This manuscript provides a thorough description of techniques combining very high-resolution imagery, ground photos, and Landsat scale imagery to classify landcover class and fraction cover of photosynthetic veg, non-photosynthetic veg and bare ground in rangeland over a region in East Africa.  The methods described here provide a useful template for broader scale implementation.  

A few suggestions:

L185:  Here it states data over four decades were used in this study.  However, the title and Fig. 3 suggest that data from 2000 to 2022 were used.  Please clarify.

L205-206:  The thermal bands are not atmospherically corrected with LaSRC, but rather with MODTRAN.  Since the thermal bands are not used in this study, suggest not mentioning them or their correction approach to avoid confusion.  Focus discussion on the VSWIR bands.

L234:  pixels that *were* not included

Fig. 3:  Explain the drop in %cloud-cover pixels in 2013.

L248-294:  These are not complete sentences.  Suggest: “The methodology to build long-term time series of LCC and VFC for rangelands in Eastern Africa is divided into three major steps: first, the development of a training/testing dataset from VHR imagery; second, the LCC classification; and third, the VFC classification.”

L254-259:  Long sentence. Suggest breaking up. The next sentence is confusing too and could use a rewrite.

L294:  Have these 8 classes been tabulated before this point within the main manuscript? If not, suggest doing so or including a table with some primary characteristics.  After reading further, I see this in Table 1.  Suggest introducing or referencing this table earlier in the text.

Fig 10: What specifically is the cause of the low percentage valid transition years of 2006, 2011 and 2018?  Was this discussed?

Fig 14:  It would be interesting to see a time sequence of this plot, perhaps 2-year intervals?  Do annual maps of change in fractional cover highlight regions of interest?

This study produced a land cover and vegetation fractional cover data in the arid and semi-arid Kenya, Ethiopia, and Somalia using a machine learning approach. The logic of the manuscript is clear but there are a few points that need to be cleared. First, this is a typical land cover classification study and the ecological implications are limited at this stage Thus, the “health indicators” in the title seem confusing. Where did the manuscript describe the implications of the ecological aspects of rangeland health? From what I have read so far, it provides a spatial-temporal distribution of different land covers. If rangeland health needs to be included, readers would like to see the driving forces analyses. Second, the methodology part is the most important part but it is not clear. In section 2.1, could you explain more to clearly show the three steps? Specifically, how were the reference data generated and used, given the importance and the intensive labor needed to compile such a dataset? I am curious also because I am not sure how these samples could be used to train the model to separate some of the very similar land cover types. Other minor suggestions:

1. Line 164, the full term of NGA should be shown when the abbreviation appears for the first time.
2. Line 165: HR or VHR?
3. Line 175, what is SD?
4. Lines175-178: Model trained by the samples during this period has been applied for classifying historical land cover characteristics. This approach was called transfer learning in some studies, which reported that it might have potential problems in capturing land use changes. Will the study area also be affected?
5. Line 179, “area of interest”: AOI
6. In section 2.2, could you show the spatial distribution of the samples, including the groups used for model calibration and validation, as well as the LC types?
7. Figure 9, suggest to use color and line type to show the trends. Also, the trends could be added.
8. Figure 15, show trends please
9. Line 607, this is an annual product. What do you mean the high temporal resolution here? Which dataset did you compare to?

The authors present a remote-sensing derived dataset on rangeland health for a region in Africa.

While I cannot judge the merit of the published data, the methods are well described, and the methodology seems sound.

It remains somewhat unclear why the specific study area was chosen, and not a larger area (e.g., all of Sub-saharan Africa).

A larger area would increase the usefulness of the published data.

I recommend a minor revision.

Further comments:

Fig. 2: very nice visualization of the spatio-tenmporal distribution of VHR data scenes used for training data generation. Would it be possible

to provide a similar figure for the Landsat scenes used for inference?

Fig. 7: It is surprising that elevation has the highest feature importance. Is there a strong dependence between class occurence and elevation? This could be tested in a quick analysis, based on the elevation at the training / validation samples?

(e.g. showing stacked barcharts of the class proportions per elevation slice). 

Also, what are the implications of this regarding model generalizability? 

Would your model also work in flat areas? It would be nice to read some more elaboration on generalizability of this method for other study areas - any expansion planned for the whole of sub-saharan Africa, or similar?

Table 1, Fig. 8: Could you spell out the class names (and maybe rotate them in the confusion matrix)? It is hard to read this table / Figure when the reader has to switch between tables to decypher the class names.

Fig. 14. Great visualization with striking spatial patterns.