Floristic vs. Dominant Classification Approaches Applied to Geospatial Modeling of Mixed and Broadleaf Forest Types in the Northwestern Caucasus (Russia)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

See attached file

Comments for author File: Comments.pdf

Author Response

We would like to express our sincere gratitude to the Reviewer for the time and attention given to our manuscript and for the valuable comments and suggestions. All major edits are highlighted in the updated manuscript file, and answers to your questions are provided below.

 

Comments 1: [Overall, the text is unusually long (introduction, methodology, results). In the abstract, describe the subject (context and issues) and present the methodology in a concise manner, but with a clear thread running through it. Expand on the results and present them in a progressive manner. Add a conclusion and clear implications regarding the results obtained.]

Response 1: [The Abstract was slightly modified to be consistent with the edits in the main text.] 

Comments 2: [With regard to the introduction, it is better to discuss the vegetation first, before talking about its mapping. Overall, the introduction should be concise and seek to answer the following questions: what is known about the subject under discussion? What is missing? How and in what ways (approaches) will this study fill this gap? Why is the case of the south of Krasnodar Krai and the Republics of Adygea and Karachay-Cherkessia (west of Mount Elbrus) so relevant to this topic?]

Response 2: [The structure of Introduction was highly modified.] 

Comments 3: [A clear hypothesis is missing.]

Response 3: [As a core idea of the study is a comparison of two approaches under the same data conditions, it barely can be expressed in the form of hypothesis (like “one approach should be better than another”), but the clarifications of study objectives were added to the Introduction.] 

Comments 4: [Reduce the description of the study area, considering two axes: the biophysical framework and location; the socio-economic framework.]

Response 4: [Done.

Comments 5: [The criteria for establishing the 100 m² plots are not explicitly stated. Were plots placed subjectively in homogeneous forest areas, or was a systematic or randomized scheme used? The statement that data were collected in "the least fragmented forest areas" is vague. It requires a clear, operational definition. How was "fragmentation" quantified or qualified? The description of the three vegetation layers is clear, but specifying the height thresholds used to define the "canopy layer" and "understory layer" would add precision and aid reproducibility.]

Response 5: [The clarifications were added to the corresponding section. Shortly, it was expertise-based (purposive) sampling, without a strict statistical justification. ]

Comments 6: [The reduction of the initial 558 plots to 515 is explained, but a table summarizing the reasons for exclusion (e.g., "X plots excluded due to land cover change," "Y plots due to spatial proximity") would greatly improve transparency]

Response 6: [The information was added to the text at the beginning of the Results chapter. ]    

Comments 7: [The description of the floristic classification process relies heavily on expert judgment ("we expertly judged the species ecological similarity"). While expertise is invaluable, this makes the method difficult to reproduce objectively. A more detailed description of the iterative table-sorting process in Juice, or a reference to a standardized protocol, would strengthen this section. The use of a 40% frequency difference threshold is a good step towards objectivity, but its justification could be expanded upon.]

Response 7: [The more detailed description was added. Still, the certain degree of “subjectiveness” is unavoidable in case of the floristic classification.] 

Comments 8: [The grouping of tree species into "dark coniferous," "light coniferous," "hard-leaved," and "soft-leaved" is a critical step that feels somewhat arbitrary and rooted in regional forestry tradition. A clear ecological or functional justification for these specific groups is needed.]

Response 8: [The clarification was added. This stratification was originally developed based on wood density (which deter-mined the economic value of tree species), but it is also valid in ecological context, as it naturally combines high-level botanical classification (coniferous/broadleaf) with forest succession stages. Particularly, light coniferous and soft-leaved groups con-sist of the early-successional species, while dark coniferous and hard-leaved groups are populated by the late-successional ones.] 

Comments 9: [Furthermore, the reference fraction patterns in Table 1 appear to be predefined without a strong statistical or ecological rationale (e.g., why are certain 0.5/0.5 combinations labeled "Mixed coniferous" and others "Mixed broadleaf"?).]

Response 9: [Yes, these fractions are intended to be strictly formal and, thus, universal, without any data-driven or expertise-based adaptations, in contrary to the floristic classification. The note about labels is unclear. In the context of the given table of fractions it can be only one straightforward interpretation of these labels – Mixed coniferous is an equal mixture of dark and light coniferous species, and Mixed broadleaf is an equal mixture of hard-leaved and soft-leaved broadleaf species.]

Comments 10: [For the dominant classification, it is stated that small classes (<5 plots) were "excluded from further analysis." It should be clarified if these plots were removed entirely from the modeling dataset or merged into other classes, as this affects the final sample size and class structure.]

Response 10: [The algorithm description was clarified.]

Comments 11: [A significant issue is the mismatch in spatial resolution between the predictor variables (HLS/DEM at 30m, SoilGrids at 250m, WorldClim at 1km) and the field plots (representing a 100 m² area). The potential ecological fallacy of using coarse-grained climate and soil data to predict fine-grained vegetation patterns is not discussed. An explanation of how this scale discrepancy was considered, or a justification for why it is acceptable in this mountainous context, is necessary.]

Response 11: [The note about scale mismatch was added to the Study Limitation subsection of the Discussion chapter. Such mismatches are generally unavoidable and directly affect the results of feature selection and the resulted model accuracy, but in most cases, we have nothing to do but to accept these data conditions, as we lack any relevant data of higher plot size or spatial resolution. But, as we can see, in our and in many other cases, even 1-km data can provide a sufficient amount of useful variation for 10-m plots. The larger plots and the more detailed geospatial data may (or may not) work better, so the specific studies must be carried out for better understanding of these issues.] 

Comments 12: [The choice of a 29-day window with a 15-day step for creating HLS composites is not justified. Why were these specific temporal parameters chosen over others? A brief justification, perhaps relating to the phenological cycle of the dominant species, would be helpful.]

Response 12: [The clarification was added. Shortly, they were empirically chosen for uniform time series smoothing and the processing simplification. Phenology-based periods are very tricky to use for the mountainous territory. ]    

Comments 13: [The text correctly identifies a key limitation: that global DEMs model the surface (including canopy) and not the terrain, and that SoilGrids are themselves models with inherent uncertainties. However, these critical limitations should be more prominently highlighted as they directly impact the interpretation of the results, especially the low importance of DEM-derived variables.]

Response 13: [The respective note was added to the Study Limitation subsection of the Discussion chapter.] 

Comments 14: [The calculation of 45 normalized ratios from 10 bands is mentioned. While comprehensive, it risks creating a high number of correlated, redundant features. A sentence acknowledging this and explaining that the subsequent FPCA and feature selection are designed to handle this redundancy would be useful.]

Response 14: [The respective note was added.] 

Comments 15: [The FOCI algorithm is applied to the entire dataset before cross-validation. This can lead to optimistic bias, as information from the entire dataset influences the variable selection, which is then evaluated on CV folds derived from the same dataset. A truly rigorous approach would perform the entire feature selection process within each training fold of the nested CV to avoid this form of data leakage. The current methodology may overestimate the model's performance when applied to new, unseen data.]

Response 15: [It's true. The respective note was added to the Study Limitation subsection of the Discussion chapter.] 

Comments 16: [The threshold for filtering low-variance variables (>95% identical values) is strict and reasonable, but its specific value is not justified.]

Response 16: [The primary idea, of course, was to drop constants, but the probability threshold was adjusted to be consistent with all other threshold values (variable correlation, FPCA variation) used in our study. The note about arbitrary nature of the value was added to the text. ]

Comments 17: [The results section is very data-dense, primarily describing what was found. It would benefit from a more narrative structure that interprets the results as they are presented. The results for the detailed classifications (MCC ~40-53%) are presented but not critically discussed until the Discussion section. The very low accuracy for several detailed classes (e.g., F23, D13) should be highlighted here as a key finding, emphasizing the practical challenge of mapping at this fine thematic resolution with the available data.]

Response 17: [The Results chapter was extended. ]  

Comments 18: [In Table 6, it is stated that certain variable sets provided "significantly better performance." The reader must refer to the methods to recall that this is based on a Wilcoxon test on MCC values. A succinct mention of the statistical test directly in the results text or a table footnote would improve clarity.]

Response 18: [Notes added in text and in table footer.

Comments 19: [While the confusion matrices (Tables 7-10) are provided, they are extremely complex and difficult to interpret in a manuscript format. Consider moving the detailed matrices to the supplement and replacing them in the main text with simplified, visual representations (e.g., heatmaps of the major confusion flows) to convey the key patterns more effectively.]

Response 19: [Done.] 

Comments 20: [The discussion of shared variables in Figure 3 is informative but could be more concise, focusing on the main takeaway.]

Response 20: [It’s slightly shorter now. ] 

Comments 21: [The discussion mixes results, limitations and implications. Reorganise it into subsections: (i) methodological limitations; (ii) main results; and (iii) implications and avenues for future research. Create a subsection dedicated to limitations (critical scale mismatch between plots (100m²) and climate variables (1km); Sampling bias in ‘low fragmentation areas’; Subjectivity of floristic classification vs. dominant algorithmic approach; Risk of data leakage in variable selection (FOCI across the entire dataset)) and in this section explicitly link these limitations to the modest performance of the detailed classes, then discuss the impact on the generalisability of the models. With regard to the main results, the key conclusions are buried in the text, without in-depth analysis. Emphasise that the lack of concordance between classifications is systemic; present the superiority of floristic models as a fundamental result and highlight the unexpected effectiveness of bioclimatic variables as a key discovery. Finally, the implications (theoretical, methodological and practical) for conservation vs forest management are underdeveloped, while future research avenues are too vague.]

Response 21: [The Discussion section was reorganized.]  

Comments 22: [Conclusion. In this section, (i) reiterate the objective and methodology; (ii) compare the results with the hypotheses; (iii) state the limitations and conclusion; and (iv) finally, discuss the implications and avenues for future research. The text should be presented concisely.]

Response 22: [The Conclusion section was slightly modified.] 

Comments 23: [There are still many old references.]

Response 23: [ Some old references were deleted or replaced.] 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper is very interesting and well written. The scientific problems are presented in a sound way and the goal of the research is clear and of scientific relevance. The most important point which I see needs improvement is the justification of using machine learning. this is a very powerful tool, but its use implies some requirements, and also justification in relation to the advantages in comparison to "traditional" statistic. Also, a deeper discussion of the ecological phenomena shaping the results is welcome. You test many variables, an explanation why these were selected is welcome. Finally, study limitations such as having a single study area should be better acknowleged

Author Response

We would like to express our sincere gratitude to the Reviewer for the time and attention given to our manuscript and for the valuable comments and suggestions. All major edits are highlighted in the updated manuscript file, and answers to your questions are provided below. 

Comments 1: [1) Lines 50-58 – Please include references.]

Response 1: [The Introduction section was reorganized and references added.]

Comments 2: [2) Line 127 – please consider using the scientific name of the species.]

Response 2: [Done.]

Comments 3: [3) Lines 138-151 – objectives should be better described as well as the research questions.]

Response 3: [The core idea and study objectives were clarified.]  

Comments 4: [4) Line 167 – reference of NaturalEarth is missing]

Response 4: [There were temporary text references in the manuscript, as the final reference list may be heavily changed after reviews. The updated manuscript version has all references in the proper format.]  

Comments 5: [5) Lines 175-194 – English is not clear. Please revise.]

Response 5: [Done.]  

Comments 6: [6) Line 205 – systematic rout survey or systematic sampling by transect?]

Response 6: [Transects, but not systematic. The respective subsection about field plots establishment was revised and edited.]  

Comments 7: [7) Line 212 – how was cover measured? Please include further details in the text.]

Response 7: [Visually. The respective subsection was updated with further details.]  

Comments 8: [8) Line 241 – expertly judge or expert analysis? 9) Line 244 – why 40%? Please include further details in the text.  10) Line 266 – why Phi>30? Please include further details in the text.]

Response 8: [The entire subsection about floristic classification was revised and edited with more details.]  

Comments 9: [11) Line 275 – projective canopy cover or crown horizontal projection? Also, the percent value of the crown horizontal projection is usually termed crown cover.]

Response 9: [Replaced with “total crown cover (horizontal projection)”.] 

Comments 10: [12) Lines 296-303 - English is not clear. Please revise.]

Response 10: [The entire subsection about dominant classification was revised and edited with more details.] 

Comments 11: [13) Line 317 – how was normalisation done?]

Response 11: [Both metrics are internally normalized and adjusted during their calculation, so our statement in manuscript is the clarification about their interpretation, not about additional processing step.] 

Comments 12: [14) Lines 321-341, 392, 404, 413, 420, 432 – the references of the software and image products are missing.]

Response 12: [See Comment 4.] 

Comments 13: [15) Lines 350-352 – it is not clear how many images were used and the dates of the images would be of importance.]

Response 13: [The image counts are added in the text, the counts of valid observations per temporal interval are added as separate table in Supplementary materials.] 

Comments 14: [16) Lines 353-362 – more details are needed.]

Response 14: [The entire subsection about satellite data was revised and edited with more details.] 

Comments 15: [17) Line 413 – the names of the 19 bioclimatic variables should be included in the text.]

Response 15: [Done.] 

Comments 16: [18) Lines 443-448 – English is not clear. Please revise.]

Response 16: [Revised: The initial composition of variables directly affects the predictive performance of the model and can also determine the limitations and potential of its further application. In our study, the source data for geospatial variables vary in origin, spatial resolution and the nature of the represented features. To analyze the impact of such differences on the performance of the resulting model, we tested several initial combinations of variables, namely: …] 

Comments 17: [19) Lines 464-474 – English is not clear. Please revise.]

Response 17: [The entire subsection about variation and correlation-base filtering was revised.] 

Comments 18: [20) Lines 472-473 – why did the authors used Pearson and Spearman correlation coefficients? More details in the text are needed.]

Response 18: [The clarification added: For this procedure, we computed both classical Pearson and Spearman correlation coefficients and used the higher of the two values as a pairwise correlation measure. Such approach allows to consider both linear and rank-based similarities in variables’ distributions, which leads to a smaller number of clusters found and, consequently, provides more effective filtering. ] 

Comments 19: [21) Lines 519-523 - please consider including the R package used.]

Response 19: [mlr3 package is cited two lines lower. Featureless model for reference performance assessment is one of its core functionalities.] 

Comments 20: [22) Lines 524-525 – these procedures are for all methods or for those in lines 519-523?]

Response 20: [For all machine-learning-related methods, including lines 519-523 (as part of models training, tuning and performance assessment)] 

Comments 21: [23) Line 528 – why 5-fold?]

Response 21: [Formally, we can say, that the five folds provide the minimal number of groups of exactly the same size (515 / 5 = 103) in our case. But generally, there is no strict rules or common methodology for evaluation of the objectively best number of folds. There is common rule of thumb, that the more data you have, the less folds in CV you need, but the exact number of folds is just a matter of available computing resources and researcher’s taste. Five folds for CV is commonly used default value, and it is suits well for our sample size, so it is nothing specific here.] 

Comments 22: [24) Line 565 – why 20 repetitions?]

Response 22: [Same as with CV-folds. Nothing specific, just a balance between time and computing resources to obtain enough data for statistically sound analysis of performance metric distributions.] 

Comments 23: [25) Lines603-607 – the name of the orders should be in italics.]

Response 23: [Done.]

Comments 24: [26) Table 2, 3 – sample size refers to the number of plots? Also, please define sample fraction.]

Response 24: [Tables’ footers were updated. E.g., Table 2 footer: Table designations: FG – class IDs for the generalized classification variant, FD – class IDs for the detailed classification variant, Sample size – number of field plots per class, Sample fraction – percentage of the total number of field plots per class. Complete syntaxon names are given in Table S2 of Supplementary Materials.] 

 Comments 25: [27) hi-res or Hires?]

Response 25: [In the context of our article both abbreviations generally refer to the same data, by have different functionality. HiRes used only in tables as a label for the initial set of high spatial resolution variables, and hi-res used only in the main text as a widely accepted abbreviation for "high resolution” data (so, it is less specific). ] 

Comments 26: [28) Lines 882-906 – the discussion of the results needs to be confronted with published references.]

Response 26: [Two relevant references are added.] 

 

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript with the title “Floristic vs. dominant classification approach applied to geospatial modelling of mixed and broadleaf forest types in the North-Western Caucasus (Russia)” deals with modelling forest use with machine learning regression. The manuscript is well structured but needs further details on the methods, and discussion (see comments). Also, English of all text needs to be revised, the parts of the text that might result in a biased interpretation of the manuscript are referred in the comments. Therefore, major changes are recommended.

 

Comments

1) Lines 50-58 – Please include references.

2) Line 127 – please consider using the scientific name of the species.

3) Lines 138-151 – objectives should be better described as well as the research questions.

4) Line 167 – reference of NaturalEarth is missing

5) Lines 175-194 – English is not clear. Please revise.

6) Line 205 – systematic rout survey or systematic sampling by transect?

7) Line 212 – how was cover measured? Please include further details in the text.

8) Line 241 – expertly judge or expert analysis?

9) Line 244 – why 40%? Please include further details in the text.

10) Line 266 – why Phi>30? Please include further details in the text.

11) Line 275 – projective canopy cover or crown horizontal projection? Also, the percent value of the crown horizontal projection is usually termed crown cover.

12) Lines 296-303 - English is not clear. Please revise.

13) Line 317 – how was normalisation done?

14) Lines 321-341, 392, 404, 413, 420, 432 – the references of the software and image products are missing.

15) Lines 350-352 – it is not clear how many images were used and the dates of the images would be of importance.

16) Lines 353-362 – more details are needed.

17) Line 413 – the names of the 19 bioclimatic variables should be included in the text.

18) Lines 443-448 – English is not clear. Please revise.

19) Lines 464-474 – English is not clear. Please revise.

20) Lines 472-473 – why did the authors used Pearson and Spearman correlation coefficients? More details in the text are needed.

21) Lines 519-523 -  please consider including the R package used.

22) Lines 524-525 – these procedures are for all methods or for those in lines 519-523?

23) Line 528 – why 5-fold?

24) Line 565 – why 20 repetitions?

25) Lines603-607 – the name of the orders should be in italics.

26) Table 2, 3 – sample size refers to the number of plots? Also, please define sample fraction.

27) hi-res or Hires?

28) Lines 882-906 – the discussion of the results needs to be confronted with published references.

Comments on the Quality of English Language

English needs to be revised

Author Response

We would like to express our sincere gratitude to the Reviewer for the time and attention given to our manuscript. All major edits are highlighted in the updated manuscript file, and answers to your questions are provided below.

Comments 1: [The paper is very interesting and well written. The scientific problems are presented in a sound way and the goal of the research is clear and of scientific relevance. The most important point which I see needs improvement is the justification of using machine learning. this is a very powerful tool, but its use implies some requirements, and also justification in relation to the advantages in comparison to "traditional" statistic.]
Response 1: [The utilized machine learning methods, namely, Random Forest and CatBoost, are well-known and widely-used for their robustness and versatility without the specific data requirements and assumptions. Also, we used the "classical" LDA and kNN methods for comparison. The size of the used reference sample, while not really huge, but still sufficient for more complex methods of model training. And, while being more complex, the used machine learning methods are entirely based on "traditional" statistic - they just use the fully automated randomized procedures for decision trees' training. So, we are confident that there is no urgent need for additional justification of the methods used.]
Comments 2: [Also, a deeper discussion of the ecological phenomena shaping the results is welcome. You test many variables, an explanation why these were selected is welcome.]
Response 2: [We just used the all-available relevant geospatial data for our study area to identify the most useful combinations (in terms of potential model accuracy) for model training. As we can see, such approach provides the meaningful results, as the optimal variable sets for different variants of forest type classification are turn out to be very different, and, somewhat, unexpected.]
Comments 3: [Finally, study limitations such as having a single study area should be better acknowledged.]
Response 3: [The Discussion section was updated with the specific subsection for the study limitations.]      

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revision has resulted in a clearer and better-structured manuscript. However, several issues still need to be addressed:

1. The abstract remains rather dense and descriptive. The presentation of results and implications is still too general.

2. In the introduction, the text states an objective rather than a testable hypothesis.

3. In the Methods section, some empirical choices are still insufficiently justified — for instance, the use of a 29-day window instead of 30 (see Comment 12) and the 0.95 variance threshold.

4. In the Discussion, although the structure is now subdivided, the text remains mainly descriptive. There is little critical analysis explaining why the floristic models perform better or what the theoretical implications of this finding are.

5. The conclusion is still too brief; it does not explicitly restate the methodology or summarize the main quantitative results.

6. Additionally, while the main matrices are placed in the appendix, the figures remain overly dense and difficult to read.

Author Response

Thank you for the recommendations. We have edited and improved some paragraphs in the text (the renewed lines are marked with blue fill). And other notes we will take into account in our further work.

Comments 1: [The abstract remains rather dense and descriptive. The presentation of results and implications is still too general]

Response 1: [Since the prescribed size of the abstract is quite limited, we do not see possibilities to include more details in it. As the article is largely methodological in nature, it is important to reflect in the abstract the most principal methodological problems and approaches to solving them].

Comments 2: [In the introduction, the text states an objective rather than a testable hypothesis]

Response 2: [This article presents the results of the early stage of development of geospatial model and forest mapping, when only the objective of the work is clear, and due to this any testable hypotheses can be formulated so general that looks trivial (e.g. “The reliability of geospatial modeling strongly depends on the initial classification of the reference data used for model training”). The concrete hypotheses and specifying questions may appear only after finishing this stage, whereas the desire to generate hypothesis obligatorily at the early stage only leads to them being adjusted to the result obtained. Nevertheless, we suppose the experience obtained at the stage is important enough to present it in a special article, even without a specific hypothesis in the base of the study]

Comments 3: [In the Methods section, some empirical choices are still insufficiently justified — for instance, the use of a 29-day window instead of 30 (see Comment 12) and the 0.95 variance threshold].

Response 3: [The explanations were expanded. As for 0.95 threshold, it is a widely applied level that does not need a special explanation, we suppose. In the article, the same threshold is also applied in other statistical procedures used]

Comments 4: [In the Discussion, although the structure is now subdivided, the text remains mainly descriptive. There is little critical analysis explaining why the floristic models perform better or what the theoretical implications of this finding are].

Response 4: [In the Section 4.4, we discuss that a similar result was obtained in another (plain) region. However, this is not yet enough to theoretically explain the advantages of floristic classification in geospatial modeling, and research in other regions is needed to discuss this issue more thoroughly]

Comments 5: [The conclusion is still too brief; it does not explicitly restate the methodology or summarize the main quantitative results].

Response 5: [In the earlier revised version of the Section 5 (Conclusions), the main methodological approach was indicated that is the comparison of two different vegetation classifications in their suitability for geospatial modelling purposes. At this, the main model results were formulated qualitatively (in both classification, the generalized level demonstrates higher model accuracy than the detailed level; and in total, floristic classification approach yields in higher model accuracy than dominant approach). We suppose the incomplete quantitative details of our complex results won’t be informative enough, whereas the complete details will overload the Conclusion section that should be a retelling of results]

Comments 6: [Additionally, while the main matrices are placed in the appendix, the figures remain overly dense and difficult to read]

Response 6: [Now, we have Figures 2-4 enlarged to the size that the page allows. The origin graphic files are up-scaled too]

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The second version of the manuscript answered most of the questions of the reviewer and clarified the text. It is recommended to be accepted for publication.

Comments on the Quality of English Language

English needs to be revised

Author Response

Thank you for the recommendations