Deep Convolutional Neural Network for Plume Rise Measurements in Industrial Environments
Round 1
Reviewer 1 Report
Summary/Contribution: This work proposes a novel framework based on Deep Convolutional Neural Networks to monitor smokestack plume clouds and make long-term, real-time measurements of the plume rise, which outperforms widely-used networks in smoke border detection and recognition.
Comments/Suggestions: 1. The authors are invited to include some numerical results in the abstract. 2. Line 125: "Then, section ?? present" ===> The authors must pay attention to missing references and English mistakes. General and deep proofreading is needed. 3. The authors need to clearly identify the limitations of existing similar works in order to emphasize the originality of their work. 4. Line 18: "Smokestack Plume Cloud (PC) rises due to momentum and buoyancy" ===> These two notions may be new to a wide number of readers. The authors are invited to provide more details about them. 5. Line 18: It is a bit strange to use the word "Finally" in the first line of the paper.
6. The authors are invited to provide more details about the different types of atmospheric conditions that the collected dataset has been classified into, and how many valid images were ultimately obtained from the 35K images collected.
7. The authors should include a paragraph on how formal methods might be used to verify AI-based solutions, notably data collection and processing.
8. For this purpose, the following references may be included:
Proofreading is needed.
Author Response
We have attached the response report to reviewer 1.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
1. The research on related work is not up-to-date enough, and it is recommended to add the latest research techniques to the cited literature.
2. An error character appears on line 125.
3. There is limited analysis of Figure 11. Why is it easier for some methods to have accuracy, recall, and other indicators of 0 when a higher number of test images are input? What is the reason for this?
4. How many test images were used in Figure 12, and what changes were made compared to Figure 11? Why did the accuracy and recall indicators not show 0?
5. There is no comparison of the accuracy of the experimental results in Table 3. How accurate are these values? How to determine whether it is accurate and valuable?
6. It is recommended to add MIOU indicators in the comparison results.
Minor editing of English language required.
Author Response
We have attached the response letter to reviewer 2.
Author Response File: Author Response.pdf
Reviewer 3 Report
The manuscript presents a three-layer framework to estimate plume rise from smokestacks. The proposed method is utilised to calculate the PC from digital images in real time. The main contributions are clearly explained, including a novel benchmark dataset that will facilitate further research in the plume cloud detection. The proposed model is compared against relevant state-of-the art.
The manuscript is well structured and provides a sound solution. However, some points to consider:
Line 125: missing reference to section 4.
Figure 2 could improve to distinguish processes from inputs.
It is understood that the result of the third block is the coordinates of the NBP in 2D (image). These are referred to here as (x, z). If these are the usual 2D coordinates, it is unclear why z is used here instead of a traditional y variable unless a top view is adopted where is becomes height (rise) information. Equation 8 that estimates the NBP location operates on x and y variables. Sometimes z is used, other times y is used as coordinates. Could this be clarified?
The proposed model does not rely on Brigg’s estimate and the manuscript critiques the method to be unreliable and very inaccurate. Delta z was not calculated via the fluxes in Equation (1). Why the focus on this formula then?
It is unclear what anchor coordinates are.
Although the approach is clear, the sentence explaining binary classification of upper and lower boundary points (lines 272-274) is very confusing. Consider rephrasing in a more precise way.
It would be helpful if the asymptotic function used to estimate the NBP is justified or explained in terms of the central line or PC behaviour. It seems an arbitrary choice unless clarified.
How were the parameters in Equation 8 identified? Also, the y variable is the estimated coordinate of the NBP. Should this be z in accordance with Equation 8?
Y_NBP is declared as the “depth” of the NBP in real life. How is this value determined? Is this a direct transformation of the y coordinate from the physical model?
It would be helpful is all quantities are consistently referred to in the text and in figures.
What is the link between the NBP and the transformation? What is the result of the NBP extraction system? It is expected these produces (x, z) image coordinates of the buoyance point. How does this link to the following transformation section?
Update: Figure 8 resolved some of the confusion regarding coordinate systems and variable names. Consider having it earlier than this point in the manuscript.
Figure 9 uses a a-b-c coordinate system.
Consider using a consistent coordinate system and variable names across the manuscript.
Tests against deep models for PC detection is done with classic suitable metrics. What about estimation of the NBP? Do method exist to verify (even roughly) how accurate are these estimations? Can this be discussed?
Overall, organised and structured paper. Good use of figures to support the theory. Could use more clarification and consistency explaining the model. Analysis and validation are relevant but could use more argument defending the system. Limitations and shortcomings should be discussed.
Author Response
We have attached the response report to reviewer 3.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The authors considered my comments and suggestions. Good luck.
Can be improved.
