GENERAL COMMENTS BY REFEREE

This is an extremely useful dataset based on two decades of MODIS surface reflectance for all of South America, with interesting example applications clearly presented by the authors. 

The introduction provides a succinct explanation of view-and illumination angle effects, for Modis satellite images, on the reflectance of a textured forest canopy.  This problem is removed by an empirical inversion that requires several images close in time with different view and illumination angles. This is the Nadir Adjusted Reflectance (NAD) product which the authors provide, with the additional benefit of state-of-the-art MAIAC cloud removal algorithm. It is already standardized to a fixed nadir view and a fixed illumination angle, facilitating its use by a much larger number of educators and scientists.

The authors then up their game by extracting useful information from this view and illumination angle "artifact", rather than just treating it as something to be removed.  This is their anisotropy (ANI) product:  the difference between reflectance under a standardized back-scatter geometry and a standardized forward scatter geometry. This is like the difference between the brightness of a highly irregular textured surface photographed with the sun behind the photographer and the same surface with the photographer facing the sun.  Intuitively, the difference in reflectance (or in vegetation indices) will be greater for more irregular surfaces and lesser for smoother surfaces. They show this ANI difference is useful for detecting canopy height in the Amazon, presumably because a canopy with tall trees and large crowns is more irregular than a canopy of shorter trees of similar height, that make a smoother canopy.

The paper provides three interesting examples of applications. First, they show that the Anisotropy attribute, as expressed in a single month of EVI vegetation index, distinguishes three Amazon forests which are not separable using the typical nadir Adjusted EVI.  They then show that their very novel Anisotropy product is useful for estimating forest height across the entire Amazon, by comparing to GEDI lidar heights. Finally, they show that each of nine distinct leaf phenology regions of the Amazon (from an independent study) are corroborated by distinct ANI and NAD seasonal curves for the EVI vegetation Index.

ANSWERS TO REVIEWER GUIDANCE QUESTIONS IN CAPS

Are the data and methods presented new?  YES

Is there any potential of the data being useful in the future? VERY HIGH

Are methods and materials described in sufficient detail? YES

Are any references/citations to other data sets or articles missing or inappropriate? NO

Is the article itself appropriate to support the publication of a data set? YES. THE ARTICLE PROVIDES EXAMPLES OF VERY USEFUL APPLICATIONS. THEIR ANISOTROPY PRODUCT WILL VERY LIKELY LEAD TO A SUITE OF NEW PAPERS ON FOREST STRUCTURE AND PHENOLOGY

Check the data quality: is the data set accessible via the given identifier?

YES, I accessed the main Zenodo datasets and the two auxilliary sets. The latter allow calculating several indices based on hotspot and darkspot, that are described in Table 3 of the ESSD submission. All datasets are explained succinctly on Zenodo and in item 5 of the ESSD submission. I was also able to access the Earth Engine repository containing two Image Collections, one for anisotropy of EVI and one for Nadir-adjusted EVI.  Both worked fine, using the sample code provided.

Is the data set complete? Are error estimates and sources of errors given (and discussed in the article)? Are the accuracy, calibration, processing, etc. state of the art? Are common standards used for comparison?

REPLY: The authors provide the number of observations per month as a proxy for error estimation. More observations provide not only more complete data but also a more reliable BRDF inversion. The cloud masking algorithm is state-of-the-art and its originator is among the authors.

Is the data set significant – unique, useful, and complete?

VERY SIGNIFICANT for scientists and educators that make use of MODIS reflectance for vegetation studies. The BRDF problem with MODIS data has been a subject of much discussion and controversy relating to Amazon forest resilience in the face of normal and extreme droughts. Here the authors not only provide corrected data, but they also turn lemons into lemonade by showing that forest structure (including canopy height) and forest leaf phenology in the Amazon are detectable by exploiting the BRDF as a measure of the anisotropic reflectance properties of canopies. So the data is also very useful. Because so much processing time is required and because few studies have previously explored the anisotropy as a useful property rather than as noise or bias, the data is unique. It is spatially complete, covering all of south America.

Consider article and data set: are there any inconsistencies within these, implausible assertions or data, or noticeable problems which would suggest the data are erroneous (or worse). If possible, apply tests (e.g. statistics). Unusual formats or other circumstances which impede such tests in your discipline may raise suspicion.  NO PROBLEMS DETECTED HERE

Is the data set itself of high quality? YES

Check the presentation quality: is the data set usable in its current format and size? Are the formal metadata appropriate?  THE DATA IS ACESSIBLE IN POPULAR FORMATS

Check the publication: is the length of the article appropriate? ARTICLE IS WELL WRITTEN WITH KEY EXAMPLES OF DATA APPLICATION

Is the overall structure of the article well-structured and clear? CLEAR AND CONCISE 

Is the language consistent and precise? GOOD WRITING STYLE

Are mathematical formulae, symbols,abbreviations, and units correctly defined and used? YES

Are figures and tables correct and of high quality? YES.

Is the data set publication, as submitted, of high quality? YES

LINE BY LINE COMMENTS

Lines 179-180    You obtained RTLS BRDF inversion parameters from pixels observations (having different view and solar angles) within eight day periods. What is the minimum number of pixel observations required to run the inversion in an eight-day period?

Lines 195-197  Are these pixel observations required for RTLS BRDF inversion conceptually identical to the "per-pixel number of samples (or observations) for each monthly composite", which is provided as ancillary data?

Figures 3 and 5 -- Topographic effects on ANI? In Figure 5, ANI data linearly predict forest height with R2 = 0.55, presumably because the more coarsely textured surface of tall-tree canopies makes the shaded sides of trees and of large crowns occupy a greater fraction of a pixel viewed in forward scatter situation, if compared with a smooth canopy such as grassland or more even-height dicotyledon forest canopy. But your data are at 1 km resolution, so there will also be topographic irregularities within each pixel, which might also contribute to higher ANI. Have you looked into the grain and/or amplitude of topographic roughness (from SRTM) as additional explanatory variables for your scatter-plot relating ANI to forest height (Figure 5)? Do you think this will in fact be relevant?

In the lower three panels of Figure 3, Tapajós and Xingu Park are flat relief, so the ANI should be showing differences in canopy texture (which generally increases with canopy height in the Amazon), not topographic effects. However, I looked at the forested areas at or near the sample site in the São Felix window using SRTM and see that it is a mixture of some patches of flatter and others of more irregular relief.  The ANI data there is also patchy in regions of intact forest. Is there some direct effect of topographic relief on ANI taking place in the São Felix site? Or is the patchy mosaic of low and high ANI within forest there mostly related to patchy change in canopy height/smoothness?

Lines315-320  Fascinating. Your data is opening up new avenues for understanding leaf phenology

Line 315  Fix the grammar

Line 320  change to "the central"

Line 350   change "on" to "for"

Line 355   change "consists in" to "is"

Line 366   change to "in the northwest"

Figure 6   Great figure, Distinct mutual relationships between the two indices in each pheno-region lend credence to the pheno-region classification of Xu et al (2015)

This study introduced the AnisoVeg product consists of monthly 1-km composites of ANI and NAD surface reflectance obtained from the MODIS over the entire South America. The paper needs a minor revision before it can be considered for publication.

1. The MODIS product MCD43 relies on multiple observations over a 16-day period, while in this study the period is extended to a month. A period of month may represent significant changes in surface especially during the vegetative stage. That would cause an inaccuracy anisotropy information of surface. Explain the possible implications of this change.

2.To generate accurate surface anisotropy, the weight of different observations should be inconsistent in the retrieval of surface BRDF by RTLSR model. The quality and the time of the observation need to be considered together.  

The manuscript “AnisoVeg: Anisotropy and Nadir-normalized MODIS MAIAC datasets for satellite vegetation studies in South America” by Dalagnol et al. describes the production of a data set on vegetation anisotropy derived from MODIS data for South America. There is considerable potential for such a data set to provide useful information about the state of the land surface, and one tantalising hint that the authors provide is the result that the ANI is able to explain R²=0.55 of the variability in the GEDI canopy height signal (Fig 5., c.f. R²=0.16 for the normalised reflectance). Overall, I think this is a well written paper which describes a potentially important data set. I do have a few grumbles, mostly minor, but I hope the authors take these in the spirit in which they are intended – I am only seeking to improve the manuscript, and I am not suggesting any complex changes.

Main comments:

I find it strange that MCD43 isn’t mentioned anywhere. The implicit claim is, I assume, that the MODIS MAIAC data is far superior to the MOD09 and MY09 which is used to derive MCD43. I don’t disagree with this, but one could also use the MCD43 data to produce similar data (not to mention that MCD43A4 contains an NBAR product which is similar in concept to the NAD in this paper). Some discussion of this in the introduction is necessary.  

Fig 3 does not convince me that there is complimentary information in the NAD and ANI data. Some additional metric to show how much different information is in there would be useful. For example, if the authors calculated the principle components, how much variance would the second component explain? [note – I am convinced of this by some of the later results, but I it should ideally be demonstrated here too.]

L409: Table 3, captions says: “Examples of other multi-angular anisotropy indices that can be further calculated using layers of the AnisoVeg product.” Initially I thought this wasn’t possible as earlier the manuscript gives the impression that the layers are only ANI and NAD, however I see now this isn’t the case. I suggest including a table explaining what the actual layers of the AnisoVeg product are. On a related note, although the authors call “H” the hotspot in this part of the paper, the algorithm apparently doesn’t compute the value in the hotspot direction and instead use 35 degrees (see Line 217 - and I agree this is a sensible thing to do). The authors should only call this “back-scattering” so as not to give the wrong impression – it is not in the hotspot.

Minor comments:

L59 – I think this sentence could cause confusion between what the definition of anisotropy is, and what causes it. Anisotropy is defined as the departure from Lambertian scattering, it is caused by the physical structure of media through which photons pass. I am certainly not doubting that the authors know this, but I think it could be made clearer to the reader. I am also not sure about the use of the word “mechanical” in this sentence.

L73 – the Foody and Curran reference is a bit of an odd one to include to support this statement. Their paper doesn’t really look at the influence of biophysical properties on the surface anisotropy, although it does include a correction for the influence of terrain on the observed radiance. With no disrespect to either Foody or Curran, there are many more relevant papers that could be included here. Suggest finding some different references.

L110 – Again, I do not think the Foody and Curran reference is the best choice of references here. The totality of what it says on this subject is: “Terrestrial land cover surfaces are typically non-Lambertian reflectors and may exhibit a class- specific angular reflectance response. Thus data acquired at different angular geometries may help to identify and characterize land cover classes in both optical (Barnsley, 1994) and microwave (Foody, 1988) wavelengths.”  Whereas the current manuscript attributes “the use of multi-angular information to obtain metrics of anisotropy and extract information on forest structure” to that paper. I think this is a bit of a stretch. Suggest finding some different/additional references.

L113 Whilst the Sandmeier et al., 1998 reference is appropriate here, it is most definitely not the “first” example of this type of work. It is an early example though, and perhaps that would be a better way of describing it.

L179 Another strange reference. The Lucht and Lewis paper refenced presents a really nice results around the so-called “weights of determination” of the kernel BRDF models, but as a general reference for the RTk-LSp model it is an odd choice. A more obvious paper would be, for example, Wanner et al. (1995).

L184: Eqn 1 – why are the labels for the kernel weights superscripted (e.g. k^v) and the kernel values subscripted (e.g. F_v)? Ultimately, it doesn’t greatly matter, but it would be better if these were made consistent, unless there’s a good reason for not doing this.

L184: Eqn 1 – I find it odd that the with kernel values are given the symbol “F” and the kernel weights are given  the symbol “k”. Traditionally in the literature it has been the other way around see, for example Wanner et al. (1995) or, indeed, Lucht and Lewis (2000). This tripped me up whilst reading the paper, and a later statement appeared wrong to me due to this, so it could cause confusion. I strongly suggest changing this so that it adheres to the convention.

L200: “0.009107388 degrees” – this is quoted too precisely - 0.000000001 of a degree is a fraction of a millimetre. The text goes on to say that it is “approximately equivalent to 1 km” so really only needed to quote to that precision (say 4 or 5 d.p. in degrees).

Typos etc:

L98 product -> products

L205: Here an astrix has been used as a multiplication sign, whereas in Eqn 1. an actual multiplication sign was used. Suggest making consistent.   

L315 “The EVI_NAD and EVI_ANI are seasonal variability and…” this doesn’t scan. Should it say “The EVI_NAD and EVI_ANI are seasonally variable and…”?

References:

Wanner, W., Li, X. and Strahler, A.H., 1995. On the derivation of kernels for kernel‐driven models of bidirectional reflectance. Journal of Geophysical Research: Atmospheres. 100(D10), pp.21077-21089.

comments orgnized in the attached supplement