Forest Burned Area Detection Using a Novel Spectral Index Based on Multi-Objective Optimization

Round 1

Reviewer 1 Report (Previous Reviewer 5)

This is the second version of the manuscript that i have already reviewed hence my comments are based also on the review and the responses we had already. 

1) Reg NBR formula, either you apply the correction or call it with another name. It is fine to use an index with whatever form you like or fits with you but to modify a published index and keep calling it with its initial name in my opinion is not correct. I understand that the NBR does not fit with your assumption of positive coefficient of B12 but that is it. NBR is (B8-B12)/(B8+B12). For my point of view either you change it to the correct form and all the resulted images or do not call it NBR but with another name i.e. modified NBR or whatever (https://www.fs.usda.gov/rm/pubs_series/rmrs/gtr/rmrs_gtr164/rmrs_gtr164_13_land_assess.pdf) 

2) Regarding the visual interpretation of burned areas, authors need to justify in the manuscript how did the deal with surface fires that are not visible under the crown

3) Finally, regarding their response to my comment about matlab command, I really appreciate their recommendation about ScenceDirect and how to search for a topic. After 27 years in Academia sector I really enjoy learning new things. However, since I will not be the only reader of the publication and this is not a technical report/manual, I would suggest them to avoid software commands and then I would advise them to click their link, pick an appropriate reference and insert it in their text. The reference number 29 [Rouse, J. W.; Haas, R. H.; Schell, J. A.; Deering, D. W. Monitoring vegetation systems in the Great Plains with ERTS. Proceedings 668 of the Third ERTS Symposium, Freden, S. C., Becker. M., Eds. NASA SP-351, National Aeronautics and Space Administration, 669 Washington, D.C.] in line 337 about integer programming algorithm in not relevant.

 

 

Author Response

Responses to Referee 1:

First of all, we would like to take this chance to thank you for your insightful comments, which would help us to significantly improve our paper. In light of your comments, we have revised the entire paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comment point:

This is the second version of the manuscript that i have already reviewed hence my comments are based also on the review and the responses we had already. 

1. Reg NBR formula, either you apply the correction or call it with another name. It is fine to use an index with whatever form you like or fits with you but to modify a published index and keep calling it with its initial name in my opinion is not correct. I understand that the NBR does not fit with your assumption of positive coefficient of B12 but that is it. NBR is (B8-B12)/(B8+B12). For my point of view either you change it to the correct form and all the resulted images or do not call it NBR but with another name i.e. modified NBR or whatever

(https://www.fs.usda.gov/rm/pubs_series/rmrs/gtr/rmrs_gtr164/rmrs_gtr164_13_land_assess.pdf).

RE: Thanks for your suggestion. We have changed the NBR into Modified NBR (MNBR) as described in Line 221-223.

 

2. Regarding the visual interpretation of burned areas, authors need to justify in the manuscript how did the deal with surface fires that are not visible under the crown.

RE: As we know, remote sensing determines the surface parameters by measuring the characteristics of electromagnetic radiations coming from the Earth's surface. The sensor may not be able to detect the fire event (after the fire is extinguished) if the surface fire does not trigger a crown fire since the dense crowns can prevent the sensor from detecting the fire event. In this case, our method and other spectral indices-based methods may be ineffective because they are built based on the spectral reflectance of different land covers.

Moreover, we have presented a brief interpretation in the manuscript to discuss this problem in Line 526-529. The detected burned area is slightly less than that of the actual field investigation. One possible reason is that the surface fires in some places may not be detectable to optical images due to dense tree crowns, which unavoidably produces the omission error.

3. Finally, regarding their response to my comment about matlab command, I really appreciate their recommendation about ScenceDirect and how to search for a topic. After 27 years in Academia sector I really enjoy learning new things. However, since I will not be the only reader of the publication and this is not a technical report/manual, I would suggest them to avoid software commands and then I would advise them to click their link, pick an appropriate reference and insert it in their text.

RE: We have added some references in Line 347-348 in the revised manuscript on mixed-integer linear programming algorithms according to Matlab documents. The references were also listed as follows:

[34] Carl-Henrik Westerberg, Bengt Bjorklund, and Eskil Hultman, “An Application of Mixed Integer Programming in a Swedish Steel Mill.” Interfaces February 1977 Vol. 7, No. 2 pp. 39–43

[35] Cornuéjols, G. Valid inequalities for mixed integer linear programs. Mathematical Programming B, Vol. 112, pp. 3–44, 2008

[36] Nemhauser, G. L. and Wolsey, L. A. Integer and Combinatorial Optimization. Wiley-Interscience, New York, 1999.

 

4. The reference number 29 [Rouse, J. W.; Haas, R. H.; Schell, J. A.; Deering, D. W. Monitoring vegetation systems in the Great Plains with ERTS. Proceedings 668 of the Third ERTS Symposium, Freden, S. C., Becker. M., Eds. NASA SP-351, National Aeronautics and Space Administration, 669 Washington, D.C.] in line 337 about integer programming algorithm in not relevant.

RE: The sequence of references has been updated. The correct reference should be No. 33.

Author Response File: Author Response.docx

Reviewer 2 Report (Previous Reviewer 1)

The necessary corrections have been made. Only the clarity of the graphs in Figure 3 and Figure 6 can be improved.

Author Response

Responses to Referee 2:

First of all, we would like to take this chance to thank you for your insightful comments, which would help us to significantly improve our paper. In light of your comments, we have revised the entire paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comment point:

 

• The necessary corrections have been made. Only the clarity of the graphs in Figure 3 and Figure 6 can be improved.

RE: Thanks for your suggestion. We have re-plotted Figures 3, 4 and 6 for clarity in the revised manuscript.

Author Response File: Author Response.docx

Reviewer 3 Report (Previous Reviewer 3)

Line 48-51: ... and numerous methods including spectral band segmentation, spectral indices, image supervised classification and deep neural network have been attempted to delineate burned areas over the last decades. Please add some references relative to each method. Also please review your paper and add enough references.

Author Response

Responses to Referee 3:

First of all, we would like to take this chance to thank you for your insightful comments, which would help us to significantly improve our paper. In light of your comments, we have revised the entire paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comment point:

 

• Line 48-51: ... and numerous methods including spectral band segmentation, spectral indices, image supervised classification and deep neural network have been attempted to delineate burned areas over the last decades. Please add some references relative to each method. Also please review your paper and add enough references.

RE: Thanks for your suggestion. The references relative to each method were specified in Line 51-52. Moreover, nine additional references have been reviewed and added to the revised manuscript, bringing the total number of references to 50.

Author Response File: Author Response.docx

Reviewer 4 Report (Previous Reviewer 4)

The authors have already significantly improved the manuscript.

However, there are still several minor corrections to be made.

First, there are typing errors that were not present in the previous version of the manuscript. For example, the numbers that refer to affiliations are not in  superscript as they should be. There are also spacing errors.

Scientific (Latin) names of tree species are not correctly written.

Correct:

Pinus tabuliformis Carriere

Populus alba L.

Larix Mill.

Pinus yunnanensis Franch.

Prunus spinosa L.

Acacia confusa Merr.

Quercus palustris Münchh.

Names of genera and species should be in italic, but the rest (name of  author) is not in italic.

As for Larix Mill, there is only the name of the genus.

The authors have already expanded the reference list, but they could add some more references.

Author Response

Responses to Referee 4:

First of all, we would like to take this chance to thank you for your insightful comments, which would help us to significantly improve our paper. In light of your comments, we have revised the entire paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comment point:

• The authors have already significantly improved the manuscript.However, there are still several minor corrections to be made.

RE: Thanks for your encouragement.

• First, there are typing errors that were not present in the previous version of the manuscript. For example, the numbers that refer to affiliations are not in superscript as they should be. There are also spacing errors.

RE: Thanks, we have checked the entire paper and made corrections in terms of the typing errors and grammar mistakes in the revised manuscript.

• Scientific (Latin) names of tree species are not correctly written.

Correct: Pinus tabuliformis Carriere, Populus alba L., Larix Mill., Pinus yunnanensis Franch., Prunus spinosa L., Acacia confusa Merr., Quercus palustris Münchh.

Names of genera and species should be in italic, but the rest (name of author) is not in italic. As for Larix Mill, there is only the name of the genus.

• RE: The names of tree species are corrected in the revised manuscript in light of the comment, and they are also listed below.

“Pinus tabuliformis Carriere” has been corrected to “Pinus tabuliformis Carrière”;

“Populus alba L.” has been corrected to “Populus alba L.”;

“Pinus yunnanensis Franch.” has been corrected to “Pinus yunnanensis Franch.”;

“Prunus spinosa L.” has been corrected to “Prunus spinosa L.”;

“Acacia confusa Merr.” has been corrected to “Acacia confusa Merr.”;

“Quercus palustris Münchh.” has been corrected to “Quercus palustris Münchh.”;

“Larix Mill.” has been corrected to “Larix decidua Mill.”;

• The authors have already expanded the reference list, but they could add some more references.

RE: Some related references have been reviewed and added, and the total number of references reaches to 50 in the revised manuscript.

Author Response File: Author Response.docx

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments for author File: Comments.pdf

Author Response

First of all, we would like to take this chance to thank you for your insightful comments, which have helped us to significantly improve our paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comments point by point: 1.the method of the study have been extensively discussed, but the results are limited. Re: In revised version, we tried to enhance the results. 2.you can use wildfires instead of forest fires Re: corrected. 3.Added a reference to line 68, pp2 Re: Have added 4.In reviewing the references related to the introduction, mention the case studies that have used this index and their general results. (line 87) Re: I’ m sorry I can not catch what’s your meaning. Would you please to clarify it. 5. table 1 can be moved to the appendix Re: Considering table 1 provides the band information, which helps to understand the spectral indices described in Tab.3, we kept the Table 1. 6. It is enough to mention the bands used in this study. Re: I can not understand what’s you meaning? As described in line 240, all the above12 bands are used to build the ABAI index. Here in table 2, we tried to demonstrate the usefulness of the proposed method for different scenarios. 7.In line 133-152, in introducing the three study areas, try to write them in the same way Re: we have revised the paragraph in line to your suggestion, and the maps are re-plotted. 8.You can use color differentiation in this figure Re: this figure has been re-plotted using different colors, and the variances of different land covers were also been given. 9.This assessment paragraph appears to belong to the Materials and Methods section, not the Results section Re: this paragraph has been moved to the Materials and Methods section, thank you. 10.Unlike this table, where the results are clearly separated for all three areas, in figures 5 to 8, at least in their captions, it is not specified these results are related to which of these three areas. Re: we have changed the captions for clarity. 11.The discussion section is written very briefly and the results of this study have not been compared with the results of others. Re: we have rewrote the discussion section to provide more information.

Author Response File: Author Response.docx

Reviewer 2 Report

Review is in the attached PDF file.

Comments for author File: Comments.pdf

Author Response

First of all, we would like to take this chance to thank you for your insightful comments, which have helped us to significantly improve our paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comments point by point: Comment 1: Abstract and Manuscript Title The abstract and the title of the manuscript do not contain keywords related to the subject of this journal. In particular, there are no such important words as forest, or woods. Re: Our motivation is to develop a novel spectral index for post-fire scars detection, which is vital importance to support post-fire management strategies and account for the environmental impact of forest fires. We considered that the context of this paper is related to the subject of this journal. Comment 2: Keywords It is curious to note that mandatory keywords section is missing. Re: keywords have been added in revised version. Comment 3: Section 2.1. Data and Study Area The description of the forest area is presented extremely briefly. The land cover map, forest structure, canopy features, tree height and ages are not specified. It should be noted that the rate of forest fire spread and the degree of forest damage strongly depend on the type of trees growing in the area. The manuscript does not specified not only the dominant growing species of woody vegetation, but also there is no general category of woodland: broad-leaved, coniferous forests or mixed forests. Also note that the resinous forests, such as pine, fir, and spruce forests, as well as the ether-containing forests, such as eucalyptus forests, burn especially strongly. Re: You are right. We have added some information both in data description and final results, such as predominately growing species or general category of woodland, forest fire spread, and fire-affected areas etc. Comment 4: The description of wildfires is also presented too briefly. When the authors describe forest wildfires, the type of fire is not indicated: a crown fire, a ground underbrush fire, or a subsoil (peat) forest fire. Also, the stage of the wildfire at the moment of the Sentinel-2A and 2B satellite flight it is not indicated. Thus, from the manuscript text it is unclear: are the wildfires in the active phase or in the smoldering phase, in which the fallen trunks and stumps burn, or has the fire gone out or has been extinguished? Moreover, it looks strange that in study area 1 (Qinyuan County), the fire started on March 29, 2019, while the Sentinel-2B image corresponds to (2018-06-10), and the second Sentinel-2A image corresponds to (2019-06-10). It is not clear why the authors chose June 10 on orbit R118. The ignition time in study area 3, in Sicily, Italy, is not presented at all. Thus, without a mask of fires, with indication of the period and intensity of wildfires, any interpretation of using indexes NDVI (Normalized Difference Vegetation Index), NBR/NBR+ (Normalized Burn Ratio Plus), BAI/BAIS2 (Burnt Area Index for Sentinel 2), MIRBI (Mid Infrared Burn Index), and ABAI, as indexes related to fire activity, are impossible. It has to be reminded that anomalies in the vegetation land cover, if they occur, may occur due to anthropogenic factors, such as acid rain, or due to biological impact, e.g. a forest can be damaged by bugs or caterpillars. Similarly, dump polygons, coal outcrops, dark rocky slopes, volcanic ashes (Sicily), and eroded peaty soils give anomaly signals. It is impossible to 2 exclude the impact of the local climate, especially in the mountain gorges and coastal zones. To confirm their correctness, the authors should provide the spatial distributions of the studying indexes direct before and after wildfires. Re: You are correct in describing fire ecology, fire processes and fire behaviors. However, it is not our intention to analyze fire behavior in details in this paper. Also, the types of fires and their occurrence factors are beyond the scope of this paper. As a post fire scars detection method, the main contribution of this paper is to proposed a new spectral index for burned area detection after a fire event. We argued that our theme maybe also interesting and important to forest management and restoration. Minor Comment: The Sentinel-2A and Sentinel-2b sensors are allocated on board of slow moving climate satellites. Thus Sentinel-2A, and its partner Sentinel-2B, re-passes over the specified domain after 10 days. The spatial resolution of Sentinel-2 is equal to 10–60 m, depending on the spectral bands. On the other hand, a wildfire can pass 60 m at short temporal interval, which can be estimated in minutes– hour, so the slow-moving “lazy” Sentinel-2A and 2B sensors are not suitable for studying fire processes. Re: You are right. But this paper does not focused on active fire monitoring, and we have no intention to study fire processes. Instead, we focus on the development of the spectral index for burned area detection. In addition, usually the climatic Burned Area Product is distributed with a temporary resolution of a month. That is, within a month, only the three Sentinel-2A images and, accordingly, in addition, three Sentinel-2B images are available to Sentinel Team researchers. Taking into account the smoke and clouds, it is extremely small. Re: we have explained why and how we selected the images in revised version. Moreover, if the Sentinel Team researchers for analyzing land cover can use the combined Sentinel 2A + 2B images, then for analyzing of vegetative and fire dynamics, researchers should use only one of the sensors, otherwise they should conduct cross–satellite validation and prove that changes, for example in NDVI, are caused by processes in the vegetation cover, and not by the decalibration of the satellites themselves. Therefore, the Sentinel-2A and 2B sensors are poorly suited for the study of fire processes, and the statements given in the Abstract and underlined below are not entirely correct. Abstract “Mapping burned area and burned severity are the very basic information for modeling the impact of fires on ecosystem dynamics. Rapid and accurate extractions of fire-affected areas are thus of vital importance to support post-fire management strategies and account for the environmental impact of fires. The availability of high spatial and temporal resolution remotely sensed images enables us to develop effective algorithms for promptly extracting burned areas.” Re: we have no intention to study fire processes in this paper. We think may be the abstract mislead the reviewer, so we revised the abstract.

Author Response File: Author Response.docx

Reviewer 3 Report

1)       Line 22: Please add there some relevant references.

2)       Lind 45-49: please add some examples, especially the studies that used spectral indices to delineate burned areas and then in the next step highlight your innovations.

3)       Line 52: … several studies. Yes, they are many studies. So please add more references.

4)       Line 61: why the construction of spectral indices is a cost-effective method. You used the freely available remote sensing products and nowadays, access to fire reference datasets is no longer difficult and there are various sources that can be easily used.

5)       Which new advancements in remote sensing image acquisition? Landsat started in 1972 and has always been used. Needs additional explanation.

6)       Table 2: Why is the acquisition date of your datasets so diverse? Is there a specific reason? Please explain.

7)       Line 136: … close to 60%.  Add a reference.

8)       Line 132-152: Why did you decide to do research in three separate areas? Does it work well for evaluating how well your strategy works? please explain. 

9)       Figure 3: Did you consider standard deviation? Based on previous studies, it can provide good information about spectral seperabillity. I suggest you also show the standard deviation in your chart or generate a new chart based on it.

10)   Line 322-325: Yes, the green band can help you to detect the water content and maybe enhance spectral separability between burned areas and water bodies. But how about other LULC classes? Such as bare land and buildings. Based on figure 4 you have 3 classes with a high level of similar spectral properties including burned area, road, and building. How about them?

11)   Line 346: How do you generate your samples?

12)   What reference dataset do you use for accuracy assessment?

13)   Line 379-380: … that the proposed index significantly outperforms other spectral indices. Did you use a statistical test? How did you get a significant difference?

14)   Line 408-410: It can be observed that all the methods can extract most of the burned areas, but some clouds and shadows are mixed to the burned area for the NBR+, and especially for BAI, as highlighted with green circles. It’s a good topic. Please consider these results in your discussion.

15)   Discussions and Conclusions should be split into two separate sections.

Author Response

First of all, we would like to take this chance to thank you for your insightful comments, which have helped us to significantly improve our paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comments point by point: 1.Line 22: Please add there some relevant references. RE: We have added some relevant references in the paper, but not in Line 22, due to Line 22 is in abstract section. 2.Lind 45-49: please add some examples, especially the studies that used spectral indices to delineate burned areas and then in the next step highlight your innovations. RE: We have highlighted the innovations. 3.Line 52: … several studies. Yes, they are many studies. So please add more references. RE: some references have been added. 4.Line 61: why the construction of spectral indices is a cost-effective method. You used the freely available remote sensing products and nowadays, access to fire reference datasets is no longer difficult and there are various sources that can be easily used. RE: the use of spectral indices for burned area detection is cost-effective, because this method only requires simple band algebraic algorithm. Compared to supervised classification methods, they need adequate training samples. However, collection of training samples is cost and laborious. We have not used remote sensing products, but the free datasets. The information of burned areas was extracted from our method. Our aim is to provide option tool for related researchers for particular applications. 5. Which new advancements in remote sensing image acquisition? Landsat started in 1972 and has always been used. Needs additional explanation. RE: This sentence was removed do to it possible misleads to readers. In fact, satellite data has indeed achieved thorough achievement compared to Landsat started in 1972 in terms of spatial, temporal and spectral resolutions. Remote sensing has seen new platforms that cover different levels of observation, from high to low altitudes. Levels of observation have included satellite data, aerial data, including aircraft data, unmanned aerial vehicles (UAV) data, and ground-based observation (e.g., geophysical equipment). 6.Table 2: Why is the acquisition date of your datasets so diverse? Is there a specific reason? Please explain. RE: The acquisition date is mainly selected according to the image quality and available satellite data after the active fire event extinguished. Since the selected datasets are undertaken the aim of validating the effectiveness of the developed method, there is no specific reason for acquisition date. 7. Line 136: … close to 60%.  Add a reference. RE: The description of used data has drastically changed, and a reference was added. 8.Line 132-152: Why did you decide to do research in three separate areas? Does it work well for evaluating how well your strategy works? please explain.  RE: As a newly developed index, it is important to test its generalized ability. Therefore, we tried to demonstrate the usefulness of the proposed method for different scenarios. Also refer to #6 9.Figure 3: Did you consider standard deviation? Based on previous studies, it can provide good information about spectral seperabillity. I suggest you also show the standard deviation in your chart or generate a new chart based on it. RE: You are right that a standard deviation will provide more information for spectral seperability. According to your comments, we have re-plotted this figure. 10.Line 322-325: Yes, the green band can help you to detect the water content and maybe enhance spectral separability between burned areas and water bodies. But how about other LULC classes? Such as bare land and buildings. Based on figure 4 you have 3 classes with a high level of similar spectral properties including burned area, road, and building. How about them? RE: This is a quite insightful comment. The main motivation of this work is to develop a new spectral index, such that we can separating burned areas from other LULC classes. As can be found in our method section, we do try to accomplish this ambitious aim by splitting burned land from other 7 typical covers. However, we have no intention for selecting special band, where the included bands is just the results derived from multi-objective optimizing problem. In other words, the involved bands and its formulation of the spectral index construction is a data-driven method. An intuitive explanation for the results maybe water is the most confused material to burned land, and accordingly green band to be the most informative band according to the optimizing results. In order to split other LUCC, an alternative method may firstly mask out water, and we remove the corresponding objective function of water in Eq. 4. 11.Line 346: How do you generate your samples? RE: the samples are stratified randomly selected. 12.What reference dataset do you use for accuracy assessment? RE: we used visual interpretation as reference due to the actual datasets are not available. 13.Line 379-380: … that the proposed index significantly outperforms other spectral indices. Did you use a statistical test? How did you get a significant difference? RE: From the quantitative assessments shown in tab.5, we can find that the overall accuracy and kappa of the ABAI are much superior to other methods. A statistical test based on the McNamara’s test was implemented to test the significant difference. 14.Line 408-410: It can be observed that all the methods can extract most of the burned areas, but some clouds and shadows are mixed to the burned area for the NBR+, and especially for BAI, as highlighted with green circles. It’s a good topic. Please consider these results in your discussion. RE: We have added it to the discussion section to enhance the results according to your suggestion. Many thanks. 15.Discussions and Conclusions should be split into two separate sections. RE: They have been split into two separate sections

Author Response File: Author Response.docx

Reviewer 4 Report

The list of references is too short. Accordingly, the introduction and especially the discussion must be expanded.

The figures are not entirely clear (eg, Figure 1 - position of study areas in wider geographical areas missing).

The results are too preliminary, as the authors themselves state in the Discussions and Conclusions.

The method requires more testing on a regional and a global scale.

Author Response

First of all, we would like to take this chance to thank you for your insightful comments, which have helped us to significantly improve our paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comments point by point: 1.The list of references is too short. Accordingly, the introduction and especially the discussion must be expanded. Re: we have added some references in revised version, and the introduction and discussion were expended. 2.The figures are not entirely clear (eg, Figure 1 - position of study areas in wider geographical areas missing). Re: the figures are re-plotted for clarity. 3.The results are too preliminary, as the authors themselves state in the Discussions and Conclusions. Re: we have revised the Discussions and Conclusions 4.The method requires more testing on a regional and a global scale. Re: It is our interesting to validate the method for a regional and a global scale. However, it is difficult to complete this work within the limited revised time span. Validations on a regional or global scale will be done in future work. Thank you.

Reviewer 5 Report

I find the topic in keeping with the research interests of the journal. The authors propose a new index for burned area identification named Analytical Burned Area Index called ABAI based on Sentinel-2 images. They approach the development of ABAI as a multi-objective linear optimization problem based on the values of band ratios. The whole application i.e. development of newly is quite interesting, innovative and worth publishing, however, I have some major remarks that should be addressed.

I found that the literature review on mapping burned area from remote sensing images is limited while a lot of recent (i.e. the last 5 years) pubs on the topic are missing.

Regarding the Sentinel-2 data authors do not mention the level of processing of their data. I assume they downloaded L2 (BOA) since this is available. However, they mention both B8 and B8A in their dataset even if B8 is omitted in L2 due to overlapping bandwidth and the lower resolution. Please explain if you chose L1 why is that, or if you chose L2 why you kept B8? If the latter is the case, did you perform radiometric correction only to this band?

The authors review some of the main indices and they conclude lines 181-183: “It can be found from the existed spectral indices shown in Table 3 that the common law is the B12 (SWIR 2) is larger or equal to 0, and so the band SWIR 2 is used as the primary band for subsequent processing”

Firstly, all bands’ values (i.e. reflectance) are larger or equal to 0 (probably not equal το 0 unless is a blackbody). Maybe authors want to say something else but in a wrong way.

Assuming that you want to indicate the usefulness of B12(SWIR) in burned area indices, in my opinion the extensive literature the last 2 decades have already proven the usefulness of SWIR in burned area mapping and that is due to the spectral response of the burned vs unburned area presented extensively in the very well know graph i.e https://un-spider.org/sites/default/files/Spectral_responses.jpg. That means that developed indices are based on the above reflectance response and not vice versa.

At table 3 NBR function is wrong. The correct is (NIR-SWIR) / (NIR +SWIR). Your reference [15] also describe the NBR function in a wrong way. Hereinafter, all NBR images are wrong (inverted sign)

Section 2.2.2. Reflectance analysis

It is unclear a) Which method did authors used to extract these classes and what is the accuracy of their method. For which image is the figure 3?

Lines 237-238 “As a result, it is difficult to use two bands to construct an effective burned land index”, I personally disagree with that statement given the widely usage and performance of NBR and other indices.

Lines 242-245: “Since the burned area may present similar reflectance characteristics as the water and cloud, and even the building and bare land as reported in the literature, it is usually necessary to apply a mask to remove the water bodies and clouds for the detection of burned areas. However, it is difficult to obtain accurate templates to mask different land covers.”

Please explain. There are numerous ways to mask water i.e. NDWI or clouds i.e. QI of L2 Sentinel-2 accurately

With the Eq 1 authors make an assumption that the algebraic sum of bands results to an index reflecting the burned area. Once again they used the B12 given that it is larger or equal to 0. see my comment above.

It's totally unclear to me why the sum should be greater than 0 only for burnt areas and less than 0 for the rest of the classes.

Lines 306-307: “Using the function ‘intlinprog’ in Matlab optimization toolbox, we can solve the multi-objective optimizing problem”.

I do not have Matlab or any other blackbox and I am used to develop my code. How can I replicate you method? What is this “intlinprog’.

The section 2.2.4 are results. Move them to the correct section

Lines 336-339:” Consequently, the pixels related to the clouds in the resulting ABAI image present certainly low (negative) values and therefore would not be confused with those related to the burned area, which presents high values and tends to be white.”

This is what you aimed with optimization. It is not a conclusion from the graph. The graph is the result from the optimization. From my point of view according to figure 4 burned area is well discriminated with NBR* and NBR+. It is the threshold that changes. In ABAI is 0 because you asked for it.

*Be aware that NBR graph should be the opposite according to my previous comment about the error in NBR

In the results section authors describe the accuracy metrics and how we collect reference data however they do not describe how they collected their reference data. it seems that they haven't collected any reference. They state about “…the referred map…” [correct to reference map] however they do not describe how this have been derived.  If it has derived by visual interpretation then how did they managed the fire to different vegetation strata i.e. surface fires?

At lines 384-386: “The extracted results by threshold segmentation to distinguish the burned area and non-burned area can be obtained, as shown in Fig.6, and the assessment of each resulting burned area map can be made by means of visual inspection” How this thresholds have been derived for each index? It is arbitrary? Usually when we have a binary classification, we plot the sensitivity and specificity for different thresholds and select the threshold where specificity and sensitivity intersects.  

Line 386: “Obviously, the classification result derived from the proposed index is the most similar to the referred map…” According to Figure 6, it is not clearly, thus “obviously” can not be used.

In general and according to the visual comparison the issues that authors raises and propose the new index opposed to the previous it is the discrimination of the clouds and the water however this targets can be identified easily with other indices developed for this specific targets.

Finally, authors perform a quantitative analysis based on the sample of 400 points which I assume they derived them visually. Once again, the results show that ABAI has the best performance, based on the missinterpratation from the other indices of the clouds and water. This is quite expected due to the nature of the index where each class including water and clouds are included in the optimization. However, if the clouds and the water have been masked firstly, the results would be significantly different.

Author Response

First of all, we would like to take this chance to thank you for your insightful comments, which have helped us to significantly improve our paper. To ensure that the revised parts are clear, all the changes have been highlighted in red. The following summarizes how we have addressed your comments point by point: 1.I find the topic in keeping with the research interests of the journal. The authors propose a new index for burned area identification named Analytical Burned Area Index called ABAI based on Sentinel-2 images. They approach the development of ABAI as a multi-objective linear optimization problem based on the values of band ratios. The whole application i.e. development of newly is quite interesting, innovative and worth publishing, however, I have some major remarks that should be addressed. Re：we are appreciate your positive comments. 2.I found that the literature review on mapping burned area from remote sensing images is limited while a lot of recent (i.e. the last 5 years) pubs on the topic are missing. Regarding the Sentinel-2 data authors do not mention the level of processing of their data. I assume they downloaded L2 (BOA) since this is available. However, they mention both B8 and B8A in their dataset even if B8 is omitted in L2 due to overlapping bandwidth and the lower resolution. Please explain if you chose L1 why is that, or if you chose L2 why you kept B8? If the latter is the case, did you perform radiometric correction only to this band? Re: L1C datasets were used for experiments. A preprocessing procedure has been conducted with the plugin tool ‘Sen2Cor’ and SNAP software, including radiometric correction/atmospheric correction and resampled spatial resolution to 10m. 3.The authors review some of the main indices and they conclude lines 181-183: “It can be found from the existed spectral indices shown in Table 3 that the common law is the B12 (SWIR 2) is larger or equal to 0, and so the band SWIR 2 is used as the primary band for subsequent processing” Firstly, all bands’ values (i.e. reflectance) are larger or equal to 0 (probably not equal το 0 unless is a blackbody). Maybe authors want to say something else but in a wrong way. RE: I think maybe the sentence mislead you. Our actual meaning is the coefficient in front of the B12 (SWIR 2) is larger than 0 or equal to 0 (i.e., the band 12 is not included). We have clarified this sentence. 4.Assuming that you want to indicate the usefulness of B12(SWIR) in burned area indices, in my opinion the extensive literature the last 2 decades have already proven the usefulness of SWIR in burned area mapping and that is due to the spectral response of the burned vs unburned area presented extensively in the very well know graph i.e https://unspider.org/sites/default/files/Spectral_responses.jpg. That means that developed indices are based on the above reflectance response and not vice versa. Re: there are many evidences of the usefulness of B12(SWIR) in burned area indices in literatures from the perspective of experimental and physical views. Here we’d like to induce the same conclusion from math statistical view. In fact, this is not a important issue, we think that we can obtain the similar result by selecting any band as denominator to conduct band ratio operation. 5.At table 3 NBR function is wrong. The correct is (NIR-SWIR) / (NIR +SWIR). Your reference [15] also describe the NBR function in a wrong way. Hereinafter, all NBR images are wrong (inverted sign) Re: You are absolutely right that the NBR function is wrong. Here for facilitating to compare different indices, we modified the formulation by adjusting its sign, such that if the pixel was affected by fire, its index increased as other spectral indices. We considered that ‘wrong’ of the reference [15] is for the same reason. 6.It is unclear a) Which method did authors used to extract these classes and what is the accuracy of their method. For which image is the figure 3? Re: The samples were randomly selected from each classes by visual interpretation. To show the possible accuracy of the samples, this figure has been re-plotted using different colors, and the stand deviations of different classes were also been given. 7.Lines 237-238 “As a result, it is difficult to use two bands to construct an effective burned land index”, I personally disagree with that statement given the widely usage and performance of NBR and other indices.2 Re: This maybe an arguable sentence. Indeed, the NBR is a widely accepted index for burned area, especially for early satellite images, but is also often argued not to provide enough accuracy in complex background. If more bands can be applied, there has a trend to include more band for spectral index. 8.Lines 242-245: “Since the burned area may present similar reflectance characteristics as the water and cloud, and even the building and bare land as reported in the literature, it is usually necessary to apply a mask to remove the water bodies and clouds for the detection of burned areas. However, it is difficult to obtain accurate templates to mask different land covers.” Please explain. There are numerous ways to mask water i.e. NDWI or clouds i.e. QI of L2 Sentinel-2 accurately Re: The main motivation of our paper is to develop a framework for constructing spectral index that has ability to suppress certain background LUCC. It should be noted that although we can apply masks to remove some certain land covers before the use of other indices for burned areas detection, e.g. normalized difference water index (NDWI) for the water bodies and normalized difference building index (NWBI) for buildings, this strategy inevitably poses error accumulation because totally accurate masks are impossible. Moreover, the use of NDWI or NDBI for mask increases complexity and uncertainty due to additional threshold to be optimized. Finally, the ABAI is a more conciseness than the method with two-step process (mask first, then process). 9.With the Eq 1 authors make an assumption that the algebraic sum of bands results to an index reflecting the burned area. Once again they used the B12 given that it is larger or equal to 0. see my comment above. It's totally unclear to me why the sum should be greater than 0 only for burnt areas and less than 0 for the rest of the classes. Re: the objective function can be broadly classified two classes. One is the burned areas (including forest fire-affected land, glass fire-affected land, etc), and the other is other LUCC (mixed with water, road, buildings, etc). Therefore, the construction of multi-objective functions is to split them with the maximal class separating capacity. The purpose that we assumed the burnt areas are greater than 0 is to align to other spectral indices, where a brightness pixel to be fire-affected point. Another advantage is to optimize the segmenting threshold to be 0. 10.Lines 306-307: “Using the function ‘intlinprog’ in Matlab optimization toolbox, we can solve the multi-objective optimizing problem”.I do not have Matlab or any other blackbox and I am used to develop my code. How can I replicate you method? What is this “intlinprog’. Re: the ‘intlinprog’ is a function in Matlan to solve the problem of the mixed-integer linear programming. There are several solutions for this problem. If you are interest in this problem, please refer this website for detail. https://www.sciencedirect.com/topics/engineering/mixed-integer linear programming. 11.The section 2.2.4 are results. Move them to the correct section Re: You are right, It has been moved to correct section. 12.Lines 336-339:” Consequently, the pixels related to the clouds in the resulting ABAI image present certainly low (negative) values and therefore would not be confused with those related to the burned area, which presents high values and tends to be white.” This is what you aimed with optimization. It is not a conclusion from the graph. The graph is the result from the optimization. From my point of view according to figure 4 burned area is well discriminated with NBR* and NBR+. It is the threshold that changes. In ABAI is 0 because you asked for it. Re: Yes, both NBR and NBR+ achieved relatively good results in figure 4. The ABAI used 0 as threshold because we asked for it, while the threshold for NBR and NBR+ required to be determined. This is an other advantage of the proposed ABAI for image segmentation. In our experiments, the segmented thresholds for other indices were determined by trial and correction, and the best results were reported in figure 4. 13.Be aware that NBR graph should be the opposite according to my previous comment about the error in NBR Re: please refer to comment #5 14.In the results section authors describe the accuracy metrics and how we collect reference data however they do not describe how they collected their reference data. it seems that they haven't collected any reference. They state about “…the referred map…” [correct to reference map] however they do not describe how this have been derived. If it has derived by visual interpretation then how did they managed the fire to different vegetation strata i.e. surface fires? 15.Re: the reference data were basically obtained by visualized interpretation with the help of multi-temporal datasets. 16.At lines 384-386: “The extracted results by threshold segmentation to distinguish the burned area and non-burned area can be obtained, as shown in Fig.6, and the assessment of 3 each resulting burned area map can be made by means of visual inspection” How this thresholds have been derived for each index? It is arbitrary? Usually when we have a binary classification, we plot the sensitivity and specificity for different thresholds and select the threshold where specificity and sensitivity intersects. Re: In our experiments, the segmented thresholds for other indices were determined by trial and correction, and the best results were reported. According to your suggestion, a sensitive analysis has been added in revised version. 17.Line 386: “Obviously, the classification result derived from the proposed index is the most similar to the referred map…” According to Figure 6, it is not clearly, thus “obviously” can not be used. Re: we have deleted the word 18.In general and according to the visual comparison the issues that authors raises and propose the new index opposed to the previous it is the discrimination of the clouds and the water however this targets can be identified easily with other indices developed for this specific targets. Re: please refer to comment #8 19.Finally, authors perform a quantitative analysis based on the sample of 400 points which I assume they derived them visually. Once again, the results show that ABAI has the best performance, based on the missinterpratation from the other indices of the clouds and water. This is quite expected due to the nature of the index where each class including water and clouds are included in the optimization. However, if the clouds and the water have been masked firstly, the results would be significantly different. Re: You are right if we used a mask as prior informant. It would improve the final results. However, as aforementioned in comment #8, it is still merit to develop new spectral index that has ability to suppress certain noises, such as water, shadows, and clouds, etc.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have made corrections according to the reviewers' recommendations and the new version is accepted. 

Reviewer 2 Report

Manuscript ID: forests-1900020 (resubmitted) Authors: Bo Wu, He Zheng, Zelong Xu, Zhiwei Wu, and Yindi Zhao Titled: “An analytical framework to construct spectral index for burned area detection with multi-objective optimization”
The authors of the manuscript forests-1900020 have improved the manuscript. In this manuscript authors developed and investigated the Analytical Burned Area Index (ABAI), which was proposed for mapping the burned areas from the newly launched Sentinel-2A and 2B satellite images. However, as I already wrote early, the manuscript maintenance does not corresponded subjects of the MDPI Forest Journal. Why is it so? The Title, Abstract and Conclusion of the resubmitted version do not have any mentions about forest or wood category. Further, in the revised version of the manuscript the keywords list is present and has following words: wildfires, burned area mapping, Sentinel 2, multi-objective optimization, spectral indices. These keywords are not the keywords that have to be in case of the studies of forest wildfires. Moreover, the ABAI indexes, which developed in this study, have categories: burned land, bare land, road shadow, vegetation, water, buildings, and cloud; please see e.g. in Figure 3 and Table 4. Thus, vegetation category is present in the major calculation of the ABAI scheme. The vegetation category is common and it is not equal to forest category. In line 522–527 it was written: “Using the ABAI method, we can then estimate the total fire affected areas in the Xichang fire event are about 2811.96ha, which is very close to the actual conflagration area of 3047.78ha released by officers [35]. With the support of local land cover data [24], it can also be inferred that the disaster types are predominately by closed evergreen needle-leaved forest (41%) and grassland (29%).” The reference to [24] is not enough to submit manuscript to MDPI Forest. Once more, I note that this study is relevant and may be of interest to readers. However, the manuscript the forests-1900020 can not be accepted due to formal reasons, since the subject of this study is out of scope of the MDPI Forest Journal. Again, as previously, I recommend to send this manuscript to more specialized MDPI Journals, such as Remote Sensing, Fire or to the special issue of GeoHazards, involved to "Advances in Applied Wildfire Research".

 

Reviewer 4 Report

The authors significantly improved the manuscript. They gave the necessary explanations and revised the conclusions. All that contributed to a significantly better assessment of their work.

However, some details should be improved in the new version as well. Authors have cited additional references, but the number of cited references is still small. Moreover, there is not a single reference in the Discussions chapter.

The authors mention some tree species but only their common names. For example: "oil pine, poplar, and larch" (rows 147 and 148); "Yunnan pine, blackthorn, Taiwan acacia and oak" (rows 168 and 169). The authors should also provide scientific (Latin) names of these tree species.