Post-Processing for Shadow Detection in Drone-Acquired Images Using U-NET

Round 1

Reviewer 1 Report

The authors implemented post-processing methods related to automatic thresholding and binary mask refinement. The proposed study aimed to improve shadow detection results in drone-acquired images.

Unfortunately, the paper contains the following problems.

    • The paper language needs polishing since it contains errors.

    • All figures (except figures 2-4 containing photos) inside the text must contain a small paragraph explaining them. This paragraph must be placed next to each figure’s caption.

    • The authors need to divide section 4 into two sections (“Results” and “Discussion”).

    • In the “Results” section, a table summarizing the experiments’ parameters must be added. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors proposed an approach to automatic shadows detection in images acquired by drones. The novelty of this method is marginal. It is composed of well-known algorithms and operations like U-net convolutional network, opening, closing, CRF and automatic thresholding. I think the paper may be reconsidered for publication after the authors compare their method with other known approaches (comment 1) and if the proposed approach proves better. The experiments carried out so far are not sufficient.

Comments

1.      In subsection 2.2.3 the authors write about AISD public dataset. Why was this dataset not used to validate the authors' method? If it is possible please compare your method with other methods from the literature on AISD or other publicly available datasets like the ones from which the authors took several images to form the new dataset. Such comparison is important especially because the dataset used in the experiments consists only of 50 images (including only nine images for testing). It is not enough to properly validate a method.

2.      Please provide an image visualizing the structuring elements mentioned in subsection 3.4.2 (cross, rectangle and ellipse).

3.      Please define w, k and K from equation 5.

4. Each of the results was evaluated quantitatively using several evaluation metrics explained in subsection 3.4.2, and qualitatively by observing and comparing the visual result.". It should be "subsection 3.5.2".

 Language

The following words and sentences require correction:

·         "Detecting shadows is one of the problems that are widely discussed in the field as shadows frequently visible in images captured by drones and becoming obstacles." - Something is wrong with this sentence. Maybe it should be "are frequently visible" and "become obstacles"?

·         "Thresholding method is easy to implement, unfortunately, a lot of weaknesses" - It should be "implement. Unfortunately".

·         "The drawbacks of these methods are that it depends mainly on the knowledge of the presence of shadow in different channels of color model, which also give some inaccuracies in the result as it cannot be detected perfectly." - I think it should be "they depend", "which also gives"

·         "added the result with another result from image segmentation" - It should be "to another result"

·         "shadow analysis needs to be done such as clustering analysis." - It should be "to be done, such as".

·         "Otsu’s thresholding method also has been commonly implemented" - It should be "has also been".

·         "CRF also has been implemented as a" - It should be "has also been"

·         "Otsu’s thresholding for automatic thresholding" - I think it should be simply "Otsu’s thresholding using OpenCV function".

·         "Figure 2. Sample Image of SenseFly Drone Dataset" - It should be "image" (small letter).

·         "the geometric transformation data augmentation technique is applied which includes 90 degrees clockwise and 90 degrees counterclockwise." - I suggest writing "the geometric transformation is applied as data augmentation, which includes rotation by 90 degrees clockwise and 90 degrees counterclockwise.".

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

"Response 1: Instead of using AISD dataset, a new dataset is prepared for this study to focus on low altitude aerial images as images in AISD dataset are taken at high altitude (mentioned in section 2.2, 3.1.1)"

The comparison with other methods should be performed. It does not matter on what dataset (AISD was just an example). As I stated before, the experiments carried out so far are not sufficient to properly validate a method.

The other comments have been properly addressed.

Author Response

Comment by reviewer:

The comparison with other methods should be performed. It does not matter on what dataset (AISD was just an example). As I stated before, the experiments carried out so far are not sufficient to properly validate a method.

Response:

Thank you to the reviewer for the comment. We have been able to include some changes to reflect the suggestion made by the reviewer to perform comparison with other method. We provided the comparison of U-Net with PSPNet as in subsection 4.2.

Round 3

Reviewer 2 Report

The comparison with one method has been made. It is only one method, but it is better than nothing. I still think the novelty of this paper is low (maybe low to average). However, if the editor thinks this is sufficient I think the paper can be published in present form.