Estimation of Aboveground Vegetation Water Storage in Natural Forests in Jiuzhaigou National Nature Reserve of China Using Machine Learning and the Combination of Landsat 8 and Sentinel-2 Data

Round 1

Reviewer 1 Report

This manuscript presents the classification a regression results obtained by different Machine Learning (ML) methods in the Jiuzhaigou National Nature Reserve (JNNR) area to estimate the Aboveground vegetation water storage (AVWS).

General comments regarding the manuscript

• The authors may justify why they use just one image in the experiments (one Sentinel-2 (S2), one Landsat-8 (L8)). In principle, I don't see any reason to repeat the experiments using at least two or three images to further validate and give more consistent results.
• Concerning the above, it is also not clear why the authors use two images separated by more than one month in time (2022-11-21 L8, 2022-12-31 S2).
• I miss more in deep description on the classification procedure applied after the SVM classifier. This seems like a key step to obtain all the classes shown in Fig. S1.
• Related with the previous point, I also miss quantitative results on the classification obtained in all classes, not only forest vs non-forest.
• As the authors raise in the discussion, typically physical RTM models are used to estimate AVWS. I miss some baseline results obtained with physical models just to have an idea of how well the ML models are performing. From a pure ML perspective a R^2 of 0.57 does not seem a very good result, but it could be the case (I don't know) that the AVWS estimates obtained with ML models are actually good, or equally good, compared to the ones obtained by classical physical models.

In detail comments (the number in the left is the line number in the submitted manuscript).

158 - It is not clear the pre-processing carried out in the S2 bands. Some bands need to be upscaled from 10m or 20m to 30m, other downscaled from 60m to 30m. What are exactly the methods used in that steps?

171 - As I stated before, the authors should give more details on the processing carried out after the SVM classifier, i.e. majority/minority analysis, clump and sieve processing, and how these allow to obtain the final classification map.

206 - The formula in Table S4 is note clear. I don't understand the 'i' appearing at the beginning of all formulae, for instance in the mean, sum of i(P_{i,j}), the first 'i'. It this 'i' the same index of the summation? Could you explain further?

207 - Pearson correlation only establishes linear relationships between variables, but if there are non-linear associations you would miss them. A quick workaround would be to use Spearman correlation instead. Also, should this paragraph be part of next subsection, 2.7. Feature selection?

246 - The end of this sentence (high dimension data, ...) does not seem grammatically correct.

278 - The authors report the kappa coefficient (KC) and overall accuracy (OA) of the classification of forest vs non-forest. Related with my previous comment on describing in more detail the processing done after the SVM classification, here the authors should give results on the quality of the final classification, reporting also KC and OA for the classes shown in Fig. S1.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

I like the work, and I feel it is very well written. A few minor comments follow:

** Title: A geographical location of study area (e.g., China), and vegetation characteristics (e.g., Natural or plantation forest, or needle-leaved or broadleaved or mixed forest) should be added.  Do the same for keywords. 

** Line 63: Add some recent citations about this statement (Be sure, I am not one of the authors of following paper), e.g.:

Moradi, F.; Darvishsefat, A.A.; Pourrahmati, M.R.; Deljouei, A.; Borz, S.A. Estimating Aboveground Biomass in Dense Hyrcanian Forests by the Use of Sentinel-2 Data. Forests 2022, 13, 104. https://doi.org/10.3390/f13010104

** Paragraphs fifth and sixth of Introduction: There is no connection between them. 

** Study area: Add information about the climate and meteorological parameters of the study area. 

** Lines 109-111: The third part of each scientific name should not be italic.  

** Lines 117-118: The maximum road edge 100 m, is enough and just need to add some citations to shows 100 m is common. See, for example, these papers: 

 https://doi.org/10.3390/f12121805

https://doi.org/10.1007/s10342-018-1138-8

** Line 137: How about the Post Hoc analysis? 

 ** 2.8.4. Model assessment: Nash or Kappa coefficients would be nice to consider. 

** Table 1 and rest of papers: Its not true with ANOVA, you need to use post hoc test (e.g., Duncan, Tukey, etc) to show "a" "b" and other signs. 

** Figure 3: Improve the quality of this figure. Also, add a sample number for each panel. It would be pleasant to separate four scenarios (Coniferous Forest, Broad-Leaved Forest, mixed forests, and all forest), and prepare this figure for each of them.  

Lines 392 to 394: This is not a research limitation. You have to reconsider it. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

As almost all changes in the current version of the manuscript are related with my previous comments and questions, I focused this revision on those.

Concerning the reasons to justify the use of a single image for the experiments, only the third one, difficult to obtain cloud-free images in the study are, sounds convincing. Giving that it seems not possible to provide better and consistent results using more than one experiment, in my opinion the authors should explain these problems in the manuscript.

Regarding the description of the classification procedure done after the SVM classifier, I appreciate the explanations given the authors in their reply. However, in the manuscript they only have included the parameters of the algorithms used in ENVI. While this information is quite valuable and worth to be included in the final version of the manuscript for the sake of reproducibility, in my opinion they should also include at least a summary of the descriptions they give in their answer. This would  make the final manuscript more complete and easy to understand to readers.

My next suggestion was to include a reference to results obtained with physical models (RTMs). The answer provided by the authors is worth to be included also in the manuscript, adding the propper references to support their claims.

Concerning the process or upscaling and downscaling the S2 image to 30 m, again the authors can include the information that they used bilinear resampling for this operation, as stated in the anwser.

Finally, regarding the formulae in Table S4, the only thing that has changed in the updated version is the caption of the table that now begins with 'The i represents the row number and the j the column number'. I'm afraid this do not makes sense. Take for instance the mean, if 'i' is the row number, then are you weigthing the pixel values (Pi,j) by row number?

Probably we have some misundertanding here. Again taking as example the 'mean', by i(Pi,j) in the expression you mean i * Pi,j? It can't be that.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The revised version of the manuscript has been significantly improved. I agree with the publication of the article.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf