analysis is more convincing and explain the variation in the forest tree cover and why you think that VCF is good based on Figure 1.

4. Simplify the issues around definitions and detection of "trees". If you have no height data you can't really comment on this. Make sure that the thrust of the Results and Discussion is clear and clean and does not jump around between phrases about height definition when the imagery is seeing tree canopies not heights. If you can better describe the height distributions and representativeness of the sites, then discussion of the height definition can be useful.

This study evaluates the bias of MODIS VCF with field data and implements a correction for VCF. The manuscript is well written and presented with nice figures. But I am more interested in the method part you developed to do comparisons than the corrected VCF map and the findings of underestimated tree cover which are not surprising. As I get worried about the feasibility of using the limited number of samples  to correct MODIS VCF maps covering the whole tropics, which will further make the related results pale. For the IGBP map, please specify year used and the spatial coverage of IGBP for grouping. My suggestion is to just focus your analysis at the site level or to limit your study area to where the TROPBIT data are representative.

Small questions,

1, specify the meaning of the terms “tree cover distributions”

2, should consider moving the details about the TROBIT field data to methods section

3, please specify the rationality of using the equation 1.