Network Collaborative Pruning Method for Hyperspectral Image Classification Based on Evolutionary Multi-Task Optimization

Round 1

Reviewer 1 Report

1.  The hyperspectral image (HSI) classification model proposed in this paper seems to be an integration of several mature modules, e.g., EMTO, genetic operation, knowledge transfer, etc. Although such research is fine, the author should clearly explain the advantage of such integration on HSI classification and discuss the reason.

2. The author mentioned about the data redundancy problem in the paper. I suggest that the author could discuss the potential advantage of the proposed method in dimension reduction and solving the problem of data redundancy.

3. Hyperspectral image should be abbreviated at the first time it is mentioned in L20. And should be abbreviated in the following text in L26, 28, 90, 96, 109, 124, 128, etc. A list of abbreviations may help address this problem.

4. The author should consider citing the following paper when mention the AVIRIS data.

Green, R. O., Eastwood, M. L., Sarture, C. M., Chrien, T. G., Aronsson, M., Chippendale, B. J., Faust, J. A., Pavri, B. E., Chovit, C. J., Solis, M., Olah, M. R., Williams, O., 1998. Imaging Spectroscopy and the Airborne Visible/infrared Imaging Spectrometer (AVIRIS). Remote Sens. Environ. 65, 227–248. https://doi.org/10.1016/S0034-4257(98)00064-9.

5. The author should consider mention NASA-JPL in the acknowledgement for providing the AVIRIS data.

Author Response

Many thanks for your kind recognition and professional suggestions of our works. We sincerely appreciate your valuable comments and helpful suggestions which have helped us to significantly improve the quality of our paper. According to all the comments, we try our best to correct and improve the issues.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposed a new image classification approach for hyperspectral image based on pruning solution. The main innovation is exploring the pruning algorithm during the process. However this work is not very clear. The whole process was more like a strategy of several general processes, although it can achieve better results than recently CNN methods. The core algorithm needed to be strengthened and detailed. Therefore, in the current state, the paper does not deserve to be published. Here (just for indication) some ways to improve the current version of the paper:

 

1. The article uses CNN so that it would be sufficient to start by discussing some issues related to applying CNN in hyperspectral classification, as the language is redundant. Furthermore, the article aims to address resource consumption issues caused by complex neural networks, and relevant work by other scholars in solving these issues should be listed. There are no references or relevant work in Section 2.1 related to the problem to be solved, and it does not describe the background of the method as described in the title.

2. Five comparison methods in this article belong to the ordinary CNN method, and the model is based on Neural network pruning, and other Neural network pruning-based methods should be added to verify the effectiveness and innovation of the model. Furthermore, Only five CNN methods are briefly listed, but the accuracy table (such as Table 2 and Table 3) and the classification chart below appear in the method called 3DDL that is not introduced.

3. Several comparative methods are simple and fast-running CNN models. It is meaningless to simplify the models through pruning on this basis. Instead, more complex models should be selected for comparison.

4. The article states in the contributions that model pruning based on multi-objective optimization can enable network deployment on mobile platforms with different hardware resources. However, the entire article does not test on other devices, and it is unclear how this conclusion was reached.

5. Besides changes in accuracy, are there any changes in parameter count, classification time, training time, etc. after the network is pruned?

6. Moreover, the experimental settings are not detailed, such as learning rate, number of training epochs, etc

 

Some contents listed needs to be improved.

1. In lines 3-4: The introduction of pruning methods appears abruptly and does not connect well with the preceding text.

2. In lines 18-21:The description of this paragraph is not necessary, please introduce HSI directly.

3. In lines 128-141: The description of this paragraph is very similar to the first two paragraphs of the introduction and does not enumerate references in terms of the issues to be addressed.

4. In Lines 400-439 and 458-468: Did all experimental results only conduct one trial? The results of a single experiment are subject to chance (such as in Table 2 and Table 3, where the three accuracy indicators of DCCN and Pruned methods differ by only about 0.02% with no significant difference.

 Threre are minor editing of English language required to be improved.

Author Response

Many thanks for your professional suggestions of our works. We sincerely appreciate your valuable comments and helpful suggestions which have helped us to significantly improve the quality of our paper. According to all the comments, we try our best to correct and improve the issues.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have proposed a network collaborative pruning method for hyperspectral image classification based on evolutionary multi-task optimization. The manuscript is complete, and the authors try to prove the progressiveness of the algorithm through experiments. However, there are some problems that need to be revised. The comments are as follows

1.      First, the computational complexity of the algorithm needs to be analyzed and compared with SOTA algorithm.

2.      How about the adaptability of the algorithm to different number of training labels, especially small labels. Please compare with the SOAT methods.

3.      The references used in the paper are relatively old, so it is recommended to update them. In addition, some more methods regarding remote sensing using graph-based methods should be investigated in your introduction, e.g., Semi-Supervised Locality Preserving Dense Graph Neural Network With ARMA Filters and Context-Aware Learning, Unsupervised Self-correlated Learning Smoothy Enhanced Locality Preserving Graph Convolution Embedding Clustering, Self-supervised Locality Preserving Low-pass Graph Convolutional Embedding, Multi-scale Receptive Fields: Graph Attention Neural Network, Multi-feature Fusion: Graph Neural Network and CNN Combining, MultiReceptive Field: An Adaptive Path Aggregation Graph Neural Framework.

4.      What is the adaptability of the algorithm proposed by the authors to image noise? Please use experiments to prove the progressiveness of the algorithm. That is to say, what is the classification performance of the algorithm when images are injected with different noises. In addition, when the classes are unbalanced, what is the classification effect of the algorithm.

The algorithm proposed in the manuscript should be compared with the SOAT classifiers. Additionally, I believe the proposed classifier should be validated on larger datasets, such as Houston, Hong An.

Author Response

Many thanks for your professional suggestions of our works. We sincerely appreciate your valuable comments and helpful suggestions which have helped us to significantly improve the quality of our paper. According to all the comments, we try our best to correct and improve the issues.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

 I have reviewed your changes and I am pleased to inform you that I have accepted your manuscript for publication.Your changes have greatly improved the clarity and coherence of your manuscript, and I believe that your study will make a valuable contribution to the field.

Reviewer 3 Report

No more comments. The paper can be accepted.