A Vision-Based Detection and Spatial Localization Scheme for Forest Fire Inspection from UAV
Round 1
Reviewer 1 Report
Very well done.
The only remark I have is the order of 10^-5 in Abstract and Conclusion. 10^-5 of what? I guess degress but it needs to be stated. Also, to avoid confusion, add units to the header of Table 3.
Section 5 could be extended a bit but no necessarily.
Author Response
Point 1: The only remark I have is the order of 10^-5 in Abstract and Conclusion. 10^-5 of what? I guess degrees but it needs to be stated. Also, to avoid confusion, add units to the header of Table 3.
Response 1: Yes, this is measured in degrees. We have made adjustments in the corresponding sections of the article and added explanations to Table 3.
Point 2: Section 5 could be extended a bit but no necessarily.
Response 2: After comprehensive consideration, we did not expand in Section 5. Thank you very much for your suggestions.
Reviewer 2 Report
The paper presents a novel approach to image processing for the discrimination of wildland fires within images. The processing methodology is well presented but the results presented are quite limited and offer something close to a best-case scenario for detecting the fires as there is little in the way of smoke to obscure the flames the algorithm was trained to identify. For a real wildfire there would be significant amounts of smoke and water vapor providing obscuration. While this just presents another challenge to train the AI to cope with, it should be noted that the test case presented is a best-case scenario that may not represent actual wildland fires.
Some additional points: What is the UAV speed along the flight path for the test case? What is the size of the target used (the iron pot), both the physical size and size in pixels in the images? How much does performance degrade if UAV flew at a height of 50m? How does the algorithm perform for oblique detection? The frames shown appear to have the target centered in the frame. For scanning a large area oblique detection would be vital.
Author Response
Point 1: The processing methodology is well presented but the results presented are quite limited and offer something close to a best-case scenario for detecting the fires as there is little in the way of smoke to obscure the flames the algorithm was trained to identify. For a real wildfire there would be significant amounts of smoke and water vapor providing obscuration. While this just presents another challenge to train the AI to cope with, it should be noted that the test case presented is a best-case scenario that may not represent actual wildland fires.
Response 1: Our test experiment was relatively idealized, and primarily focused on the performance of forest fires detection with flames. We have done vision-based smoke detection before, but the results are not great in practice, with a lot of false alarms. Therefore, the problem of how to detect wildland fires in consideration of the obscuration of smoke and vater vapor is also one of our future research.
Point 2: What is the UAV speed along the flight path for the test case?
Response 2: In this test case, the flight speed is roughly 1~2 m/s.
Point 3: What is the size of the target used (the iron pot), both the physical size and size in pixels in the images?
Response 3: The physical diameter of the iron pot is about 60cm, and its pixel size in the images is roughly 40×40.
Point 4: How much does performance degrade if UAV flew at a height of 50m?
Response 4: Although we have not conducted the relevant experiments, we guess that if the target remains small, like in the test case, there will be a significant loss of performance with the same camera resolution. In order to achieve controllability and not cause accidents, the physical size of the equipment used for our experiments is relatively small. The number of pixels representing the fire area in the image decreases as the distance increases, and much detailed information is lost, resulting in blurred targets and weaker recognition. Therefore, if the fire target could be larger, there would be almost no loss of performance, but this would be dangerous and uncontrollable for the experiment.
Point 5: How does the algorithm perform for oblique detection? The frames shown appear to have the target centered in the frame. For scanning a large area oblique detection would be vital.
Response 5: For oblique detection, the performance is the same as long as it is within the camera's receptive field, regardless of whether the target is in the center or the edge of the image. Since it is a vision-based solution, the actual possible situations such as tree occlusion will cause interference, which needs to be solved by developing a suitable flight path.