An Efficient and Lightweight Detection Model for Forest Smoke Recognition
Round 1
Reviewer 1 Report
Comments and Suggestions for Authors1. The relevant studies cited in the Introductionn section are on flames, and there is a lack of discussion of relevant studies on smoke detection.
2. Carefully review the classification of deep learning-based detection methods. For example, YOLOv3 belongs to one-stage instead of two-stage.
3. In the 2.2, the FPN and PAN parts should be worded carefully. For example, who is the comparator of their lower recognition accuracy.
4. In the 2.3, for the explanation of public announcement, the relevant words should be symbols instead of letters.
5. The explanation of Equations 4, 5 is not detailed enough.
6. Some of the symbols in Figure 3(b) are not explained in detail.
7. In the 2.4, for the explanation of the public announcement, the relevant words should be symbols instead of letters.
8. why are the pixels of the training images in Table 2 inconsistent with the pixels described in the dataset?
9. The evaluation metric in Table 3 should be mAP instead of Map.
10.Line 267 should be figure 7(b). In addition, it should be explained why the detection overlap is produced in figure (a).
11. Add comparative experiments as appropriate.
Comments on the Quality of English Language
Quality of English Language need to be improved.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.
Comments 1: The relevant studies cited in the Introductionn section are on flames, and there is a lack of discussion of relevant studies on smoke detection.
Response 1: Thank you for pointing this out. We agree with this comment. Therefore, I have changed the discussion of relevant studies on smoke detection. This change can be found on page 2, second paragraph, line 56.
“Toreyin et al. [4]used the texture features of smoke for detection, calculated the wavelet energy of the suspected motion region by wavelet transform, and judged whether it was smoke or not based on the trend of wavelet energy. Lecun [5]used structural wavelet transform to represent the smoke texture image, calculated the different scales of wavelet transform by using GLCM, and used neural networks to classify the smoke candidate regions. Chen [6]proposed a block-based interframe differencing and LBP-TOP combination method for smoke dynamic characterization. To reduce false alarms, a smoke histogram was constructed to record most recent classification of smoke candidate regions. Traditional smoke detection methods have made some contributions. However, the method of extracting specific smoke features from fire scenarios is not well suited for diverse detection scenarios.”
Comments 2: Carefully review the classification of deep learning-based detection methods. For example, YOLOv3 belongs to one-stage instead of two-stage.
Response 2: Agree. We have, accordingly, revised the classification of deep learning-based detection methods to emphasize this point. Mention exactly where in the revised manuscript this change can be found – page 2, third paragraph, and line 70.
“Currently, end-to-end object detection methods based on deep learning can be divided into two types. Two-stage object detection based on candidate regions, such as Faster R-CNN [7], R-FCN [8], and Libra R-CNN [9]. One-stage regression-based object detection, such as YOLO [10-12], SSD [13], EfficientDet [14], and RetinaNet [15]. Specifically speaking, although two-stage algorithms have high detection accuracy, they are not suitable for real-time detection tasks. Single-stage object detection typically requires only a single forward computation, but two-stage methods may still have some advantages for object tasks that require high accuracy.”
Comments 3: In the 2.2, the FPN and PAN parts should be worded carefully. For example, who is the comparator of their lower recognition accuracy.
Response 3: Thank you for pointing this out. We agree with this comment. Therefore, we have changed the wording of the FPN and PAN parts. This change can be found on page 5, second paragraph, line 154.
“FPN [23] is a top-down approach. It upsamples the coarse location information but semantically stronger feature maps at the higher pyramid level to pro-duce higher resolution features. PAN [22] complements FPN by down-sampling from the bottom up, so that the top-level features contain image location information. Because FPN uses summation for feature fusion, some detail is lost in the fusion process. Therefore, it may not perform as well as PAN for scenes that re-quire high-precision detection. PAN uses a cascading approach to feature fusion, which can preserve more detail but increases computational complexity.”
Comments 4: In the 2.3, for the explanation of public announcement, the relevant words should be symbols instead of letters.
Response 4: We gratefully appreciate for your valuable suggestion. We agree with this comment. Therefore, we have revised the discussion of relevant studies on smoke detection. This change can be found on page 2, second paragraph, line 56.
“Eq.1 is self-attention, where aggregation M1 on context X is performed after calculating the attention score between query and target through interaction process T1. In contrast, Eq.2 is focal-modulation, where the context features are first aggregated using M2 at each location i, then the query interacts with the aggregated feature based on T2 to form yi. Where q denotes the query mapping function. ⨀ indicates the element-wise mul-tiplication. m is the context aggregation operation, which consists of two steps: hierarchical semantics in Eq.4 and gated aggregation in Eq.5.”
Comments 5: The explanation of Equations 4, 5 is not detailed enough.
Response 5: Thank you for pointing this out. We agree with this comment. Therefore, we have modified the explanation of Equations 4, 5. This change can be found in the 2.3.
“In Eq.4, Where is the context function of the l-th layer, generated by deep-wise convolution with kernel size kl and GeLU activation function. Hierarchical semantics extracts context information from local to global through different levels of granularity. In Eq.5, where G∈RH×W×1 is a slice of G for the level l. Specifically, we use a linear layer to obtain a spatial- and level-aware gating weights G = fg(X) ∈ RH×W×(L+1). Then, we perform a weighted sum through an element-wise multiplication to obtain a single feature map Zout which has the same size as the input X. Gated aggregation condenses context features at different levels of granularity into a single feature vector, the modulator. Combining the previous interaction and aggregation, the focal modulation formula can be expressed as Eq.6, where and are the gating values and visual features at position i of Gl and Zl, respectively.”
Comments 6: Some of the symbols in Figure 3(b) are not explained in detail.
Response 6: We gratefully appreciate for your valuable comment. We agree with this comment. Therefore, we have changed the symbols in Figure 3(b). This change can be found in the 2.3.
“Detailed explanation of context aggregation (b) in Focal Modulation (a). The aggregation procedure consists of two steps: hierarchical contextualization to extract contexts from local to global ranges at different levels of granularity and gated aggregation to condense all context features at different granularity levels into the modulator.”
Comments 7: In the 2.4, for the explanation of the public announcement, the relevant words should be symbols instead of letters.
Response 7: Thank you for pointing this out. We agree with this comment. Therefore, we have modified the explanation of the public announcement in the 2.4. This change can be found in the 2.4.
“Where = wtt + bt and = wtxi + bt are linear transforms of t and xi,t and xi denote the target neuron and other neurons in a single channel of the input feature, respectively. i is index over spatial dimension and M = H×W is the number of neurons on that channel. wt and bt are weight and bias the transform, computed with the following. and denote the mean and variance, respectively, computed for all neurons in the channel except t.”
Comments 8: Why are the pixels of the training images in Table 2 inconsistent with the pixels described in the dataset?
Response 8: Agree. We feel sorry for the inconvenience brought to the reviewer. We have, accordingly, revised the pixels of the training images in Table 2 and deleted the pixels described in the paragraph.
Comments 9: The evaluation metric in Table 3 should be mAP instead of Map.
Response 9: Thank you for pointing this out. We agree with this comment. Therefore, the revised evaluation metric mAP in Table 3.
Comments 10: Line 267 should be figure 7(b). In addition, it should be explained why the detection overlap is produced in figure (a).
Response 10: Thank you for pointing this out. We agree with this comment. Therefore, we have changed the explanation why the detection overlap is produced in figure (a). This change can be found in figure 7.
“Figure 7 compares the effect of the missed detection problem when monitoring ground smoke. The improved YOLOv8s can detect forest smoke more accurately. However, the detection results of the unimproved YOLOv8s are less accurately, missing one smoke target (Figures 7b, d). It may be that the improved YOLOv8s have more robustness and better target detection performance.
Comments 11: Add comparative experiments as appropriate.
Response 11: Thank you very much for your comment. We totally agree with your opinion.
In this study, we focus on the performance of the improved YOLOv8s in practical forest smoke detection scenarios. The study mainly compares the performance of different attention mechanisms, different object detection models, and different modules in the improvement process, which are listed in Table 3, Table 4, and Table 5, respectively.
We would like to thank the reviewers again for taking the time to review our manuscript!
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript addresses an interesting issue. The article has enough innovation. I have some suggestions as follows to improve the manuscript:
1. More numerical results should be mentioned in the abstract.
2. The importance of research and research should be stated in the introduction. Also innovation and research objective
3. ROC index can be used to evaluate accuracy.
4. In the discussion section, comparison with past research and future suggestions will be discussed.
Minor editing of English language required
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.
Comments 1: More numerical results should be mentioned in the abstract
Response 1: Thank you very much for your suggestion. We totally agree with your opinion. Therefore, we have added the numerical results in the abstract, as follows:
“The experimental results show that the mean Average Precision of the improved model is 90.1%, which is 3% higher than the original model. The number of parameters and computational complexity of the model are 7.79MB and 25.6GFLOPs (giga floating-point operations per second), respectively, which are 30.07% and 10.49% less than the unimproved YOLOv8s.”
Comments 2: The importance of research and research should be stated in the introduction. Also innovation and research objective.
Response 2: Agree. We gratefully appreciate for your valuable comment. The importance of research and research should be stated as follows:
“If a fire occurs on a large scale, it will cause incalculable casualties and economic losses. Therefore, there is an urgent need for the development of a fast and effective fire detection solution. We propose an efficient and lightweight forest smoke detection model to successfully detect and localize smoke in the early stages of forest fire spread. Edge device with limited computing resources and memory can achieve real-time object detection. New ideas and insights are provided for forest fire detection.”
Comments 3: ROC index can be used to evaluate accuracy.
Response 3: Thank you very much for your comment. We totally agree with your opinion. ROC curve provides a visual tool for evaluating the performance of a model at different thresholds and is typically used for binary classification problems. Object detection is a multi-objective problem, and mAP provides a comprehensive evaluation of detection accuracy, localization accuracy, and multi-objective detection performance. Although it is possible to transform the object detection task into a binary classification problem and evaluate it using ROC curves, some important information about the object detection task is lost. Therefore, mAP, a more specific and comprehensive metric for evaluating the object detection task, is preferred for the improved YOLOv8s.
Comments 4: In the discussion section, comparison with past research and future suggestions will be discussed.
Response 4: Thank you very much for your comment. We totally agree with your opinion. In the discussion section, the past research and future suggestions as follows:
“The actual forest fire scenario has extremely complex features, such as trees, terrain, and other disturbing factors, which increases the difficulty of object detection [33]. More past research is presented in the introduction section.
In the future, we need to further optimize the forest smoke detection algorithm to better adapt to the practical application scenarios of edge computing platforms. Firstly, we will focus on the model compression algorithms for neural network models on edge devices. This includes techniques to reduce the memory and computational complexity of the model, such as Model Pruning [34], Knowledge Distillation [35], Quantization [36], and Low-rank Decomposition [37]. In addition, the great success of deep learning relies heavily on increasingly large training data sets. Dataset compression can be used to construct a minimal subset of training data from the entire training data set without significantly affecting the performance of the model [38]. Finally, we will explore automated neural network architecture design for neural network architecture search. This approach allows for adaptive generation of optimal network structures based on specific scenario requirements, thus improving the applicability and efficiency of the algorithms. These research directions aim not only to reduce computational cost and memory requirements, but also to improve the efficiency and utility of object detection. By implementing efficient wildfire detection algorithms on edge devices, we can respond to fire threats in more timely manner and minimize losses.”
We would like to thank the reviewers again for taking the time to review our manuscript!
Reviewer 3 Report
Comments and Suggestions for AuthorsThe paper presents a comprehensive study on improving forest smoke detection using a modified YOLOv8 architecture. The paper is quite good, but it could be improved. Below are some comments and suggestions:
- The abstract and introduction effectively highlight the significance of early forest fire detection and the challenges in current methods. Emphasizing the novelty of proposed approach in the context of existing literature strengthens the introduction.
- The review of existing methods provides a good foundation. However, it would be beneficial to discuss more recent advancements in the field including research done in others countries, particularly European, because most references are from China.
- The paper gave adetailed explanation of the YOLOv8-based model, the introduction of BiFPN, SimAM, and Focal Modulation provides clarity. It would be helpful to further elucidate why these specific enhancements were chosen and how they specifically address the challenges of forest smoke detection.
- The creation of a forest smoke dataset is commendable. The dataset used in experiments MUST BE publicly available on certain open Dataset platforms if the paper will be published. In its final version it must include link to Forest Smoke Dataset of 8176 images, so that anybody could repeat and check authors research contributions. This is ‘Conditio sine qua not’ for papers like this. Additional details about the dataset, like diversity in weather conditions and smoke densities, would be also valuable. Also, explaining the rationale behind the chosen division of training, validation, and test sets enhances the methodological rigor.
- The comparison with other models like Faster R-CNN, SSD, and EfficientDet is insightful. Further discussion on why your model outperforms these could provide deeper insights into the effectiveness of your modifications.
- The ablation study is thorough. It would be useful to include a discussion on the trade-offs involved in these modifications, particularly regarding computational efficiency and real-time applicability.
- Discussing how this model can be implemented in real-world scenarios, its compatibility with various edge devices, and its performance in different environmental conditions would be useful. Also, acknowledging any limitations and potential areas for future improvement would provide a balanced view.
- The conclusion effectively summarizes the paper's contributions. Expanding on potential future work, such as adaptation to different types of smoke and fires, or integration with other fire detection technologies, would be beneficial.
Overall, your paper addresses a critical and timely issue with a novel approach. Its strengths lie in the technical innovation and comprehensive experimental evaluation. Enhancing it with the suggestions above could make it an even more valuable contribution to the field of forest fire detection and environmental monitoring.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.
Comments 1: The abstract and introduction effectively highlight the significance of early forest fire detection and the challenges in current methods. Emphasizing the novelty of proposed approach in the context of existing literature strengthens the introduction.
Response 1: Thank you very much for your suggestion. We totally agree with your opinion. We propose an efficient and lightweight forest smoke detection model to successfully detect and localize smoke in the early stages of forest fire spread. Edge device with limited computing resources and memory can achieve real-time object detection. New ideas and insights are provided for forest fire detection.
Comments 2: The review of existing methods provides a good foundation. However, it would be beneficial to discuss more recent advancements in the field including research done in others countries, particularly European, because most references are from China.
Response 2: We gratefully appreciate for your valuable suggestion. Therefore, I have revised the discussion of relevant studies in others countries. This change can be found on page 2, second paragraph, line 56.
“Toreyin et al. [4]used the texture features of smoke for detection, calculated the wavelet energy of the suspected motion region by wavelet transform, and judged whether it was smoke or not based on the trend of wavelet energy. Lecun et al. [5]used structural wavelet transform to represent the smoke texture image, calculated the different scales of wavelet transform by using GLCM, and used neural networks to classify the smoke candidate regions. Chen et al. [6]proposed a block-based interframe differencing and LBP-TOP combination method for smoke dynamic characterization. To reduce false alarms, a smoke histogram was constructed to record most recent classification of smoke candidate regions. Traditional smoke detection methods have made some contributions. However, the method of extracting specific smoke features from fire scenarios is not well suited for diverse detection scenarios.”
Comments 3: The paper gave adetailed explanation of the YOLOv8-based model, the introduction of BiFPN, SimAM, and Focal Modulation provides clarity. It would be helpful to further elucidate why these specific enhancements were chosen and how they specifically address the challenges of forest smoke detection.
Response 3: Thank you for pointing this out. We agree with this comment. Why these particular enhancements were chosen, and how they specifically address the challenges of forest smoke detection, as follows:
“To improve the feature fusion capability in forest smoke detection, we fuse a simple yet efficient weighted feature fusion network into the neck of YOLOv8. It also greatly optimizes the number of parameters and computational load of the model. Then, the Simple and parametric-free Attention Mechanism (SimAM) is introduced to address the problem of forest smoke dataset images that may contain complex background and environmental disturbances. The detection accuracy of the model is improved and no additional parameters are introduced. Finally, we introduce focal modulation to increase the attention to the hard-to-detect smoke and improve the running speed of the model.”
Comments 4: The creation of a forest smoke dataset is commendable. The dataset used in experiments MUST BE publicly available on certain open Dataset platforms if the paper will be published. In its final version it must include link to Forest Smoke Dataset of 8176 images, so that anybody could repeat and check authors research contributions. This is ‘Conditio sine qua not’ for papers like this. Additional details about the dataset, like diversity in weather conditions and smoke densities, would be also valuable. Also, explaining the rationale behind the chosen division of training, validation, and test sets enhances the methodological rigor.
Response 4: Thank you very much for your suggestion. We totally agree with your opinion. As the dataset and code are confidential and of interest to the company, please forgive us if we cannot make them public. In addition, the weather conditions and smoke densities have little, though possibly valuable, connection to the main goal of this study - efficient and lightweight object detection model. In general, the data set is split 8:1:1, and a larger training set helps the model learn general features of the data and improves the generalization ability of the model.
Comments 5: The comparison with other models like Faster R-CNN, SSD, and EfficientDet is insightful. Further discussion on why your model outperforms these could provide deeper insights into the effectiveness of your modifications.
Response 5: Thank you for pointing this out. We agree with this comment. Therefore, we have changed the discussion on why my model outperforms other models as follows:
“Our improved YOLOv8s model achieves significant results, with a model mean Average Precision as high as 90.1% and an F1 score of 0.89, which is significantly higher than other object detection models. And the improved YOLOv8s has stronger generalization performance and robustness.”
Comments 6: The ablation study is thorough. It would be useful to include a discussion on the trade-offs involved in these modifications, particularly regarding computational efficiency and real-time applicability.
Response 6: Thank you very much for your comment. We totally agree with your opinion. The improved YOLOv8 model has a 3% improvement in mean Average Precision, the 30.07% reduction in the number of parameters, and the 10.49% reduction in computational complexity, and the FPS is 100. It is beneficial to computational efficiency and real-time applicability.
Comments 7: Discussing how this model can be implemented in real-world scenarios, its compatibility with various edge devices, and its performance in different environmental conditions would be useful. Also, acknowledging any limitations and potential areas for future improvement would provide a balanced view.
Response 7: Thank you very much for your comment. We totally agree with your opinion. The actual forest fire scenario has extremely complex features, such as trees, terrain, and other disturbing factors, which increases the difficulty of object detection. In addition, the detection sensitivity of the algorithm will be different for large forest fires and small forest fires. The next work should continue to expand the data set to further improve the performance of the model. A balanced ratio of positive and negative samples helps the model learn the object and background more comprehensively, improving model robustness and generalization performance.
In the future, we will focus on the model compression algorithms for neural network models on edge devices. This includes techniques to reduce the memory and computational complexity of the model, such as Model Pruning, Knowledge Distillation, Quantization, and Low-rank Decomposition. In addition, dataset compression can be used to construct a minimal subset of training data from the entire training data set without significantly affecting the performance of the model. Finally, we will explore automated neural network architecture design for neural network architecture search. This approach allows for adaptive generation of optimal network structures based on specific scenario requirements, thus improving the applicability and efficiency of the algorithms. These research directions aim to reduce computational cost and memory requirements and to improve the efficiency and utility of object detection. By implementing efficient smoke detection algorithms on edge devices, we can respond to fire threats in more timely manner and minimize losses.
Comments 8: The conclusion effectively summarizes the paper's contributions. Expanding on potential future work, such as adaptation to different types of smoke and fires, or integration with other fire detection technologies, would be beneficial.
Response 8: Agree. We gratefully appreciate for your valuable comment. In this study, compared to the original YOLOv8s, the mean Average Precision of forest smoke detection is improved by 3%, and the number of parameters and computational complexity of the model are reduced by 30.07% and 10.49%, respectively. The proposed detection model significantly outperforms other existing object detection networks in the self-built forest smoke detection dataset, and also has advantages in lightweight models. The improved model can be applied to edge devices, such as mobile devices and UAVs, to realize real-time monitoring and early warning, and improve the response speed and accuracy of fire events, which is of great significance for early detection and response to forest fires.
We would like to thank the reviewers again for taking the time to review our manuscript!
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe citation [37] is not shown in the main text.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.
Comments 1: The citation [37] is not shown in the main text.
Response 1: My sincere apologies for putting quote [37] at the beginning, probably due to starting a new page. The citation [37] can be found on page 15, second paragraph, line 418.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors accepted only part of my comments in the modified version of the paper, and for most of them they only gave explanations. Some of the explanations are acceptable, some are not.
- I especially disagree with the explanation on point 4 related to the public publication of the Forest Smoke Dataset, which was used for training and testing of the new smoke detection procedure. The author's answer was: "We totally agree with your opinion. As the dataset and code are confidential and of interest to the company, please forgive us if we cannot make them public." which is not acceptable for me. The authors want to publish their paper as scientific paper. The scientific paper should be such that everyone, who is interested, can completely repeat their experiment and get the same results that the authors got and then maybe even improve them with their own modifications. If the database of images is not published publicly, this is not possible, and the work cannot be scientific paper. We can only believe that the obtained results are correct, and one thing is belief, and the other is the possibility of verifying them. The work can be accepted as a scientific paper only if:
a) The database of images for training and testing (8176 images in Forst Smoke Dataset) will be made public or
b) The authors repeat the entire procedure, but this time using an existing publicly available database of wildfire smoke images.
The code does not have to be published, but all the parameters of the network should be published so that everyone can repeat the experiment and work on possible improvements.
This is the essence of scientific work. Scientific work is not only about advertising your results, but also about contributing to the scientific community in such a way that other scientists can not only objectively check the results, but also continue to improve the procedure, in this case wildfire smoke detection.
- Another thing that should be fixed are the changes made according to Comments 2: The review of existing methods provides ... The existing changes are not appropriate. Only three works related to existing detection methods were mentioned. There are many more of them, so they should definitely be analysed in more detail. It looks like the analysis of existing methods was done just for the sake of order, and not with an effort to give a deep overview of what scientists have done so far. I especially disagree with the statement: "Traditional smoke detection methods have made some contributions. However, the method of extracting specific smoke features from fire scenarios is not well suited for diverse detection scenarios." In the period of the last 20 years or so, various standard and non-standard methods of detecting the smoke of emerging wildfires have had significant results and successes. Of course they could be improved. This is the point of science. Start with existing state of art and give some improvements, but on the way that all scientific community could benefit of their work.
Author Response
Thank you very much for taking the time to review this manuscript. Please find the detailed responses below and the corresponding revisions in track changes in the re-submitted files.
Comments 1: About dataset
Response 1: On behalf of our research team, we sincerely apologize. About sharing the forest smoke dataset. In this regard, I need to clarify to you that our research relies on sensitive information within the company and there is a very high demand for confidentiality of the data. In this case, we will ensure that company policies and regulations are followed to protect the security of confidential information. While we are unable to make the entire dataset publicly available, a detailed description of the dataset will be provided to ensure reproducibility and transparency of the research. We hope you understand the constraints we face.
In addition, "The authors repeat the entire procedure, but this time using an existing publicly available database of wildfire smoke images." I have added experiments to the original article, which is a dataset from a paper titled "An automatic fire detection system based on deep convolutional neural networks for low-power, resource-constrained devices" at link https://github.com/gaiasd/DFireDataset. I have repeated the experiments on this public dataset with the EfficientDet , YOLOv3-tiny, YOLOv4, YOLOv5, YOLOv8s, and improved YOLOv8s models, and the related conclusions and the table can be seen in the revised article.
Comments 1: The review of existing methods provides ... The existing changes are not appropriate
Response 1: Thank you for pointing this out. We agree with this comment. Changes as follows：
“In conclusion, while the above studies have introduced various methods to detect texture features of smoke, traditional smoke detection involves complex image processing methods with poor generalization performance and robustness.”
