Cloud Detection of Gaofen-2 Multi-Spectral Imagery Based on the Modified Radiation Transmittance Map

Round 1

Reviewer 1 Report

1. The motivation of the study needs to be more clear. Kindly rewrite it with more emphasis on the usefulness of the present work in real time applications.

2. What are the references for Eq.9 and 10?

3. Please list some of the shortcomings of K-Means method as well here.

4. Can you discuss more about the Grey Level Entropy in Table 1?

5. Section 4.2 (Limitations and future scope) should come after the conclusions.

6. Conclusion in its present form needs to be addressed. Make it more simple and may be point wise so that readers can follow it easily.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Review of

 

Cloud Detection of Gaofen-2 Multi-spectral Imagery based on the Modified Radiation Transmittance Map

 

By Yi Lin, Lin He, Yi Zhang, Zhaocong Wu.

 

Overview

This paper presents an interesting approach for cloud detection based on the radiance information estimated from the remote sensing images. In general, the authors developed an algorithm based on the radiance estimation corrected by the spectral features obtained from the multispectral (R, G, B, and NIR channel) images. These images were processed via a simple K-means segmentation and subsequently corrected via several image post-processing methods for extracting the unwanted object from the generated segmentation clusters. As a result, the authors presented the cloud cover maps for tested images from the Gaofen-2 and Gaofen-1 imagery and compared the results to a different cloud cover detection method and ground truth images. The results are assessed via two main metrics, Pixel accuracy and advanced F-score. The authors conclude that their method shows a better accuracy for the Gaofen-2 images than the compared method and that their algorithm reaches 95% cloud clover detection accuracy for the tested images. 

 

Major comments

The detection of the thin cloud layer within the remote sensing image is nowadays really pressing, and the authors show an interesting approach for its detection. The overall aim of the paper is quite clear. However, from my point of view, the overall structure is a bit inconsistent, and the paper contains several issues that the authors should address. The following lines cover my major comments and recommendations for the authors. 

 

1. In the extensive Introduction, the main contribution of the paper and the main problem that the paper aims to solve are missing or unclear. I would recommend focusing on that in the latter part of the Introduction and emphasizing the contribution of the paper. 

2. Consider dividing the “Materials and Methods” only into “cloud detection” and a “post-processing” section. I believe the additional splitting of the subsections is unnecessary. Also, the division of the Results section seems to be unnecessarily detailed.  

3. In the section on K-means segmentation, you mention, “K-means algorithm is operated on the extinction coecient map…”. How is this extinction map generated? It would be beneficial to see how this map/image is generated. Please explain this, ideally mathematically, and include an example image in Figure 3. 

4. There are several number results of the proposed method within the text of the “Results” section 3. Personally, I would recommend summarizing these numbers in the separate Table or Tables in the mean ± variance format. It will simplify the investigation of the method performance and comparison to other methods presented within the text. It will also improve the understanding of the presented numbers, whether they belong to the F-Measure or Pixel accuracy, which is sometimes unclear.  

5. Complementary to the previous point, there are several mentions of the accuracy or pixel accuracy in the text. I would recommend mathematically defining this accuracy in the section where the metrics (eq. 16-18) are presented.  

 

The following section contains my comments, suggestions and questions in a more detailed manner.
 

Minor comments

 

• What does “good performance" mean when referring to the citation [9]?

 

• I believe you wanted to use “color channels” instead of “color spaces” in the Introduction. 

 

• “low frequency bands” - In terms of spectral processing, consider replacing this with “infrared bands”.

 

• Figure 1 is not referenced within the text.

 

• Consider renaming the section “Materials and Methods” only to “Methods” or “Methods and Algorithms”.  

 

• What mathematically represents “(o)” and “(p)” symbols in eq. 1? 

 

• “China Centre for Resources Satellite Data and Application” I would recommend adding a footnote URL to this site.

 

• What is meant by “DN” in eq. 7, and what does this symbol represent in this equation?

 

• Please, ensure that all symbols used in equations are explained in the text. 

 

• There are several empirical thresholds, such as i_mean and i_max. What are the typical values of these constants? 

 

• “Five pixel sets will be generated as A1, A2, A3, A4 and A5, and Ai represents the set of pixels in i-th cluster”.How do the pixel sets differ from the generated clusters?

 

• What represents the parameters i_v and i_d, and how were they obtained?

 

• What does “filtering the cluster results” mean? Please explain in the text. 

 

• “The detection results of coarse cloud with cluster number of two to five in Fig. 3 are 0.8327, 0.9209, 0.9443 and 0.9449, respectively.” How did you calculate these results? 

 

• What are your computation requirements for the presented algorithm? Maybe a graph showing an increase of time calculation requirements against the increased number of clusters and the increased clustering precision against the increased number of clusters could better explain the reason why the number cluster selection is equal to 5. 

 

• How do you define “Discriminability” in the Texture feature selection? Please explain.

 

• “Parameter decision” section: Were these parameter settings tested on unseen data? 

 

• Figure 5: The results of your algorithm after brightness hyperparameter selection should be included in this figure to compare them against the ground truth. Otherwise, this figure is unnecessary. 

 

• “The performance of the proposed method was evaluated mainly on Gaofen-2 high spatial resolution remote sensing images.” Did you test your algorithm on other data than from Gaofen-2 and Gaofen-1 imagery? How did it perform?

 

• How did you divide the Gaofen-2 images into categories “stratus, stratus fractus, altostratus, cumulus, cirrocumulus, stratocumulus” ? Manually? 

 

•  “stratus, stratus fractus, altostratus, cumulus, cirrocumulus, stratocumulus” Why did you select these cloud types precisely? What about other cloud types? 

 

• Did you also test the cloud detection results with different settings of the alpha parameter included in the F-measure? 

 

• How were the ground truth images of the cloud cover obtained in Fig 6 - 10?

 

• Was it necessary to divide the images into the 6 classes according to the cloud types? If the results data is placed in the table, it would be much easier to assess it. 

 

• Did you assess images that involve several cloud types in them?

 

• “Here is another example of the experiment on stratus (Fig. 7). The extinction coecient map was obtained from the radiation transmittance map modified by brightness mask and water mask. After clustering by using K-means algorithm, the grade of cloud density variation can be clearly identified from the clustering result (Fig. 6(e)). Filtering and post-processing were carried out to get the further fined cloud map.” Does this process change from the previous ones, or is it only a repetition of the algorithm? If so, I think it is not necessary to repeat it in the results section. 

 

• Figure 9: Please add the explanation of the yellow and red highlighted parts of the image to the caption of Fig. 9. 

 

• What do you mean by “loose conditions” or “relaxed conditions”? How did you change the parameters of your algorithm? Please explain in the text. 

 

• Did you test the performance of your algorithm against other present cloud detection algorithms on the Gaofen-2 data? Maybe, it would be beneficial if you could try. 

 

• Did you adjust your algorithm for the data from Gaofen-1 imagery, or did you use the same parameters as for Gaofen-2 images? 

 

• “our algorithm is compared with that of Zhang’s algorithm [17]” How does your algorithm differ from Zhang's?

 

• Figure 10: Please add a citation of Zhang's algorithm [17] in the caption of Fig. 10 

 

• Why did you select only Zhang's algorithm for the performance comparison?
  
  

Additional questions

• Is it possible to use your algorithm for cloud detection automatically and autonomously on the unseen data?

 

• What happens to the segmentation results if no cloud cover is present within the image? 

 

• Would it be possible to detect different cloud types via your algorithm?

 

Bibliography

 

• Did you search for the research papers that assess different cloud types on the images during the cloud detection/segmentation? Some could be added to the cited bibliography.
  
  

• Please add the citation of the “Koschmieder’s law” in the text (e.g. “Middleton, William Edgar Knowles. Vision through the atmosphere. University of Toronto Press, 1952.”)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Review of the paper «Cloud Detection of Gaofen-2 Multi-spectral Imagery based on the Modified Radiation Transmittance Map»

by the authors: Yi Lin, Lin He, Yi Zhang, Zhaocong Wu

 

General

I must say right away that the content of the article does not apply to my area of interest. Therefore, my remarks will be of a general nature.

The paper is devoted to the development of an algorithm for the detection of clouds against the background of the image of the Earth's surface by remote sensing from space.

A similar problem occurs in radar meteorology. In this case, one must be able to distinguish the so-called local objects from the clouds. The problem is solved simply. In cloudless weather, a reflectivity pattern is obtained, and then it is subtracted from the overall picture in the presence of cloudiness. I think that the same can be done in this paper. But for some reason, the authors complicate the task. I did not understand the meaning of such a complication.

Therefore, I would like the authors to be able to explain the essence of the problem for a wide range of readers, and not just for programmers.

 

Remarks  

1.    At the end of the line before formula (1) there is an extra dot. Before the formula, there can be either nothing, or a comma, or a colon. This remark applies to the entire text of the article.

2.     After the formulas, you must put either a period or a comma, depending on the context of the sentence.

3.    In formula (8): what do the values , ? It is clear that these are the reflection coefficients in the green and near infrared regions. But it should be written in the text of the article.

4.    Page 10. Heading 3.1 Mistake in the word "experimental".

 

Conclusion

Since the comments are editorial in nature, I think that the article can be recommended for publication after the correction of the comments made without re-reviewing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

In this paper, a novel cloud detection method is proposed for Gaofen-2 remote sensing imagery. Overall, the research theme of this paper well-fits the scopes of Remote Sensing. Reviewer appreciate all of author's effort for the manuscript. The paper is innovative to a certain extent, but there are still many problems that need further revision.

(1)   First of all, what are the criteria for distinguishing thin clouds from thick clouds, and what are the thresholds for spectra or other indices？

(2)   The extraction method in this paper can be compared and analyzed with other existing methods to further increase the research significance of this paper.

(3)   Please explain how the accuracy extracted in the paper is calculated. There is no description of the accuracy estimation method in the section of Methods.

(4)   In introduction, We should first introduce the current methods of cloud extraction and the shortcomings of high-resolution images, and then introduce the necessity of choosing Gaofen 2. The last paragraph will talk about the main problems solved in this paper, instead of introducing what each section is.

(5)   The threshold value of water extraction in different images should be different. How to determine?

The full text of the formula size and style need to be unified.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have done a very good job revising the paper. The suggestions and questions have been fully addressed and answered. I have only a few additional comments on the current state of the paper.  

 

Comments:

 

• The addition of Table 3 has improved the clearness of the results. However, in the text, several mentioned results were still not summarized in a separate table, especially the comparison with Zhang’s algorithm. Please consider adding a separate table with these results. 

. 

• In the caption of Table 3, please state to what data these results belong (Gaofen-2, I believe) and if these results are based on your algorithm. The caption of the stable should independently describe to a potential reader what kind of data is included in the Table. 

 

 

Upon attention to these small details, I believe this paper is acceptable for publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

The paper has been carefully revised and recommended for publication

Author Response

Thank you for your revisions. We further revised the article.