Improving Image Clustering through Sample Ranking and Its Application to Remote Sensing Images

Round 1

Reviewer 1 Report

In the whole, the paper respects the structure of a good paper (introduction, related work, the method, experiments and conclusions).

The paper presents a new image clustering technique that builds on and improves existing models by:

- first ranking samples within the clusters based on the confidence of the samples belonging to their current clusters and then using the ranking to formulate a weighted cross entropy 60 loss to learn to improve clustering performances.

- A method for computing the likelihood of samples belonging to their current clusters based on whether they are situated in densely populated neighbourhoods and a scheme for weighting the ranked samples.

- Extensive experimental results demonstrating that the new technique can improve the existing state-of-the-art image clustering models and can perform well on a variety of datasets of remote sensing images.

Author Response

Thank you very much for reviewing and the comments.

Reviewer 2 Report

In this paper, the authors propose a novel method by first ranking samples within each cluster based  on their confidence belonging to the current cluster and then using the ranking to formulate a  weighted cross entropy loss to train the model. For ranking samples, authors have developed a method for computing the likelihood of samples belonging to the current clusters based on whether they are  situated in densely populated neighbourhoods while for training the model, we have given a strategy for weighting the ranked samples. We present extensive experimental results which demonstrate that  the new technique can be used to improve the state-of-the-art image clustering models, achieving accuracy performance gains ranging from 2.1% to 15.9%. Performing our method on a variety of  datasets from remote sensing, we show that our method can be effectively applied to remote sensing  images. I read the article, it is well written and well presented. I have some minor concerns with the current version of paper.

1-       Problems of existing methods stated in Line 37 -46 should be supported by some references from literature.

2-       Please include the organization of paper in introduction section.

3-       In experiments, there should be separate subsection related to the details of datasets used in experiments.

4-       Please include label of y-axis in Figure 8.

5-       Please include some analysis of space complexity as well.

6-       English can be improved in some parts. For example, below sentence can be refined.

 Applying our method to various remote sensing datasets shows that our method could be 353 effectively used in the remote sensing images.--> We applied the proposed method to various remote sensing datasets, and compared the performance with SOTA methods. Through results we found that our method significantly outperform SOTA methods in most cases.

7-        Before conclusion, please state some development related problems of your approach.

8-       It would be better to list some potential application with exact names in the conclusion section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

L57-67: It's not appropriate in the introduction, so it needs to move to the concluding remark.

Table 1 shows that the performance of the k-means to TSUC is too low. Why?

Several algorithms are cited and applied, but the description of the algorithm is insufficient for the general readers in the remote sensing application fields.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf