Submit to this Journal Review for this Journal Propose a Special Issue

Article Menu

Share Help Cite Discuss in SciProfiles

Open AccessFeature PaperArticle

Peer-Review Record

Forest Fire Segmentation via Temporal Transformer from Aerial Images

Forests 2023, 14(3), 563; https://doi.org/10.3390/f14030563

by Mohammad Shahid¹, Shang-Fu Chen¹, Yu-Ling Hsu², Yung-Yao Chen³

, Yi-Ling Chen¹ and Kai-Lung Hua^1,*

Reviewer 1:

Reviewer 2:

Reviewer 3:

Reviewer 4:

Reviewer 5:

Esteban Tlelo-Cuautle

Forests 2023, 14(3), 563; https://doi.org/10.3390/f14030563

Submission received: 11 December 2022 / Revised: 5 March 2023 / Accepted: 7 March 2023 / Published: 13 March 2023

(This article belongs to the Special Issue UAV Aided Forest Fire Risk Prediction Based on Remote Sensing, Machine Learning and Cloud Computing)

Round 1

Reviewer 1 Report

In this paper, the authors proposed an architecture for forest fire segmentation termed FFS-Unet, which comprises a transformer, UNet and skip connection into a unified model to collaboratively extract low-level and global dependencies between frames. This is a meaningful study, however, there are several concerns should be considered and solved.

1. The abstract does not reflect technological innovation, and does not analyze the essential reasons. Why comprises a transformer, UNet and skip connection into a unified model can be used to solve the corresponding problems?

2. Authors have written/mentioned more about the generic information and they have focused more on their contribution in the paper.

3. What about the infrared image?

Author Response

Dear Associate Editor and Reviewers,

Thank you very much for your useful comments and suggestions to the paper. The manuscript has been through a number of modifications, and benefited a lot from your reviews. Each modification we performed in response to the associate editor and the reviewers is highlighted in yellow in the revised manuscript. Here we provide a detailed response to the reviewers’ comments and questions. In general, we have improved the writing of our manuscript and added more descriptions to the figures and formulas to make it clearer.

Enclosed is the revision of our manuscript and the responses to reviewers' comments. This manuscript has been reviewed and approved by all authors. We appreciate your assistance on this matter and await your decision and those of the reviewers on this manuscript's suitability for publication in Forest, MDPI. We feel confident that the work will make impact to the related research community once it is accepted.

Best regards,
Kai-Lung Hua
Professor
Department of Computer Science and Information Engineering
National Taiwan University of Science and Technology
E-mail: [email protected]
Phone: +886-2-2730-1066
Fax: +886-2-2730-1081

Author Response File: Author Response.pdf

Reviewer 2 Report

The description is clear. May be on raw 149 - "built a low-cost system using deep convolutional neural" has to be "built a low-cost system using deep convolutional neural network"

Author Response

Dear Associate Editor and Reviewers,

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

the Article is relevant and scientific-sounding, but there are some notes that, in my opinion, should be revised before promotion:

1. The Results section should contain methodological findings made by Authors.

2. The Discussion section is missed - it should show the advances mabe by this research over existing papres published earlier, as well as future perspectives of research in the field.

3. The Conclusions section in my opinion should not contain figures; they look better in Results section.

4. I suggest adding in the References list some international works, i.g. from Canada, USA, etc.

Good luck!

Author Response

Dear Associate Editor and Reviewers,

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper proposes an architecture for forest fire segmentation, termed FFS-Unet, able to detect fire from aerial images. The paper is interesting and generally well written and organized. However, some issues need to be clarify:

1. There are some typos in the manuscript. Some examples: a) line 280: instead of “segmentation” it is written “Segmentation”; b) lines 282-283: the sentence “Therefore …” is badly constructed; c) line 296: the sequences “UAV fire video dataset” and “No sizeable …” need to be separated using colon (i.e. “:”), d) lines 298 and 301: Flame is written either “Flame” or “FLAME”; the same happens with “IoU” which is sometimes written “iou” (see lines 292 or 293 for example) or “IOU” (line 442); e) line 389: the sentence is confusingly written; etc. Please check the entire manuscript.

2. Line 285: on what basis the two values (0.4 and 0.7) were chosen. Some details are needed.

3. Line 298: what “Internet” means in the context? Please explain or rewrite.

4. Line 377: it is written “the weighted sum of LCE + LDice” which is confusing. Also, the notation “LCE + LDice” (Table 5) is confusing. Please change;

5. Line 320: what the authors mean by “complete convergence”?

6. Line 393: please change the word “incredibly”;

7. Subsection 6.1: Order of presentation in the list (text body) is not the same as in Table 6;

8. Figures 12 and 13: Please increase the contrast and raise the font dimension. Moreover, what “IoU vs. #Param (M)” or “IoU vs. Infer-time(ms)” represent (see Figure 13)?

Author Response

Dear Associate Editor and Reviewers,

Author Response File: Author Response.pdf

Reviewer 5 Report

The article titled: Forest Fire Segmentation via Temporal Transformer from Aerial

Images, includes 59 references, a good number to support this research. Authors may search for some references published during 2023.

The manuscript looks good including 13 figures and seven tables. In the experiments, authors list some realted Works to compare the results as shown in two tables 6 and 7:

Table 6. Fire detection results compared with different method on UaV fire dataset

Table 7. Fire detection results compared with different method on the CorsicanFire dataset [59]

In both tables the six references are the same and show similar results when performing the comparison. However, authors may find more recent Works, as of 2023 to compare with the most recent state of the art.

The contribution is well justified from the Abstract: Forest fires are one of the most critical natural tragedies threatening forest land and 1 resources. Correct and early forest fire detection is essential to reduce losses and improve firefighting. 2 Conventional firefighting techniques, based on ground inspection and limited by the field of view, 3 lead to insufficient monitoring capability for a large area. Recently, due to their great flexibility and 4 ability to shield large regions, Unmanned Aerial Vehicles (UAVs) have been used to combat forest 5 fire incidents. An essential step for an autonomous system to deal with fire situations is first to locate 6 the fire in a video… Authors can find more realted Works on UAVs that include video analysis.

The result emphasized in the abstract need more justification to be considered as the last state of the art: The empirical outcomes were evaluated on the UAV-collected 14

video and Carson fire datasets. The proposed FFS-Unet enhances the performance by achieving 95.1% 15 in F1 and 86.8% in IoU, higher than previous forest fire techniques with fewer parameters…. Other related Works may perform much better using other video analysis and even other fire datasets, can it be posible to occur?

In line 43 authors mention only two cases of fire detection: Airborne forest detection 43 can be further sub-grouped into a) Human-powered helicopters and b) unmanned aerial 44 vehicles (UAV)… Are there other method that may compete with UAVs?

You should justify why mentioning tratidional UAVs methods from the following paragraph: The traditional UAV-based forest fire detection technique utilises distinguishable 56 characteristics such as colour, form, texture and motion features. For instance, Esfahlani 57 et at. [ 13 ] presented a fire detection algorithm based on color, movement attributes, and 58 temporal variation of fire intensity.

Authors should add some text between each section and subsection, as in 2. Related Work 105 2.1. Shallow models for fire detection… That way, the reader can find the abstract of the subsections. See for example Section 3. Method 186
In this section, based on the U-Net [45 ] classic segmentation network, we propose

In some subsections you need to add references from which equations are derived, as in Version December 11, 2022 submitted to Journal Not Specified 5 of … 2.4. Convolution operation, which does not include any reference.

You missed discussing what the reader can find in section 5 and section 6, in the last paragraph of section 1: The paper is organized as follows. Section 2 discusses some related works and 101
highlights some representative work closely related to our research. We explain our 102
approach in Section 3 and discuss evaluations in Section 4 before finally arriving at a 103
conclusion in Section 7.

Article Menu

Forest Fire Segmentation via Temporal Transformer from Aerial Images

Further Information

Guidelines

MDPI Initiatives

Follow MDPI