Mapping Soil Organic Carbon Stock and Uncertainties in an Alpine Valley (Northern Italy) Using Machine Learning Models
Xiangtian Meng
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsI have carefully read the manuscript entitled " Mapping Soil Organic Carbon Stock and Uncertainties in an Alpine Valley (Northern Italy) Using Machine Learning Models". On the whole, there is no problem with the logic and principle. Some meaningful conclusions have been drawn, but I still worry about the following questions:
The environmental covariates used mainly include terrain factors, why use so many terrain factors? Why not choose some indicators that can reflect the process of soil formation, such as organisms and parent materials?
Comments on the Quality of English LanguageMinor editing of English language required
Author Response
We thank the reviwer about thier revision and their segestion, please see the attachement fille with all the answers.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsDear Authors
The manuscript entitled: Mapping Soil Organic Carbon Stock and Uncertainties in an Alpine Valley (Northern Italy) Using Machine Learning Models, can be accepted to be published in LAND after major revisions.
The work has regional interest. Presentation quality can be improved if the authors could go deeper in the analysis of covariates and the obtained SOC results. The work is interesting for the reader and sound.
The authors could improve the conclusion and abstract by adding the major mentioned limitations in SOC stock ranges obtained by modeling - Line 454–473.
Also, you can present the overall average SOC stocks obtained by modeling and compare with the results in Table 2.
I think that is also of great importance to present average SOC10 and SOC30 data for different elevation ranges, aspects, and slopes. The differences must appear. You focused well on modelling approach and methodology, but maybe some reader is interested to see the deepened results.
Considering the text, you can find here some smaller mistakes and suggestions:
Abstract:
Please add the total area of Valchavienna valley in the text.
Add the part about the limitations (Line 454-473), written in discussion, into abstract – one sentence.
Introduction:
Line 86: typing error
Materials and Methods:
Line 140: Classified as ...... according to.... WRB
Line 159: What does it mean spatialization of SOC stocks?
Line 195: Where:
Line 198: I did not understood well. What covariates you used for particular model? If you did not use the same for all, than you can present in table 1, for particular model.
Results
Line 342: Predictors importance in figure caption should be written following the order in figures
Line 363: order of appereance should be logical, start with RS....
Discussion:
Line 403: typing error, superscript...
Sub-chapter 4.2: You can present in this chapter average SOC stocks for the entire valley, and for different topographic positions.... with rough explanations.
Conclusion
Add limitations about SOC stock ranges from discussion.
Kind regards
Author Response
We thank the reviewer about the revision in the attached file you may find the answers about all the comments and questions.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThis paper does make a contribution to the literature. That said it is the application of established approaches in a context that has not been assessed much. The 10-fold CV is not a robust measure of performance, but given the limited dataset it is acceptable. There are some minor issues throughout the paper that need to be addressed, and it would be improved by including performance plots for the models.
Minor grammatical errors of typographic errors such as:
- Missing periods after sentences
- Missing spaces after periods
- Superscript applied where not necessary
- One sentence paragraphs in places
Line 177: Wakley-black and LOI are both not ideal SOC methods. Future work should involve determining SOC by dry combustion. LOI tends to overestimate SOC due to water associated with clays, and Walkley-Black leads to underestimation of SOC. While there is a correction factor for Walkley-Black they are not always it is not always appropriate for a given type of soil.
Line 178: What was the criteria used for determining very high organic matter. Also given the different types of error associated with the two methods, how much agreement was there between the two methods?
Line 186: Performance of the pedotransfer function should be reported.
Table 1: This should this be topographic wetness index, not terrain wetness index.
Line 233: Explain how the categorical variables were encoded to be used in the SVM.
Line 235: What was the ANOVA done? It is not clear why in the introduction. Did the data meet the assumptions of an ANOVA?
Line 241: What type of standardization was used?
Line: 272: These two metrics don’t provide a full picture of the model performance. Also provide the bias and Lin’s concordance correlation coefficient values.
Line 302-308 and Figure 3: The objectives of this study didn’t included the land cover comparisons. The authors need to establish it as a stated objective of the paper. Also, structurally I would recommend discussing the mapping results first. I would then discuss how that explains land cover as a predictor in the RF model.
Line 326: Are these percent error rates based on average values? This needs to be explained.
Line 327: Rephrase higher results of MAE as lower RMSE, as high indicates worse performance.
Line 443: What was the distribution of points by topographical parameters. Were less points located at these higher altitudes and slopes. That would also contribute to higher uncertainty. I see this is discussed later, and adding what percent of samples were from these types of locations would be informative.
Line 461: This is a very well documented phenomenon associated with tree-based models, and a limitation with RF models. I would recommend the authors add the scatter plots for the models as a figure.
Line 470: High amount of SOC is relative concept. I would rephrase.
Comments on the Quality of English Language
The paper is well written. There are however minor issues to correct.
Author Response
We thank the reviewer about the revision, in the attachment you may find the answers on the comments
Author Response File: Author Response.pdf
Reviewer 4 Report
Comments and Suggestions for AuthorsOverall a well-written article, congratulations to the authors. I have only a few notes:
1. Section 2.2.2: this stems from an unpublished study, but can you at least mention how many data were used / what methods were used etc.?
2. Section 2.2.4: you need to provide more details about how your performed hyperparameter optimization. Which parameters were optimized per each learning algorithm and which are kept study? Did you perform grid search and if so what was the grid (what values explicitly)? What kernels did you test for SVR? Etc.
3. Quality of Figure 2 is low
4. Figure 3: did you consider keeping one figure and placing the boxplots side by side per LU to aid the comparison? (See e.g. Figure 4)
5. Table 3: Enet should be capitalized like in the rest of the manuscript. Also I would advise using bold to denote the best results
6. Why did you use a statistical test to perform feature selection and not e.g., recursive feature elimination? I would add a sentence or two to justify the selection
7. Section 2.2.5 It's not clear to me if 50 runs is 10-fold x 5 different random seeds or something else. Please state explicitly.
Minor:
There are some R2 values without superscript
Line 281 m-2 missing superscript
Consider citing caret and pinning the version used
Author Response
We thank the reviewer about the revision, please see the attachment .
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsAfter revising the manuscript, I have no further questions
Reviewer 2 Report
Comments and Suggestions for AuthorsDear Authors,
congratulations for you work.
Kind regards
