FIgLib & SmokeyNet: Dataset and Deep Learning Model for Real-Time Wildland Fire Smoke Detection
Round 1
Reviewer 1 Report
In this manuscript, a publicly-available dataset of nearly 25,000 labeled wildfire smoke images alongside a novel deep learning architecture is presented. Early detection of fire ignitions is essential to minimize environmental destruction, especially in Mediterranean-like climates. The authors used 24.800 images for training, validation, and testing of the fire smoke detection model. Overall, the manuscript is well-structured and it is apparent that a lot of effort has been put into this research. The aim and objectives are clear, the novelty of the manuscript is clearly stated, and the results-discussion is well-structured and valid.
Some minor recommendations:
- 46-52 “We also introduce SmokeyNet, a novel deep learning architecture ……..reduce the time to wildfire response.” Comment: These sentences are identical to the abstract. In my opinion, these sentences should be rephrased.
- Figure 1 and Table 1 should be moved to the following sections (3.1 and 3.2 respectively).
- Table 2 should be moved to the results.
- 249 Remove parentheses.
- 257 Add a title in this section (5.1) or simply remove the “5.1 Human Performance Baseline”.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The text presents a magnificent dataset and study of different deep learning models for smoke detection in forest fires. The article is very interesting and I believe that it provides the scientific community with a magnificent labeled dataset that will surely be very useful for future studies. In addition, it presents different methodologies for early fire detection with convolutional networks, comparing results between them, as well as with three human experts. With very interesting and promising results.
Below are some minor issues for consideration:
It would be advisable not to use exactly the same paragraphs in the abstract description that are used later in the introduction.
In Figure 1, I think it would be advisable to place it closer to your reference. In any case, I think it would be simpler if each subfigure had its corresponding letter (a, b, c, d) added to it and in the figure caption, this reference is indicated for consultation. Also, in my opinion, section 3.1 should not start with the example of figure 1 but with the dataset data and then put figure 1 as an example.
I think that table 1 could be shown just before section 3.2. In the table footer (which is usually used as a table title instead of footer) the word "with" appears twice in a row.
In section 3.2, I believe that the unusual (80/10/10) use of a low percentage of fires for training should be justified. As well as expanding the justification of the sentence: "Splitting the data by fires instead of images ensures that no data related to the test set is in the training set".
I believe that the "additional transformations" introduced in the last paragraph of section 3.2, deserve a broader and more concrete explanation. Basically it is the preprocessing that is being used in the system and at least it would be interesting to see a block diagram of it, with the different steps.
Special mention I think deserves the cropping of a very specific number of top rows introduced in the images. Why that number? I guess it will be done before splitting the images into smaller blocks right?
I think figure 2 should be closer to section 4.2 where it is referenced.
Table 2 I think should be shown after section 4.5 and not before. Also, the metrics used in Table 2, although they are the usual ones in this type of studies, I think they should be explained and/or at least referenced.
The results shown in the video and illustrated very conveniently in Figure 3, are very interesting and show perfectly what really happens in reality (although it is a pity that the fire sequences shown in the video and in the figure do not coincide temporally). As already indicated, many false alarms due to clouds are observed. I think that an interesting improvement could be a second post-processing stage where the quadrant that has a successive number of alarms detected over time is marked as a fire. In this sense, I think it is very interesting to try to include the successive time variable in deep learning. Perhaps not only with a single previous frame but with a wider sequence. Of course, other experiences with thermal imaging greatly simplify the process of detection and elimination of false alarms.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The monitoring and early warning of natural disasters based on big data and artificial intelligence is of great significance. It combines the latest technology and significantly promotes the monitoring and early warning of natural disasters such as forest fire. This study constructs a database of nearly 25000 forest fire smoke signs for fire source identification, monitoring and early warning based on deep learning method. The specific problems are as follows: (1) the contents of the discussion of the study are not enough, the results should not be put together with results and analysis, and the discussion should be independent and detailed, and further studies should be discussed. The authors should discuss the advanced parts and limitations of the study and the further work to be done in the aspects on data, methods, results and so on. (2) The content in the current conclusion and the work to be carried out in the further stage should be simplified, and more content should be put in the discussion section. There should be no more references in the conclusion section.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf