Remote Sensing Image Scene Classification with Noisy Label Distillation

Round 1

Reviewer 1 Report

This paper presents a scheme called Noise Label Distillation (NLD) in a teacher-student architecture, with the purpose of automatically labeling remote sensing images. 

The advantage of the proposed scheme is that it does not require a pre-training stage. According to the results, specifically the confusion matrix, the labels assigned by the network coincide with the real labels in most cases. 

 

Author Response

Dear Reviewer,

Thank you very much for taking the time to review this manuscript. We really appreciate all your comments and valuable feedback. If you have any further suggestions for changes, please let us know.

Sincerely,

Tao Lei

 

Reviewer 2 Report

This paper proposed a way to classify the remote sensing image scenes with noisy labels. This paper would be more scientific solid if the authors provide clear reasoning for designing the network structure and the loss function. 

First of all, the difference between NLD and DML is not obvious. The structures and loss functions are similar. It would be more understandable if more details of the designing strategies and analysis are provided. Moreover, the role of the decision network is not clear. It seems that it was designed for mutual training. The loss L_g is the combination of r1 and r2. The total loss is also the combination of L_g and L_h. It is not clear how L_g and the total loss works differently.
Second, I respectfully disagree that the purpose of training the network is multi-task learning in line 172. The network was only designed for the classification task using both clean and noisy data. There is no clear evidence that this network used for multipurpose.
Finally, the confusion matrices in Section 4 are hard to understand. It would be easier to understand if there is a better way to represent the results.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This work proposes a new efficient framework named NLD to address the noisy label problem for remote sensing image scene classification. NLD can distill the knowledge from different types of noise to improve the performance of artificial neural networks.

From my point of view, the paper seems very interesting and is suitable for this journal. However, I have some concerns:
- the authors should review some English expressions.
- there is an important mistake in the article: the authors do not specify the percentages of training, test, unlabeled and noised labels.
- Please, include more semisupervised methods.
- Please, include a discussion about the number of parameters and FLOPs.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Thank you for your answers. The novelty of the proposed algorithm is not obvious. The difference between the proposed algorithm and the existing algorithms is not clear yet. The paper would be more solid if you provide clear evidence for the distinct points of the proposed algorithm.

Reviewer 3 Report

I would suggest to try different training dataset size 10%, 20%, 30%, 40% and 50%.