A New Identification Method for Surface Cracks from UAV Images Based on Machine Learning in Coal Mining Areas

Round 1

Reviewer 1 Report

The manuscript documents a method for the identification of surface cracks from UAV imagery based on machine learning in coal mining areas, which is an improvement to the traditional laborious in-field surveys. However, there are some points that need to be improved before the manuscript being considered for publication. The methodological approach is not clearly identified and there are some flows and contradictions that needs to be addressed. Moreover, the dataset must be reconsidered.

Please read my comments and sugestions in the attached PDF.

Comments for author File: Comments.pdf

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.pdf

Reviewer 2 Report

Abstract. Page 1. The authors should avoid using acronyms in the abstract. If they are needed, they must be defined.

Abstract. Page 1. Although it is OK to put some of the results in the abstract, in this case, I believe the attention is placed too much on the results, rather on shortly describing the logic behind the proposed method to highlight the main innovative contributions.

General comment. Along the manuscript, the authors should define again all the acronyms.

Introduction. Pages 1-2. The state-of-the-art overview should be significantly improved. The authors should consider adding more reference regarding

• Crack detections using images. A few suggestions are provided below

Mohan, A., & Poobal, S. (2018). Crack detection using image processing: A critical review and analysis. Alexandria Engineering Journal, 57(2), 787-798.

Fei, Y., Wang, K. C., Zhang, A., Chen, C., Li, J. Q., Liu, Y., ... & Li, B. (2019). Pixel-level cracking detection on 3D asphalt pavement images through deep-learning-based CrackNet-V. IEEE Transactions on Intelligent Transportation Systems.

• Crack detections using other sensors, such as LIDARs. A few suggestions are provided below

Zheng, X., & Xiao, C. (2018). Typical applications of airborne lidar technolagy in geological investigation. International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 42, 3.

Glenn, N. F., Streutker, D. R., Chadwick, D. J., Thackray, G. D., & Dorsch, S. J. (2006). Analysis of LiDAR-derived topographic information for characterizing and differentiating landslide morphology and activity. Geomorphology, 73(1-2), 131-148.

Kasai, M., Ikeda, M., Asahina, T., & Fujisawa, K. (2009). LiDAR-derived DEM evaluation of deep-seated landslides in a steep and rocky region of Japan. Geomorphology, 113(1-2), 57-69.

• Application of image processing techniques to process camera images on board of UAVs, especially using machine learning techniques (such as deep learning). A few suggestions are provided below

Ammour, N., Alhichri, H., Bazi, Y., Benjdira, B., Alajlan, N., & Zuair, M. (2017). Deep learning approach for car detection in UAV imagery. Remote Sensing, 9(4), 312.

Opromolla, R., Inchingolo, G., & Fasano, G. (2019). Airborne visual detection and tracking of cooperative UAVs exploiting deep learning. Sensors, 19(19), 4332.

Zeggada, A., Melgani, F., & Bazi, Y. (2017). A deep learning approach to UAV image multilabeling. IEEE Geoscience and Remote Sensing Letters, 14(5), 694-698.

Carrio, A., Sampedro, C., Rodriguez-Ramos, A., & Campoy, P. (2017). A review of deep learning methods and applications for unmanned aerial vehicles. Journal of Sensors, 2017.

Bah, M. D., Hafiane, A., & Canals, R. (2018). Deep learning with unsupervised data labeling for weed detection in line crops in UAV images. Remote sensing, 10(11), 1690.

Introduction Page 2. Lines 51-61. The authors should provide proper references to justify their statements about processing drawbacks of object-oriented, edge detection and artificial interpretation methods.

Introduction. The authors should synthetize the main contributions provided by the manuscript to the literature.

Introduction. Figure 1. The building blocks of the proposed method are synthetized in Figure 1. However, a clear description of the figure is missing thus compromising its usefulness in understanding the proposed logic. For instance, it is not clear what do you mean by Result 1, 2 and 3, respectively.

General comment. Please, be sure that all the parameters introduced in the equations are correctly defined. For instance, in equation (1) and (2), this is not the case for w, x, b, etc.

Section 2.4.2. Pages 6-7. For the sake of clarity, the authors could also add some examples of 50x50 sub-images for both the crack and no-crack type.

Section 2. Page 7. Lines 196-207. The authors should clearly state the rationale behind implementation logic of the proposed method. For instance, what is the discriminant factor for choosing either NVDI or EVI? What is the reasoning behind the 10% level and how do you select this percentage threshold? Also the authors should specify from the beginning of the section the typology of multispectral camera clearly indicating its spectral resolution and the total number of different analyzed bands.

Section 2. Table 2. It is not clear why the number of no-crack images does not reach the 18405-value mentioned at line 190.

Section 3. While the training dataset is clearly described in the previous section, the dataset used for testing is not explicitly described. Beware that it should be composed of different images, than the ones used for training.

Section 3. In this reviewer’s opinion, it is not correct to define the percentage of correct declaration as an accuracy metric. Rather it may be defined as “success rate”.

Section 3. Table 4 and 5. Could you clarify which is the worsening in computational efficiency which is expected to be associated with an increased number of trees in Table 4? Similar question for the k-value analysis in Table 5.

Section 3. Table 9. The authors should justify why the M2GLD method has a worsening effect on the success rate of the classification process.

Section 4. In this reviewer’s opinion, section 4.2 is not necessary as it is. Indeed, some of its content, e.g., lines 411-430, should be placed in the introduction, while the remaining part should be integrated with the conclusion.

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper discusses a comparison between several methods of machine learning applied to UAV images to detect ground cracks. The authors presented their hypothesis and experiments well, and the results look convincing. I do have some minor remarks:

1. Please add a short text about GSD (Ground Sampling Distance), I think it is very important in this kind of study. Otherwise, how else should we know how much area does 50x50 pixel represents? Another idea for research would be to understand which image resolution is good for this kind of study.
2. The title mentions the word "novel", and this is repeated many times in the text... The way this reviewer sees it, not a lot of the tested algorithms is "novel", neither is the use of machine learning for tasks like this. I would suggest to change the title and the focus of the paper more on the comparison part, which for me generated the most interesting result.
3. Please revise the English language, I see the use of some uncommon aphorisms ("aeolian sandy area") and grammatical errors are still rampant. 

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.pdf

Reviewer 4 Report

The article presents a machine-learning-based approach for extracting surface crack information from UAV-acquired image data.  Specifically, the article describes the procedural steps, through comparisons of different techniques, used to arrive at the final algorithm as presented.

I freely admit that my expertise does not have significant overlap with image analysis.  However, I believe that the article is well written and presents a thorough justification of the approach taken by the authors.  I therefore feel that this article merits publication in `Remote Sensing'.  

I do have some minor comments I would like to see addressed first however:

1) The authors use a training set of 1590 images to build their machine learning based model.  How was positive crack identification conducted for the training data?

2) The final success rate of the identification algorithm is 90%.  Were the incorrect identifications false-positives, or missed cracks?  What is the ultimate impact of the 10% failure rate?

3) The fonts are too small and overall legibility of Figure 6 is poor.

4) Page 14 incorrectly refers to Figure 8 as Figure 11.

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors improved the manuscript according to the comments from the previous review. The manuscript structure is now better organized.

However, there is only one reference in the Discussion Section, the authors must compare their results with similar studies or studies with different purposes that presented a similar methodology (e.g. machine learning comparison for other purposes using UAV data).

Other comments:

Lines 90-92: replace "parts" by "sections".

Lines 98-103: Confuse, these sentences must be rephrased.

Lines 119-132: This part of the methodology is still not clear. The authors must explain better what is the "10%". Include that it reffers to 10% of the area percentage.

Include the figure of the multispectral image data of one UAV image provided in the cover letter.

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.pdf

Reviewer 2 Report

I would like to thank the authors for addressing my concerns. A few minor modifications are still required in my opinion.

Introduction. Lines 53-55. Regarding this sentence “Unmanned air vehicle(UAV) images have significant advantages, such as high resolution, flexible maneuverability, high efficiency, and low operating costs”, I would like to propose a slight modification for the sake of clarity. In fact, high resolution may be considered as a characteristic of images taken from UAVs. However, maneuverability, efficiency and low operating cost are characteristics associated to the use of UAV rather than to the use of images. So I would propose “Unmanned air vehicle(UAV) have significant advantages, such as high resolution remote sensing images, flexible maneuverability, high efficiency, and low operating costs”,

Introduction. Regarding the answer to point 3 in the original review. The authors have not added to the literature review the journal article below

Fei, Y., Wang, K. C., Zhang, A., Chen, C., Li, J. Q., Liu, Y., ... & Li, B. (2019). Pixel-level cracking detection on 3D asphalt pavement images through deep-learning-based CrackNet-V. IEEE Transactions on Intelligent Transportation Systems.

This article is important since it is a very recent publication about surface crack detections from images using deep learning. Although the application is from cars and used to detect cracks in asphalt pavement, it can be a useful reference for the image processing side.

Also, at lines 78-80, besides reference [18] (published in 2017), the authors could consider adding other suggested references to account for more recent results and applications about use of  deep learning from UAV images, to allow the reader understanding this is a very hot  topic in the literature

Ammour, N., Alhichri, H., Bazi, Y., Benjdira, B., Alajlan, N., & Zuair, M. (2017). Deep learning approach for car detection in UAV imagery. Remote Sensing, 9(4), 312.

Opromolla, R., Inchingolo, G., & Fasano, G. (2019). Airborne visual detection and tracking of cooperative UAVs exploiting deep learning. Sensors, 19(19), 4332.

Zeggada, A., Melgani, F., & Bazi, Y. (2017). A deep learning approach to UAV image multilabeling. IEEE Geoscience and Remote Sensing Letters, 14(5), 694-698.

Bah, M. D., Hafiane, A., & Canals, R. (2018). Deep learning with unsupervised data labeling for weed detection in line crops in UAV images. Remote sensing, 10(11), 1690.

Section 2. Data source. Please add the focal length of the camera as additional information.

Section 2. Line 120. Acronyms should be defined. Please add NDVI definition as also provided in the cover letter.

Section 2. The authors should add in the manuscript the explanation about the empirical selection of the 10%-threshold on NDVI as well as on the 168-threshold about the image gray level, also including appropriate reference to the literature if necessary.  It is important, as explained by the authors in the cover letter, to clarify the readers about the need to carefully select the thresholds based on the application scenario.

Section 3. The answer to point 13 in the original review is not satisfying.

“Section 3. Table 4 and 5. Could you clarify which is the worsening in computational efficiency which is expected to be associated with an increased number of trees in Table 4? Similar question for the k-value analysis in Table 5.”

In fact, I was asking a clarification about the effect of the selected number of trees and k-values on the computational efficiency (i.e., on the run time), not on the accuracy.

Section 4. The answer to point 15 in the original review is not satisfying.

“Section 4. In this reviewer’s opinion, section 4.2 is not necessary as it is. Indeed, some of its content, e.g., lines 411-430, should be placed in the introduction, while the remaining part should be integrated with the conclusion.”

I agree that a discussion about the results including highlights on current limitations is an important part of a journal article. However, as it is, section 4.2 mainly contains a summary of the manuscript content. My advice is to concentrate all the statement regarding the significance of research in the introduction. From a logical point of view this allows the readers to understand from the beginning the importance of the proposed method.

Author Response

Thank you so much for your comments regarding our manuscript entitled " A novel identification method for surface cracks from UAV images based on machine learning in coal mining areas " by Fan Zhang et al. submitted to Remote Sensing. We have revised the manuscript according to the reviewer’s comments. Please see the attachment for the specific details of the manuscript modification. Thank you for your comments again.

Author Response File: Author Response.docx