Identifying European Old-Growth Forests using Remote Sensing: A Study in the Ukrainian Carpathians
Round 1
Reviewer 1 Report
This paper evaluates the utility of a decision tree classifier to identify forest types in an Ukrainian landscape. The general context of the work is interesting as accurate evaluations of the geographic extent of old-growth forests and the distinction of forest types using remote sensing are still required. The forestry and remote sensing communities could be particularly interested in this study. However, major important improvements must be addressed in this study:
In the introduction, the motivation of this paper is relatively clear: there is need to identify old-growth forest using remote sensing in the studied landscape. It is missing:
1) What has been done to address this issue elsewhere in Ukraine, Europe, or the globe. The authors mention other studies in Europe and the USA. In many of the discussed studies they mention also the imagery used. They fail to detail what those studies found, how successful or accurate they were in their attempts to map old-growth forests, and which classifiers were used. This information is important to properly contextualize what is novel in this study. Based on that background the authors need to more clearly position their work in regards to the gaps on knowledge this study attempts to fill: using Sentinel-2 images and a decision tree classifiers to identify old-growth forests.
2) At the end of the introduction they provide three objectives. Objective one and two are unnecessary and could actually be considered just part of background preliminary information for any classification. Why would the remote sensing and forestry community want to know the spectral signatures of these forest types? Please, justify. Otherwise, this could be assumed to be background information to provide context about the necessity of using an advanced classifier to classify the landscape, not objectives themselves. Moreover, if these forest types can be differentiated by means of spectral signatures, why is an advanced classification approach necessary? Objective three if stated differently provides the basis for this study. I would suggest adding the word “classifying” and in general expand this objective to “mapping old-growth forest”, not only just to testing whether or not RF is capable of classifying the training data. I would add details about the classification objectives. For example, “This study attempts to differentiate old-growth forest of different tree species (mention the species) and also old-growth forest of mixed from other forest types.” This is not clear until later in the paper.
In methods,
3) Study area is completely missing. A section with a description of the study area should be provided. The descriptions provided with no sub-heading at the beginning of section 2 should be included in a study area section, including geographic location, and some description of the area such as topography, land-use history, etc. It is critical to mention dominant land-cover types in the study area to understand the difficulties of classifying this landscape. Also, a better description of forest types is need including tree species composition within old-growth forest and other forest types. The authors mention throughout the manuscript and in Figure 1 the class NOGF (Non-old-growth-forest). Make sure to provide a detailed description of what the NOGF encompasses. Use the information in Table 1 to describe the study area and provide details of the landscape. I would suggest adding an insert map in Figure 1 that provides details of the location of the study area in Ukraine. Substantial parts of the information provided in section 2.2 should be included in a description of the study area. Keep section 2.2 strictly as a description of the land-cover types used in classification.
4) In the description of the RF approach (Section2.4), explain clearly: the number of RF models conducted, the number of land-cover classes evaluated in each of the models, and the number of predictors in each of the models. It should be clear if the response variable is various types of OLF (characterized by different tree species) or OLF (of one or more tree species) vs. NOGF (and what this class encompasses). The reader should understand by the end of this section how many RF models are expected to be discussed in results.
5) I am not sure Table 2 (labeled incorrectly as a second Table 1 in the text) is needed. I would use most of the information in this table to describe the study area in a newer 2.1 section. In Table 3, add NOGF as a class. NOGF is used in a classification, so it is important to know the number of polygons (or the training data used) as well as all the details provided for the other forest classes.
6) The current section 2.2 has information that is more descriptive of the study area and of the forests to be analyzed. You should clearly state in this section the number of old-growth forests analyzed in this study and the NOGF class.
7) Section 2.2: Explain / justify why a buffer of 2 km was used.
8) I would suggest combining the information of Table 1 and of the six vegetation indices utilized into a single table that describes all predictor variables used in the RF models. Mention also in this table that predictors included variables from two seasons and the textural indices used. A table of predictors could make section 2.4 clearer.
9) Justify why texture was calculated of only bands B3, B8, and B12 and why you used a 5X5 window. The authors need to provide a strong argument as to why texture on only these predictors was calculated. RF could accommodate as many texture predictors as there are spectral predictors. It would only use those important in classification. In other words, having texture variables of every predictor would only benefit the models, so the authors need to justify why calculating a subset of texture indices is better. Alternatively, including textural indices of all predictors could be used, potentially improving accuracy values.
10) The authors need to justify why they want to test a series of RF models with subsets of predictors. If testing the importance of incorporating seasonal variables and texture in their models is important, they should state it clearly as an objective. RF is a robust technique that is able to handle high numbers of predictors without overfitting the model. Therefore, it is difficult to understand why starting a classification with a subset of predictors and gradually adding them is important (particularly if the basic set of predictors comprise just bands). With the entire set of predictors, RF would use just those that are important and leave the less important ones unused. It is slightly different in regards to seasonality (i.e. images from two dates) because testing the importance of seasonality implies having various images. In Europe, I would believe it is relatively well acknowledged that incorporating seasonality in any classification approach is important because it is a seasonal landscape in a temperate zone. However, how critical it is to have images of various seasons for this landscape could be informative. If similar accuracy is achieved with single and multi-seasonal models, then the authors could conclude that no images of different seasons are needed. Alternatively, the authors could conclude that to identify properly old-growth forest in this landscape it is critical to incorporate images of different seasons. I think the authors need to justify the use of a subset of predictors in their models if they want to keep the methods and results as they are. Otherwise, including all predictors is a more standard procedure in RF classification assessments. I would suggest having three models: summer, autumn, and summer + autumn only. For each, the authors could provide a variable importance graph (or table) that detail, for instance, mean decrease accuracy of predictors in each model. In that way, they could assess the importance of each of predictors in a single model. The comparison of accuracies and confusion matrices of these three models could also serve as a means to evaluate the importance of using images of multiple seasons to classify old-growth forests in this landscape. The information of Table S1 could be used for this purpose and the authors should consider having that information in the main text.
11) The authors need to justify why they are using images from just summer and autumn. Is imagery from winter and spring not available? Or do these two seasons provide the most contrasting phenological periods of the studies forests?
In results,
12) I don’t understand why the authors do not provide land-cover maps to show the performance of RF to classify the studied landscape. They concentrate in section 3 on discussing results based on the confusion matrix of the training data. RF is a robust classification approach but one must be very careful in using only accuracy of the training data as a means of assessing the performance of the classifier. RF assumes the training data is independent from each other, but in many cases training data is spatially correlated with each other. This means that, in a land-cover map, areas that are not well represented in the training data could tend to be more often misclassified. It is very difficult to completely assess the performance of RF if a land-cover map is not presented and discussed. I strongly suggest including land-cover maps for every model tested to discuss how accurate RF performs. Later in the text the authors present some differences in coniferous forests and explain these differences using elevation. A land-cover map could aid in explaining these results.
13) Figure 2 is unnecessary. It is reasonable to expect that in a seasonal landscape, such as one in Ukraine, the spectral signatures of forest classes will be different between summer and autumn. This figure does not provide any information critical to understanding classification results using RF. Likewise, the information provided between lines 216 and 228 is not the result of an analytical process.
14) I suggest changing Figure 3 to a table to more clearly see the differences in accuracies for OGF for each species analyzed and across models.
15) Table 4 provides information for just one of the RF models. The same information should be provided for every RF model to be able to properly compared models.
16) Figure 4 does not provide any information of importance to understand the performance of RF in classifying the studied landscape. I suggest deleting it. As previously mentioned, spectral signatures are not the result of any method and should not be treated as a main result.
17) A series of t-tests are reported in results. Nowhere in methods is it mentioned that t-tests will be used for any analysis. Moreover, the authors should revise the purpose and use of t-tests. I don’t think they are comparing means of two groups and therefore this analysis is not correct. Please explain / justify.
18) Figure 5 should be deleted. This is information that could be of importance to discuss differences between land-cover maps, but it is not a result of any analysis detailed in the methods section. If land-cover maps are produced and reported and if there is a tendency in the maps to misclassify some land-cover classes in areas of certain elevation, then the authors could use changes in forest structure as a function of elevation to explain the misclassification.
19) The information on Figure 6 would be more clear if presented in a table format or a scatterplot. In a table, authors could actually provide accurate numbers. In a scatterplot, they could plot accuracy as a function of number of predictors. In any case, change the name of each season to: summer, autumn and combined / bi-seasonal / summer + autumn. Specify the dates just in the text, in description of the data.
20) The information on section 3.3 should be incorporated into the previous section, where single-season results are already discussed.
21) Section 3.4 should be deleted; the visual inspection of an image is not a process that can be used to analytically assess human impact on old-growth forests. If the authors provide land-cover maps for their models, they could use them to assess changes in forest cover potentially conducted by people.
In conclusions,
22) The conclusion section is too long and repeats information already provided in results. I suggest shortening this section by about 50% to highlight just the main points that resulted from this study.
23) All through the text make sure you spell Random Forest with caps, it is a proper name of a classification technique.
24) The authors should consider adding the following literature:
Lu, D.; Weng, Q. A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens. 2007, 28, 823–870.
Breiman, L. Random Forest. Mach. Learn. 2001, 45, 5–32.
Freeman, E. A.; Moisen, G. G.; Coulston, J. W.; Wilson, B. T. Random forests and stochastic gradient boosting for predicting tree canopy cover: comparing tuning processes and model performance. Can. J. For. Res. 2016, 46, 323–339.
Cutler, D. R.; Edwards, T. C.; Beard, K. H.; Cutler, A.; Hess, K. T.; Gibson, J.; Lawler, J. J. Random forests for classification in ecology. Ecology 2007, 88, 2783–2792.
25) The authors should revise various citations where only the title and URL is provided. An author / institution / etc., should be provided.
Comments for author File: 
Comments.pdf
Author Response
Referee #1
This paper evaluates the utility of a decision tree classifier to identify forest types in an Ukrainian landscape. The general context of the work is interesting as accurate evaluations of the geographic extent of old-growth forests and the distinction of forest types using remote sensing are still required. The forestry and remote sensing communities could be particularly interested in this study. However, major important improvements must be addressed in this study:
In the introduction, the motivation of this paper is relatively clear: there is need to identify old-growth forest using remote sensing in the studied landscape. It is missing:
1) What has been done to address this issue elsewhere in Ukraine, Europe, or the globe. The authors mention other studies in Europe and the USA. In many of the discussed studies they mention also the imagery used. They fail to detail what those studies found, how successful or accurate they were in their attempts to map old-growth forests, and which classifiers were used. This information is important to properly contextualize what is novel in this study. Based on that background the authors need to more clearly position their work in regards to the gaps on knowledge this study attempts to fill: using Sentinel-2 images and a decision tree classifiers to identify old-growth forests.
Response 1- We have added information fully detailing what was done in these previous studies – what satellite imagery was used, where the study was carried out and the form of classification used. Furthermore, we have expanded our discussion of the classification accuracies these studies obtained in the Discussion section of the paper where they can be directly contrasted with our own paper. Finally, we have expanded the section explaining what is different and novel about our approach compared to previous studies.
2) At the end of the introduction they provide three objectives. Objective one and two are unnecessary and could actually be considered just part of background preliminary information for any classification. Why would the remote sensing and forestry community want to know the spectral signatures of these forest types? Please, justify. Otherwise, this could be assumed to be background information to provide context about the necessity of using an advanced classifier to classify the landscape, not objectives themselves. Moreover, if these forest types can be differentiated by means of spectral signatures, why is an advanced classification approach necessary? Objective three if stated differently provides the basis for this study. I would suggest adding the word “classifying” and in general expand this objective to “mapping old-growth forest”, not only just to testing whether or not RF is capable of classifying the training data. I would add details about the classification objectives. For example, “This study attempts to differentiate old-growth forest of different tree species (mention the species) and also old-growth forest of mixed from other forest types.” This is not clear until later in the paper.
Response 2- We thank the referee for these important suggestions. We have changed the objectives as suggested.
In methods,
3) Study area is completely missing. A section with a description of the study area should be provided. The descriptions provided with no sub-heading at the beginning of section 2 should be included in a study area section, including geographic location, and some description of the area such as topography, land-use history, etc. It is critical to mention dominant land-cover types in the study area to understand the difficulties of classifying this landscape. Also, a better description of forest types is need including tree species composition within old-growth forest and other forest types. The authors mention throughout the manuscript and in Figure 1 the class NOGF (Non-old-growth-forest). Make sure to provide a detailed description of what the NOGF encompasses. Use the information in Table 1 to describe the study area and provide details of the landscape. I would suggest adding an insert map in Figure 1 that provides details of the location of the study area in Ukraine. Substantial parts of the information provided in section 2.2 should be included in a description of the study area. Keep section 2.2 strictly as a description of the land-cover types used in classification.
Response 3: We were remiss not to include more information on the study site. We have modified our paper to account for all these suggestions. We have introduced a ‘Study Site’ section as suggested. We have added to the description of the area, including references, to past landuse and current pressures on forests. We have moved information on tree species composition from the OGF survey area to study site section. We have added an inset map of Ukraine and coordinates to Fig 1.
4) In the description of the RF approach (Section2.4), explain clearly: the number of RF models conducted, the number of land-cover classes evaluated in each of the models, and the number of predictors in each of the models. It should be clear if the response variable is various types of OLF (characterized by different tree species) or OLF (of one or more tree species) vs. NOGF (and what this class encompasses). The reader should understand by the end of this section how many RF models are expected to be discussed in results.
Response 4 – We agree that this section was confusingly worded. We therefore have rewritten this section, taking care to explicitly set out the number of polygons used in RF classification and the number of RF models considered.
5) I am not sure Table 2 (labeled incorrectly as a second Table 1 in the text) is needed. I would use most of the information in this table to describe the study area in a newer 2.1 section. In Table 3, add NOGF as a class. NOGF is used in a classification, so it is important to know the number of polygons (or the training data used) as well as all the details provided for the other forest classes.
Response 5: We thank the referee for this comment. We have added to our description of the study area, as described above. However, we prefer to retain Table 2 as we feel this provides the reader with useful information and context. Table 3 provides information on NOGF as suggested.
6) The current section 2.2 has information that is more descriptive of the study area and of the forests to be analyzed. You should clearly state in this section the number of old-growth forests analyzed in this study and the NOGF class.
Response 6 – We move this descriptive content to the new ‘Study Site’ section. We have rewritten this section to very clearly state which and how many OGF/NOGF polygons are used in the RF models.
7) Section 2.2: Explain / justify why a buffer of 2 km was used.
Response 7 – This was chosen as the minimum distance that allowed the required number of NOGF polygons to be selected. We have added “This distance was chosen as it enabled the requisite number of appropriately sized NOGF polygons to fit in.”
8) I would suggest combining the information of Table 1 and of the six vegetation indices utilized into a single table that describes all predictor variables used in the RF models. Mention also in this table that predictors included variables from two seasons and the textural indices used. A table of predictors could make section 2.4 clearer.
Response 8: We thank the reviewer for this comment. We have rewritten section 2.4 to make it clearer as suggested. With this revised section 2.2 we feel that Table 1 is clearer left as it is.
9) Justify why texture was calculated of only bands B3, B8, and B12 and why you used a 5X5 window. The authors need to provide a strong argument as to why texture on only these predictors was calculated. RF could accommodate as many texture predictors as there are spectral predictors. It would only use those important in classification. In other words, having texture variables of every predictor would only benefit the models, so the authors need to justify why calculating a subset of texture indices is better. Alternatively, including textural indices of all predictors could be used, potentially improving accuracy values.
Response 9 – We intentionally selected one band in the visible, one band in the NIR and one band in the SWIR to sample from the full spectral range for a minimal computational demand. Bands within the visible, NIR and SWIR are quite highly correlated with each other, so selecting additional bands would not add much information to our analysis. We add this explanation of our choice of textural features and window size to the paper.
10) The authors need to justify why they want to test a series of RF models with subsets of predictors. If testing the importance of incorporating seasonal variables and texture in their models is important, they should state it clearly as an objective. RF is a robust technique that is able to handle high numbers of predictors without overfitting the model. Therefore, it is difficult to understand why starting a classification with a subset of predictors and gradually adding them is important (particularly if the basic set of predictors comprise just bands). With the entire set of predictors, RF would use just those that are important and leave the less important ones unused. It is slightly different in regards to seasonality (i.e. images from two dates) because testing the importance of seasonality implies having various images. In Europe, I would believe it is relatively well acknowledged that incorporating seasonality in any classification approach is important because it is a seasonal landscape in a temperate zone. However, how critical it is to have images of various seasons for this landscape could be informative. If similar accuracy is achieved with single and multi-seasonal models, then the authors could conclude that no images of different seasons are needed. Alternatively, the authors could conclude that to identify properly old-growth forest in this landscape it is critical to incorporate images of different seasons. I think the authors need to justify the use of a subset of predictors in their models if they want to keep the methods and results as they are. Otherwise, including all predictors is a more standard procedure in RF classification assessments. I would suggest having three models: summer, autumn, and summer + autumn only. For each, the authors could provide a variable importance graph (or table) that detail, for instance, mean decrease accuracy of predictors in each model. In that way, they could assess the importance of each of predictors in a single model. The comparison of accuracies and confusion matrices of these three models could also serve as a means to evaluate the importance of using images of multiple seasons to classify old-growth forests in this landscape. The information of Table S1 could be used for this purpose and the authors should consider having that information in the main text.
Response 10: We thank the referee for this comment. As suggested, we have modified the objectives to highlight that a key aspect of our study is to test the importance of including textural features in the models. As suggested we have added graphs of mean decrease accuracy to the Supplementary Figures for the most accurate classification models(SI Fig 2 and 5.)
11) The authors need to justify why they are using images from just summer and autumn. Is imagery from winter and spring not available? Or do these two seasons provide the most contrasting phenological periods of the studies forests?
Response 11 – We did look at using winter and spring images – but in the Carpathian mountains extensive snow cover covered much of the forest from December through to April. Many of the polygons on the higher ground were completely white and provided little useful information. We add a line to the paper to explain our decision.
In results,
12) I don’t understand why the authors do not provide land-cover maps to show the performance of RF to classify the studied landscape. They concentrate in section 3 on discussing results based on the confusion matrix of the training data. RF is a robust classification approach but one must be very careful in using only accuracy of the training data as a means of assessing the performance of the classifier. RF assumes the training data is independent from each other, but in many cases training data is spatially correlated with each other. This means that, in a land-cover map, areas that are not well represented in the training data could tend to be more often misclassified. It is very difficult to completely assess the performance of RF if a land-cover map is not presented and discussed. I strongly suggest including land-cover maps for every model tested to discuss how accurate RF performs. Later in the text the authors present some differences in coniferous forests and explain these differences using elevation. A land-cover map could aid in explaining these results.
Response 12: Thank you for this important suggestion. We produce a land-cover map for all features and combined summer and autumn imagery for both OGF species and OGF/NOGF classification (Figs. 2a, b and Figs. 4a, b). We add into Discussion a paragraph related to the land cover maps, including how OGF tree species and elevation affects model performance.
13) Figure 2 is unnecessary. It is reasonable to expect that in a seasonal landscape, such as one in Ukraine, the spectral signatures of forest classes will be different between summer and autumn. This figure does not provide any information critical to understanding classification results using RF. Likewise, the information provided between lines 216 and 228 is not the result of an analytical process.
Response 13 – Figure deleted as suggested.
14) I suggest changing Figure 3 to a table to more clearly see the differences in accuracies for OGF for each species analyzed and across models.
Response 14: We have now included this information in tabular form as Supplementary Table SI1.
15) Table 4 provides information for just one of the RF models. The same information should be provided for every RF model to be able to properly compared models.
Response 15: We have changed this Table to provide information on the most accurate model.
16) Figure 4 does not provide any information of importance to understand the performance of RF in classifying the studied landscape. I suggest deleting it. As previously mentioned, spectral signatures are not the result of any method and should not be treated as a main result.
Response 16- Deleted as suggested
17) A series of t-tests are reported in results. Nowhere in methods is it mentioned that t-tests will be used for any analysis. Moreover, the authors should revise the purpose and use of t-tests. I don’t think they are comparing means of two groups and therefore this analysis is not correct. Please explain / justify.
Response 17- We accept that we should have explained this analysis in the Methods section. We are using t-tests to test for significant differences between the mean band values of the different tree species and later, the OGF and NOGF spectra. We have added a line to the Methods section explaining our analysis.
18) Figure 5 should be deleted. This is information that could be of importance to discuss differences between land-cover maps, but it is not a result of any analysis detailed in the methods section. If land-cover maps are produced and reported and if there is a tendency in the maps to misclassify some land-cover classes in areas of certain elevation, then the authors could use changes in forest structure as a function of elevation to explain the misclassification.
Response 18- We add a short description of this analysis to the methods section. While we see the reviewer’s point, and accept that this figure is not strictly necessary for Random Forest classification, we wish to keep this figure in the paper. We feel that the our finding that OGF conifer stands were lighter than NOGF stands is an interesting result, and is in contrast to previous work (Cohen). We feel our explanation for why we found this – the high altitude conifer OGF was more open than NOGF – is also of interest.
19) The information on Figure 6 would be more clear if presented in a table format or a scatterplot. In a table, authors could actually provide accurate numbers. In a scatterplot, they could plot accuracy as a function of number of predictors. In any case, change the name of each season to: summer, autumn and combined / bi-seasonal / summer + autumn. Specify the dates just in the text, in description of the data.
Response 19: We provide this information in tabular form in Supplementary Information Table SI 2 and 3. We change the season names as requested.
20) The information on section 3.3 should be incorporated into the previous section, where single-season results are already discussed.
Response 20: Incorporated into previous section
21) Section 3.4 should be deleted; the visual inspection of an image is not a process that can be used to analytically assess human impact on old-growth forests. If the authors provide land-cover maps for their models, they could use them to assess changes in forest cover potentially conducted by people.
Response 21: Deleted this section as suggested.
In conclusions,
22) The conclusion section is too long and repeats information already provided in results. I suggest shortening this section by about 50% to highlight just the main points that resulted from this study.
Response 22: As suggested we have shortened significantly, especially the first portion of Conclusion section.
23) All through the text make sure you spell Random Forest with caps, it is a proper name of a classification technique.
Response 23: changed to “Random Forest” throughout
24) The authors should consider adding the following literature:
Lu, D.; Weng, Q. A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens. 2007, 28, 823–870.
Breiman, L. Random Forest. Mach. Learn. 2001, 45, 5–32.
Freeman, E. A.; Moisen, G. G.; Coulston, J. W.; Wilson, B. T. Random forests and stochastic gradient boosting for predicting tree canopy cover: comparing tuning processes and model performance. Can. J. For. Res. 2016, 46, 323–339.
Cutler, D. R.; Edwards, T. C.; Beard, K. H.; Cutler, A.; Hess, K. T.; Gibson, J.; Lawler, J. J. Random forests for classification in ecology. Ecology 2007, 88, 2783–2792.
Response 24: we have added these citations to the Random Forest section
25) The authors should revise various citations where only the title and URL is provided. An author / institution / etc., should be provided.
Response 25: Added author names to these citations
Author Response File: 
Author Response.docx
Reviewer 2 Report
Paragraf 66
Insert Scientific Hypothesis: There is a significant difference between old OGF and NOGF forests on the remote sensing of identification.
Paragraf 73 and 74
Line spacng
Paragraf 86
Spacing between words
Paragraf 91
I do not think it's good to start a sentence with a number
Paragraf 264
New page
Paragraf 267
It is necessary that the text referring to Figure 4 is before the chart.
This is similar to Figure 5 (First you need to write the text and then display the Figure).
Author Response
Reviewer #2
Paragraf 66
Insert Scientific Hypothesis: There is a significant difference between old OGF and NOGF forests on the remote sensing of identification.
Response – we add “based on the hypothesis that there is a significant difference between the spectra of OGF and other forest types (non-Old-Growth Forest.)”
Paragraf 73 and 74
Line spacng
Response- altered
Paragraf 86
Spacing between words
Response - Changed to remove excessive space between words in Sentinel 2 images section.
Paragraf 91
I do not think it's good to start a sentence with a number
Response- Changed to “The 20m resolution…”
Paragraf 264
New page
Response: Changed
Paragraf 267
It is necessary that the text referring to Figure 4 is before the chart.
This is similar to Figure 5 (First you need to write the text and then display the Figure).
Response – Figure 4 has been removed as per the request of another reviewer. Figure 5 (and all other Figures and Tables) shifted to below their mention in the text.
Reviewer 3 Report
In general, I found the paper to raise an important research topic, showing the relevance of using satellite remote sensing in forestry and conservation. The objective of the study is supported by the method used, and the results presented and discussed. The combination of field based dataset, sentinel-2 images, and use of random forest algorithm for analysis and classification purposes could provide an exemplary approach for other forest monitoring and conservation studies. Despite its merits, there are however several opportunities for improving the manuscript that the authors should take into consideration. Comments and suggestions were divided into two categories: i) main comments – which include some general and overarching issues that should be improved and, ii) a set of specific comments that relate to some finer details, corrections, questions or suggestions.
1. General comments
Improve the manuscript in making it concise and with a proper flow, especially in the introduction and results section, as well as the flow between the sections. The results should be made to go more in-line with the objectives, which is not the case in this version of the manuscript. For example the objective states nothing about results presented on 3.4 which is about anthropogenic disturbance monitoring on OGF.
Improve the use of punctuations, citation to tables/figures, caption of tables/figures, spelling out abbreviations on captions, use of parenthesis etc. These should be checked throughout the manuscript.
The conclusion should also be improved, as in the current state most of it (line 369-387) reads more like a discussion. Please summarize your main findings in few sentences and indicate on general directions (line 387-406 reads fine for conclusion).
2. Specific comments
- L27: Old-growth forests (OGFs)..
But in the rest of the manuscript you used OGF. Please be consistent
- L21: 3.3%, 2.1%, and 1.8% respectively
- L36: describe hat the causes are for continuous loss of OGF in Europe
- L46-44 contradicts with the statement on L53. Please rephrase
- L55-59: comes in between the story of using remote sensing for OGF identification. Better o combine it with Paragraph 2 or have it as Par. 3
- L70: Determine spectral differences between OGF and “other forests”.. – better to introduce non old growth forests (NOGF) here
- L74-210 please keep the method section consistent. Now it goes from Sentinel-2 images -> OGF survey data -> Sentinel-2 image evaluation -> random forest method
Better to restructure to: 2.1. Study site (description of the site with map) -> 2.2. OGF survey data -> 2.3. Sentinel-2 images and derivable, 2.4. Random forest method
- L90: These images were then terrain and atmosphere corrected.. -> were then topographically and atmospherically corrected
- L91: correct to: “the 20 m resolution bands of S2 were...”
- L92: remove “(see below)”
- L94: Which surveys did you use for this analysis. From 2010-2017? Please mention it explicitly
- L106: (....ha) of OGF were also recorded
- L111: There is no table 2.. Correct the caption of the second table
- L129: “.. and young forest shows up as very brightly in the images”: please provide reference for this statement
- L133: “As in practice OGF often completely ringed mountains, NOGF was typically lower in altitude and lacked high montane forest.”: please check the formation of this sentence
- L141: spell out the abbreviations used on the legend, as well as OGF and NOGF. The figures should be able to stand alone
- L142: please also spell out Near IR
- L162: “(A paper [42] gives a detailed overview of the GLCM process.)” please remove brackets and rephrase the sentence.
- L162: Generally,
- L179- 180: justify why Random forest was used. E.g. Suitability to your large datasets?
“selected for its ease of use”? can you provide references here?
- L192: Caption should be Table 2
- L209-210: “Since it is fairly easy to accurately identify forest as either broadleafed, mixed or evergreen” : can you please provide a reference
- L212: Please label the figures as (a), (b) on top, provide a label for the Y-axis of the second figure as well.
- L220: 140?
- L230: Mention abbreviations used on the legend as well
- L254: “Table 5” -> “Table 4”
- L275-282: provide references
- L296-297: remove ()
- L317-318: can you say more on textural information.. e.g. in relation to shadows/vegetation structure ? Comparison with other studies? Difference between seasons?
- L320: The caption is not clear. “...for various features..”?
- L352: This section should be merged with 3.2
- L357: This section comes as a surprise. Nothing was mentioned about it in the abstract and/or the objectives. Please combine it as discussion in the other sections and part of it to the conclusion
- L368-387: sounds like a discussion. Please summarize it in few sentences.
- L368-406: Please restructure the conclusion: towards presenting the summary of the study and the way forward
Supplementary files: the figures were mentioned before the tables.. so please readjust, please check on spelling out the abbreviations as well.
Author Response
Referee #3
In general, I found the paper to raise an important research topic, showing the relevance of using satellite remote sensing in forestry and conservation. The objective of the study is supported by the method used, and the results presented and discussed. The combination of field based dataset, sentinel-2 images, and use of random forest algorithm for analysis and classification purposes could provide an exemplary approach for other forest monitoring and conservation studies. Despite its merits, there are however several opportunities for improving the manuscript that the authors should take into consideration. Comments and suggestions were divided into two categories: i) main comments – which include some general and overarching issues that should be improved and, ii) a set of specific comments that relate to some finer details, corrections, questions or suggestions.
We thank the reviewer for their positive comments on our paper.
1. General comments
Improve the manuscript in making it concise and with a proper flow, especially in the introduction and results section, as well as the flow between the sections. The results should be made to go more in-line with the objectives, which is not the case in this version of the manuscript. For example the objective states nothing about results presented on 3.4 which is about anthropogenic disturbance monitoring on OGF.
Improve the use of punctuations, citation to tables/figures, caption of tables/figures, spelling out abbreviations on captions, use of parenthesis etc. These should be checked throughout the manuscript.
The conclusion should also be improved, as in the current state most of it (line 369-387) reads more like a discussion. Please summarize your main findings in few sentences and indicate on general directions (line 387-406 reads fine for conclusion).
Response – We thank the reviewer for his detailed and informative comments. We remove Section 3.4, which is to be the subject of a different paper. We have rewritten the first half of the Conclusions section to make it more concise. We spell out all abbreviations in captions.
2. Specific comments
- L27: Old-growth forests (OGFs)..
But in the rest of the manuscript you used OGF. Please be consistent
Response: changed to Old-growth forest
- L21: 3.3%, 2.1%, and 1.8% respectively
Response: deleted space.
- L36: describe hat the causes are for continuous loss of OGF in Europe
Response: Added “from deforestation and conversion to managed plantations” to Line
- L46-44 contradicts with the statement on L53. Please rephrase
Response: Added “mostly” to L45
- L55-59: comes in between the story of using remote sensing for OGF identification. Better o combine it with Paragraph 2 or have it as Par. 3
Response: Reordered lines so that Sentinel 2 paragraph now directly follows
- L70: Determine spectral differences between OGF and “other forests”.. – better to introduce non old growth forests (NOGF) here
Response – added mention of non-Old-Growth Forest here
- L74-210 please keep the method section consistent. Now it goes from Sentinel-2 images -> OGF survey data -> Sentinel-2 image evaluation -> random forest method
Better to restructure to: 2.1. Study site (description of the site with map) -> 2.2. OGF survey data -> 2.3. Sentinel-2 images and derivable, 2.4. Random forest method
Response: Restructured as suggested
- L90: These images were then terrain and atmosphere corrected.. -> were then topographically and atmospherically corrected
Response: Corrected
- L91: correct to: “the 20 m resolution bands of S2 were...”
Response: corrected
- L92: remove “(see below)”
Response: Removed
- L94: Which surveys did you use for this analysis. From 2010-2017? Please mention it explicitly
Response: Added “and covered the survey years 2010-2017 inclusive”
- L106: (....ha) of OGF were also recorded
Response: added “also”
- L111: There is no table 2.. Correct the caption of the second table
Response: Corrected
- L129: “.. and young forest shows up as very brightly in the images”: please provide reference for this statement
Response: added citation (Wulder et al, 2004)
- L133: “As in practice OGF often completely ringed mountains, NOGF was typically lower in altitude and lacked high montane forest.”: please check the formation of this sentence
Response – Changed to “Much of the OGF consisted of high altitude forest stretching up to the treeline. The neighbouring NOGF was therefore typically downhill from the OGF, and consequently was at a lower elevation and lacked high montane forest.”
- L141: spell out the abbreviations used on the legend, as well as OGF and NOGF. The figures should be able to stand alone
Response: added the explanations of abbreviations for OGF, NOGF and S2 (we’ve done this for all figures.)
- L142: please also spell out Near IR
Response: Corrected
- L162: “(A paper [42] gives a detailed overview of the GLCM process.)” please remove brackets and rephrase the sentence.
Response: Changed to “A detailed overview of GLCM can be found here.”
- L162: Generally,
Response: Added comma
- L179- 180: justify why Random forest was used. E.g. Suitability to your large datasets?
“selected for its ease of use”? can you provide references here?
Response: Added 2 references to support ease of use and added a further cited justification for its selection – high classification accuracy.
- L192: Caption should be Table 2
Response- Changed the captions of the tables
- L209-210: “Since it is fairly easy to accurately identify forest as either broadleafed, mixed or evergreen” : can you please provide a reference
Response – changed to “RF classification was carried out separately for each of these different forest types”- this section has been extensively rewritten as per the comments of another reviewer.
- L212: Please label the figures as (a), (b) on top, provide a label for the Y-axis of the second figure as well.
Response- Figure deleted as per request of other reviewer. Other relevant figures labelled (a),(b),(c).
- L220: 140?
Response – added “%” after this number
- L230: Mention abbreviations used on the legend as well
Response – Figure deleted as per request of other reviewer
- L254: “Table 5” -> “Table 4”
Response - corrected
- L275-282: provide references
Response – provided 2 references (Crist et al. (1986) and Kimes et al. (1986))
- L296-297: remove ()
Response: Removed brackets from line “Open areas were generally..”
- L317-318: can you say more on textural information.. e.g. in relation to shadows/vegetation structure ? Comparison with other studies? Difference between seasons?
Response- Add brief discussion of reasons for differences between OGF and NOGF. We compare with the one study (Cohen) we could find that also calculated textural features for OGF (albeit not GLCM.) and one that calculated textural features for young and mature forest (Ozdemir).
- L320: The caption is not clear. “...for various features..”?
Response: Changed to “for five selected models”
- L352: This section should be merged with 3.2
Response: Merged with previous section
- L357: This section comes as a surprise. Nothing was mentioned about it in the abstract and/or the objectives. Please combine it as discussion in the other sections and part of it to the conclusion
Response: We have deleted this section – we intend to present a separate paper on OGF forest loss.
- L368-387: sounds like a discussion. Please summarize it in few sentences.
Response – We have heavily edited the first section of Conclusions, and consolidated the Paragraphs so as to only mention the important points.
- L368-406: Please restructure the conclusion: towards presenting the summary of the study and the way forward
Response: We have removed the final paragraph of the conclusions, and with the above edits we hope it is more concise.
Supplementary files: the figures were mentioned before the tables.. so please readjust, please check on spelling out the abbreviations as well.
Response – abbreviations spelled out as requested. Order of figures and tables rearranged to reflect order of appearance in text.
Round 2
Reviewer 1 Report
The authors conducted a detailed revision of the paper and responded to all of the major concerns I listed in my report. Many of my suggestions have been incorporated in the new version and those that were not have been justified. I, therefore, have no additional major revisions to suggest. I recommend some minor improvements, including:
1. Cite Figure 1 at the end of section 2.1. You could refer just to the newly provided inset map only if you prefer, e.g. (Figure 1, inset map).
2. Along with the previous suggestion, add a very short explanation of the inset map to the Figure 1 legend, e.g.
Figure 1. Relief map of the study site showing Old-Growth Forest (OGF, shown as red polygons) and 280 Non-Old Growth Forest (NOGF, shown as blue polygons.) S2 image shows the extent of the Sentinel-281 2 image used in the study. Inset map shows the location of the study area within Ukraine.
I would suggest adding the word “Ukraine” within the polygon of the country in the inset map.
3. Section 3.1. I appreciate that the authors now provide a land cover map. I am still a bit concerned about the first paragraph of this section which spends too much time describing spectral signature differences as a means of distinguishing old-growth forests of different tree species. This seems not to be of concern to the other reviewers, so I would just suggest making sure this description of spectral signatures fits well within the context of a land-cover classification using RF and that the paragraph flows well without having to quote too much information that is provided as supplemental material.
4. Line 962: add “forests” between the words “usually” and “lower”
Author Response
The authors conducted a detailed revision of the paper and responded to all of the major concerns I listed in my report. Many of my suggestions have been incorporated in the new version and those that were not have been justified. I, therefore, have no additional major revisions to suggest. I recommend some minor improvements, including: Response: We wish to thank the reviewer for all their comments, and for the time and care taken in reviewing our manuscript, both in this and the previous revision. We feel that the changes have greatly improved our paper. 1. Cite Figure 1 at the end of section 2.1. You could refer just to the newly provided inset map only if you prefer, e.g. (Figure 1, inset map). Response 1: Inserted as requested. 2. Along with the previous suggestion, add a very short explanation of the inset map to the Figure 1 legend, e.g. Figure 1. Relief map of the study site showing Old-Growth Forest (OGF, shown as red polygons) and 280 Non-Old Growth Forest (NOGF, shown as blue polygons.) S2 image shows the extent of the Sentinel-281 2 image used in the study. Inset map shows the location of the study area within Ukraine. I would suggest adding the word “Ukraine” within the polygon of the country in the inset map. Response 2: Added to Figure caption and changed Fig 1 as requested. 2. Section 3.1. I appreciate that the authors now provide a land cover map. I am still a bit concerned about the first paragraph of this section which spends too much time describing spectral signature differences as a means of distinguishing old-growth forests of different tree species. This seems not to be of concern to the other reviewers, so I would just suggest making sure this description of spectral signatures fits well within the context of a land-cover classification using RF and that the paragraph flows well without having to quote too much information that is provided as supplemental material. Response 3: We have shortened this section by cutting the last three lines in Section 3.1, first paragraph. 4. Line 962: add “forests” between the words “usually” and “lower” Response 4: Inserted as requested.