LUFFD-YOLO: A Lightweight Model for UAV Remote Sensing Forest Fire Detection Based on Attention Mechanism and Multi-Level Feature Fusion

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The paper is well written, with a lot of interesting results. I have only some suggestions. In the indroduction and until the scenes aquired by UAV are presented, you do not state in which bands the images are acquired or with which type of sensor are acquired: it can only be deducted from the images itself, I think that is necessary to state at the beginning what sesor/bands was used. Moreover, fire detection is more easily made by means of images acquired in the Thermal infrared, and TIR sensor for UAV are available. Can you give some reason why you have choose to make use of (I suppose) RGB image?

Comments on the Quality of English Language

English is corrent in the whole. Maybe only some inccuracies with the use of comparative adjectives (please check in the text)

Author Response

Please see the attached file. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Study proposes architectural changes and enhancements of YOLOv8n for the purpose of real-time fire detection in UAV remote sensing data. The proposed GhostNetV2 backbone alteration reduces overall model complexity providing enhanced real-time capabilities. Enhanced small-object detection c2f structure provides further feature information at varying levels proving to increase small fire detection. Finally, a proposed hierarchical feature integrated c2f structure is introduced and provides an improvement in performance with complex background information. The proposed model shows promising results in both a reduction in parameter size and an increase in detection precision.

The following comments should be addressed to enhance clarity.

Section 2.1:

• The study includes use of an existing UAV remote sensing dataset, M4SFWD, and discusses the alterations needed to maintain consistency in the study. However, it is unclear if the images were scaled to 640x640 or if these were cropped out of the usable images in this dataset.
• Additionally, the dataset is stated to have 3974 images, is this the count used in the study post alterations or pre alterations? If the latter, what is the total number of images trained and tested with?
• Data augmentation was leveraged for the created SURSFF dataset, however it is unclear what values for each parameter were selected hindering reproducibility.

Line 366-371: Paragraph is out of place, disrupts the flow of the paper. Doesn’t seem to fit in the “The proposed LUFFD-YOLO network section”

Figure 8 layout is unclear, consider sub captions for each row in place of A-C numbering to ensure clarity in which model is represented

Table 2: Ablation experiment identifies the key changes made to improve model performance however, Methods(1) and Methods(2) is unclear. Consider updating this to the exact change in question I.E. GhostNetV2 + ESDC2f

Line 563-575: Paragraph is unclear, is the cited works utilizing the same public dataset in which the proposed LUFFD-YOLO is being compared? Or, are these works leveraging similar architectural enhancements for forest fire detection?

Model does improve with proposed architecture alterations, however, researchers note some shortcomings for future work. Discussion on these shortcomings and examples from your study would further highlight this and allow for broader discussion of performance of proposed model.

 

 

 

Comments on the Quality of English Language

NA

Author Response

Please see the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In Figure 7, the graphs are very small, which makes it impossible to perceive the features of the graphs in detail.

In Figure 9, the images (feature maps) are so small that it is not clear at all how they differ and why there are so many of them in the Figure.

The authors concluded that they proposed “a lightweight object detection model that achieves a high level of accuracy while maintaining a good balance between its architecture and detection performance”. However, this claim is supported only for a limited data set since the results were obtained on a prepared very small dataset, which does not make it possible to generalize the results obtained. It would be advisable to test the model on real remote-sensing data obtained during UAV pilot operation, which would give a more reliable result for comparison.

Author Response

Please see the attached file. 

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

the revised manuscript is in a way better shape now. Thanks for addressing most of my comments there. 

Minor comments:

- The caption in Figure 1 has become long. Please try shortening it and instead discuss it more in the body of the paper.

- The paragraph starting at line 588 still feels out of place to me. It seems to compare the proposed network to existing works without any further differentiation. Also seems datasets are not the same here, so how can we compare scores in a meaningful way? Suggest modifying it.

 

Other than that, good job.

Author Response

Thank you for the comments. Please see the response to the comments. 

Author Response File: Author Response.pdf