Real-Time Detection of Full-Scale Forest Fire Smoke Based on Deep Convolution Neural Network
Round 1
Reviewer 1 Report
Dear authors,
Thank you very much for your manuscript. I think the initiative to utilize advance method for rapid forest fires detection is significant in the practice of sustainable forest management.
There are several items you have to address. Kindly find my comments below.
- Is there any speed of the forest fire detection that you wish to achieve or supposed to be achieve? Or is there any standard rescue operation speed that the information need to be produced.
In my opinion, if there is no standard or desired speed, it is difficult to discuss your result, especially on the implication of the findings. - About the finding in figure 10, can you explain why such differences occur? Support your argument with theoretical prior to the exact method.
Still on figure 10, I noticed that the difference in detection speed is somewhere about 10-15 fps. Is this small differences really impacted the entire process of rescuing mission or quick response mitigation initiatives? - Can you discuss the trade off each method that being tested? For instance, increase of time processing, additional cost incur due to additional requirement of storage, etc.
Highlighting this issue is so critical in real emergency response application; as the fastest method sometimes require complex setup and sometimes method with the simplest computation and setup is desired due to requirement of operationality. - Could you explain the limitation of each techniques? For instance, what is the minimum scale of forest fire can be detected and so on.
I hope that my comment is useful to improve the quality of this manuscript.
Good luck
Author Response
Dear Editors:
We are very grateful to Journal of Forests for supporting our research work. According to the suggestions mentioned in previous E-mail, the changes are made to make the paper conform to the publishing standard, and the revised paper has been submitted. As for the comments, we will give you the direct answers.
1. Is there any speed of the forest fire detection that you wish to achieve or supposed to be achieve? Or is there any standard rescue operation speed that the information need to be produced.
In my opinion, if there is no standard or desired speed, it is difficult to discuss your result, especially on the implication of the findings.
Ans:
In addition to detection accuracy, speed of FPS detection target detection processing the number of pictures is another detection indicator of the target detection algorithm. Real-time detection could be realized as long as the speed of algorithm detection is fast enough. study showed that FPS reach more than 12 frames per second which meets the requirements of real-time detection and the higher the FPS frame rate, the stronger the real-time detection. Compared to FPS, the experiment needs to be performed on the same hardware. The operating hardware platform of this experiment is Lines 267-270。
2. About the finding in figure 10, can you explain why such differences occur? Support your argument with theoretical prior to the exact method.
Still on figure 10, I noticed that the difference in detection speed is somewhere about 10-15 fps. Is this small differences really impacted the entire process of rescuing mission or quick response mitigation initiatives?
Ans:
Faster RCNN has higher detection accuracy while YOLO series is faster. Faster RCNN has a two-stage scheme to detect target. Firstly, finding out the feature using the best network and then adjusting the frame. However, those two-stage work could be completed in only one stage applying YOLO series method. The core of Faster RCNN is to find the network with best performance at present, and then assemble networks together to produce better results. Based on a multi-network fusion scheme, feature of Faster RCNN are very precise but with slow calculation which is detrimental to the real-time nature of our forest fire smoke detection.
The emergence of YOLO v3 exactly solved this problem. The most significant feature of YOLO v3 is fast and accurate as well as Faster RCNN. The forest fire smoke detection results in Figure 10 showed that, the maximum detection speed of YOLO v3 reaches 27 frames/second while The detection real-time performance of Faster RCNN is the worst among the four kinds of target detection, with detection speed of 5 frames/second. Compared to Faster-RCNN detection model requiring object proposals, the SSD method completely eliminates the stages of proposals generation, pixel resampling or feature resampling, making it easier to optimize training and to integrate the detection model into the system. Though the speed of 16 frames per second meets the real-time detection requirements, but its detection accuracy is 87.5%, which is the lowest compared with other models.
Based on the results of the Scalable Neural Network (EfficientNet), Efficient-Det combined with a new bidirectional feature network (BiFPN) and new scaling rules to achieve SOTA accuracy. Compared to the previous most cutting-edge detection algorithm, Efficient-Det’s volume is reduced to 1/9 of the original and the amount of calculation is also greatly reduced. This study developed a small EfficientDet-D0 baseline from D0 to D7 model to improve the detection accuracy gradually while the computing power is also increased. According to the experimental results show that the detection speed is 12 frames/second, and the detection accuracy is up to 95.7%.
3. Can you discuss the trade off each method that being tested? For instance, increase of time processing, additional cost incur due to additional requirement of storage, etc.
Highlighting this issue is so critical in real emergency response application; as the fastest method sometimes require complex setup and sometimes method with the simplest computation and setup is desired due to requirement of operationality .
Ans:
The trade-offs of the four target detection algorithms have been added in lines 313-340 of the article.
4. Could you explain the limitation of each techniques? For instance, what is the minimum scale of forest fire can be detected and so on.
Ans:
In the target detection algorithm , Small target object was defined as whose width and height are less than one-tenth of the original image. Most of the foreground smoke in the forest fire smoke dataset belongs to small targets while the experiment showed satisfying detection results among both large and small targets which reflects the multi-scale detection of the algorithm in this study.
Best regards
The authors
11-26-2021
Reviewer 2 Report
Consider following comments
- Provide source of the images used in the figure 1 (Dataset total 999).
- Figure 2 source?
- Reference style need to check.
- It could be better if author provide dates of all image, so reader can check with real fire dates.
- Conclusion need to say why your algorithm is better than others with reference.
- Discussion part should needed.
Author Response
Dear Editors:
We are very grateful to Journal of Forests for supporting our research work. According to the suggestions mentioned in previous E-mail, the changes are made to make the paper conform to the publishing standard, and the revised paper has been submitted. As for the comments, we will give you the direct answers.
Point 1: Provide source of the images used in the figure 1 (Dataset total 999).

Response 1: Because the dataset is relatively large, it has been open sourced on ZHENG dataset 2021 for readers to check, the position in the article lines 240-241.
Lines 240-241, “The dataset that we use in the experiments can be freely download via ZHENG dataset 2021.” was added.
Point 2: Figure 2 source?
Response 2: The dataset address has been open sourced on ZHENG dataset 2021.
Point 3: Reference style need to check.
Response 3: Reference format has been modified.
Point 4: It could be better if author provide dates of all image, so reader can check with real fire dates.
Response 4: The dataset address has been open sourced on ZHENG dataset 2021.
Point 5: Conclusion need to say why your algorithm is better than others with reference.
Response 5: Discussed on lines 313-340 of the article.
Lines 313-340, “Discussion
Faster RCNN has higher detection accuracy while YOLO series is faster. Faster RCNN uses a two-stage scheme to detect the target. The feature will be found out using the best network followed by adjusting the frame. However, the two-stage scheme can only be completed in one stage when the YOLO series method is applied. The core of the Faster RCNN is to find the network with best performance, and then assemble networks together to produce better results. Based on a multi-network fusion scheme, feature of Faster RCNN are very precise but with it yields slow computation which is detrimental to the real-time nature of the forest fire smoke detection.
The emergence of YOLO v3 solved this challenge. The most significant feature of the YOLO v3 is fast and accurate as well as the Faster RCNN. The forest fire smoke detection results in Figure 10 showed that, the maximum detection speed of YOLO v3 reached 27 frames/second while the real-time detection performance of the Faster RCNN is the worst among the four methods of target detection, with detection speed of 5 frames/second. Compared to the Faster-RCNN detection model requiring object proposals, the SSD method completely eliminates the stages of proposals generation, pixel resampling or feature resampling, making it easier to optimize training and to integrate the detection model into the system. Although the detection speed for the SSD method of 16 frames per second meets the requirements for real-time detection, its detection accuracy is 87.5%, which is the lowest compared with other models.
Based on the results of the Scalable Neural Network (EfficientNet), Efficient-Det can be combined with a new bidirectional feature network (BiFPN) and new scaling rules to achieve SOTA accuracy. Compared to the previous most cutting-edge detection algorithm, Efficient-Det’s volume is reduced to 1/9 of the original and the computation time is also greatly reduced. This study developed a small EfficientDet-D0 baseline from D0 to D7 model to improve the detection accuracy gradually while the computation effort is also decreased. According to the experimental results, it shows that the detection speed is 12 frames/second, and the detection accuracy is up to 95.7%.” was added.
Point 6: Discussion part should needed.
Response 6: The discussion section has been added to lines 313-340 of the article.
Best regards
The authors
11-26-2021
