Review Reports

Round 1

Reviewer 1 Report

In this paper the new method of cloud detection is presented. This method is based on multiscale data and the dual channel attention mechanism. Since the use of the multiscale data is usually applied, this improvement seems to be not so new. But the application of the dual channel mechanism is original. This paper is well written and presents valuable results. I have no further remarks.

Author Response

Thank you for your approval of this manuscript. We do cite the dual attention mechanism as the main contribution of this paper. In this revision, we also describe the main contributions of this paper based on the original, as follows:

·         The band information and angle information provided by the data are fully utilized. The influence of different band information and different observation angle information on experimental accuracy is also considered.

·         Use 3D U-NET as the benchmark network model. While classifying pixels, the texture information of clouds is preserved as much as possible. This benefits from the jump connection structure between encoder and decoder.

·        A dual-channel attention mechanism is proposed to extract useful information from band and angle respectively.

We also made some minor changes to the manuscript to improve it.  

Reviewer 2 Report

The manuscript describes the preliminary results for a cloud detection approach using multivariate data (observation angles and wavelengths) to feed a deep learning method.

The topic is relevant, and then it is worth publication. However, I believe the authors should wait to get more satellite data before publishing. The future paper will present very similar content but more definitive work.

The pdf file provides a few suggestions to improve the paper presentation.

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

This study proposed a novel cloud detection algorithm based on multi-scale input and dual-channel attention mechanisms. The results indicated that the proposed deep learning model was superior to the existing semantic segmentation methods in both accuracy and visual effect. This research is relevant and interesting. But, the scientific contribution and specific novelty of this paper are not edited well. The methods and results are not presented adequately, especially the figures are not clear and some figures are even not necessary. Some major revisions may be needed before further processing.

DETAILED COMMENTS:

(1)   Abstract is too general, more detailed quantitative results should be added.

(2)   The contributions and objectives of this research are not very clear, please summarize them in the introduction.

(3)   Some more descriptions about U-Net structure should be added to the figure 1. At least add some text to make the figure 1 more readable.

(4)   It’s very hard for me to understand the 3D convolution process. It is recommended to remove a lot of irrelevant background contents and directly explain the detailed 3D convolution process.

(5)   What’s the meaning of figure 5? It's not necessary to put it here if a figure doesn't make sense.

(6)   More comprehensive comparison of different cloud detection approaches should be conducted, such as the efficiency, the detection effect when the cloud is large or small, etc.

(7)   Most of the figures are not clear, please revise them.

Author Response

We have modified the description of methods and results in our paper based on your suggestions. Please see the attachment for the specific modifications.

Author Response File: Author Response.pdf