Mapping Forest Burn Extent from Hyperspatial Imagery Using Machine Learning

Round 1

Reviewer 1 Report

This paper used a Mask Region-based Convolutional Neural Network (MR-CNN) and Support Vector Machine (SVM) to determine trees and burned pixels in a post-fire forest. This overall research scheme is good and provides a method for post-fire identification using high resolution images.

The background theory of the introduction is insufficient, and the background support of the method as the introduction is easy to make others question whether these methods are necessarily the best. In the field of remote sensing, there should be many methods used to identify post-fire pixels. Some methods are chosen to better explain scientifically meaningful problems, such as post-fire identification can provide a good support for forest ecological management. I hope the author will consider this part carefully. It is suggested that the scientific theoretical background should be put first, the method part should be put later, and the overall goal should be presented finally.

Line 355, “(Table 5 and Error! Reference source not found.)”??

It is not recommended to include literature in the conclusion, only introduce the method of this paper, and the comparison part can be put in the discussion part or the method part.

Author Response

See attached document

Author Response File: Author Response.pdf

Reviewer 2 Report

No further comments.

Author Response

See attached pdf

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors clarified all my questions and improved the paper.

Author Response

See attached pdf

Author Response File: Author Response.pdf

Reviewer 4 Report

I reviewed the most probably a revised version of the manuscript for the first time. According to my review, the paper needs some minor modifications before publishing. My specific suggestions are below:

1. Introduction and Background needs further improvements. The Authors directly and only focused on SVM approach, UAV based studies and only mentioned about Landsat satellites in their literature review. However, forest burn area detection and monitoring with RS  has a long history including several pixel or object based classification methods, spectral index based approaches that applied on variety of satellite images some of which are very high resolution. For example Sentinel 2 mission has provided better spatial, spectral and temporal resolution and used recently in several forest fire mapping. Thus I recommend improving the Introduction based on a general review of previous studies related to topic.
2. The flow (actually the paragraph order) of the Introduction should be revised. Introduction will be better if it  flows as, definition of the problem and its effects on Earth, previous attempts to monitor this problem (historic development), the gaps or limitations of these studies and lastly what is the focus of this research in relation with above mentioned limitations (what is  novel or significant in current research)
3. In data collection, it should be enlarged with what is the UAV platforms, data collection altitude - corresponding spatial resolution etc.
4. Figure 5 does not seem  a good flowchart to me, I expect more hierarchial and classical one as they inform better than tihs about the processing steps.

Author Response

See attached pdf

Author Response File: Author Response.pdf

Reviewer 5 Report

General comments:

• Sentences are underlined in multiple places.
• Multiple references with similar or identic content from the main author ([1,3,4,13]).
• How the method would work in the context of non-prescribed mega-fires?
• The authors should consider an Annex or a paragraph explaining more details about the SVM method. At the same level of reproducibility as how the MR-CNN configuration is explained.

Major reviews:

• Reference [1]: Wrong page of the article (p. 107 in the document or p. 95 on the printed page).
• Reference [12]: This reference is from a presentation abstract. Not able to access a paper or more than half a page of explanation.
• Introduction: All fire statistics look to be computed from 2018 data. The authors should consider updating that information with more current numbers and references.
• Section 2.2.1.1: For reproducibility, you should add some content about:
  • The tool used to create the training polygons.
  • The tool used to human-label the polygons.
  • The SVM Kernel and the hyperparameters used.
  • The exact tools from ArcGIS classification that made all that possible.
• Table 1: Use & or AND instead of +. In logic, the + sign normally means OR, not AND.
• Line 208: Rephrase since it could be extended up into the tree crown if the canopy pixel was indeed a burned canopy pixel.
• Table 2: The authors have classified Burned and Unburned Canopy pixels as Canopy and then they are using Canopy as it was Unburned. Clarify why this selection.
• Line 355: Reference is broken.
• Lines 361-364: This explanation does not match the numbers in table 5. A threshold larger than 5600 gives better accuracy and sensitivity while keeping the specificity constant at 91%. 
• Line 414 and 419: Missing the beginning of the sentence.

Minor reviews:

• Line 139: ration -> ratio.
• Line 135: as -> and.
• Lines 210-214: Consider rephrasing, really long sentence.
• Lines 389-391: Reference to (d), (e), and (f) are missing.

 

Author Response

See attached pdf

Author Response File: Author Response.pdf

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

This paper used MR-CNN and SVM to map forest burn extent including burned surface vegetation and sub-crown burn based on Hyperspatial Imagery. The starting point of this work is good. It can accurately locate the fire and its true extent. However, the whole paper is poorly written. The logic is not clear, the research significance is unclear. If the paper is well conceived, it will show unique significance in locating forest fires with high resolution images. It is a pity that the author do not seem to have prepared very carefully.

The abstract and introduction parts lack logic, and the key points are not outstanding, so they need to be sorted out again. The expression in the paper is not concise enough. The authors need to look again to see if there are many other references to the work that can be used in the introduction and discussion section. As far as the whole paper is concerned, the authors are expressing their various ideas without a good comparison with the existing work.

Data selection and processing are not clear enough. The author should articulate each part clearly.

From this whole paper, it can be seen whether the author did not carefully organize the various links of fire mapping. It seems that the whole work is in a hurry.

The expression of the tables and images are not concise.

The discussion section is more self-talk, with little reference to anyone else’s existing work.

References are not recommended in the conclusion.

Reviewer 2 Report

The manuscript proposed a fused method that integrates MR-CNN and SVM to mitigate the misclassifying of the areas under the tree being mapped as unburnt areas. The mapped results are promising this issue using the drone-derived hi-res image. The paper is well organized and written. The paper rarely talked about the parameter-tunning and setting of MR-CNN and SVM, I recommend adding these steps for the audiences to follow your works.

Line59-60, why? A short explanation of why SVM was effective on burnt extent mapping.

Reviewer 3 Report

The manuscript addresses the use of machine learning in UAVs RGB very high spatial resolution images to mapping forest burn extent. The analysis is based on four sites located in the USA and clearly evidences an approach to expand the previous research from authors.

Firstly, the MS has a lack of theoretical background to justify the work and the research method. The research design must be improved adequately to a scientific paper.

Here, I point out some specific comments:

Abstract

Line 18-19. I suggest improving the abstract, adding quantitative information instead of only qualitative information.

Introduction

The introduction does not provide sufficient background. It's not clear for me what is the real need to follow a lot of steps of processing, also considering the highly computational/time-consuming demand to process UAS images and difficulties to implement on a large scale justified only to improve burn extent maps that and provide more detailed information for managers. There are many burn extent maps or hot spots based on orbital images currently (including for forested biomes, such as Amazon) that also provide information for managers and monitoring.

Also, the authors highlight challenges in forested biomes (lines 31-33) but conclude that the method used in the MS faces this problem “An issue that was noticed with some fires was when many unburned tree noise pixels are connected to one another in areas of closed canopy cover” (lines 507-509). It doesn’t seem logical.

Moreover, did the authors compare the burn extension results with high/medium spatial resolution orbital images to see any significant differences in terms of area (rapidEye, planet, Sentinel), and how your research can significantly improve the information for managers?

Materials and Methods

There is a lack of information about parameters defined in the algorithms, information about the extension of study areas, time of processing of all steps, etc.

Why use two different approaches (pixel and object-based) instead of image segmentation processing, for example? Also, the number of training and validation data is missing, and if the dataset is representative.

The methodological design seems to be very difficult to reproduce for different areas and seems to be a more experimental approach. Also, the steps are not very clear, and I suggest presenting a flowchart.

Figure 3d. It's difficult to see the green polygons.

Table 3. Its cited only in discussion section.

What the impacts of establishing a manual threshold calibration? Is need any evaluation in different areas?

Table 5. Repeated information shown in Figure 6.

Lines 408-415. It's strange this information in the results section.

Discussion

What about the effects of herbaceous vegetation due to rapidly growing after a fire? Maybe it could also influence the detection of burn extension, such as observed in image 8b. For the processing approach used, I believe that the period of post-fire images acquisition must be analyzed. In some cases, 6 weeks (as cited in lines 476-478) could not be enough.

Conclusion

I suggest emphasizing in the conclusion that improving your approach is only related to your study area and UASs images.