Next Article in Journal
Comparison of Field and Virtual Vegetation Surveys Conducted Using Uncrewed Aircraft System (UAS) Imagery at Two Coastal Marsh Restoration Projects
Previous Article in Journal
High-Resolution Satellite Reveals the Methane Emissions from China’s Coal Mines
 
 
Article
Peer-Review Record

Orthogonal Capsule Network with Meta-Reinforcement Learning for Small Sample Hyperspectral Image Classification

Remote Sens. 2025, 17(2), 215; https://doi.org/10.3390/rs17020215
by Prince Yaw Owusu Amoako 1, Guo Cao 1,*, Boshan Shi 1, Di Yang 2 and Benedict Boakye Acka 3
Reviewer 1: Anonymous
Reviewer 2:
Remote Sens. 2025, 17(2), 215; https://doi.org/10.3390/rs17020215
Submission received: 31 October 2024 / Revised: 27 December 2024 / Accepted: 7 January 2025 / Published: 9 January 2025

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The list of aims in lines 151-168 appears to be extensive and contains some overlapping goals. It would be beneficial to streamline this part by simplifying redundant aims to enhance clarity and focus.

The paper would benefit from a clearer articulation of the rationale behind employing capsule networks for hyperspectral image classification (HSIC) tasks. Highlighting the specific motivations and advantages could strengthen the argument for using this approach.

In line 312, a more detailed explanation of the dimensions (10000 x 20 x 28 x 28 x 3) is needed. While it is understood that 10000 represents the sample size, 20 the feature number, and 28x28 the patch size, the meaning of the final dimension '3' should be explicitly stated for better comprehension.

For a more comprehensive evaluation, the authors are encouraged to include a comparison with a transformer-based network in the experimental section.

The visualization of the pseudo color images in Figs. 4-7 appears unconventional. 

There appears to be a typographical error in the caption of Fig. 5; "HIS" should be corrected to "HSI" to accurately represent Hyperspectral Image.

The Data Availability Statement contains an incorrect link. The provided link only offers a pseudo color image from ZY-3 02, which is not relevant to the datasets used in the paper (HS-13, ZY-1 02D, HJ-1A, OHS, and GF-5). To uphold academic integrity and facilitate reproducibility, please provide the correct data download links for the datasets mentioned in the study.

Author Response

All responses are provided in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This article introduces an innovative model called OCN-MRL that combines Orthogonal Capsule Network and Meta-reinforcement Learning for small sample Hyperspectral Image Classification (HSIC). The OCN-MRL framework uses Meta-RL for feature selection and CapsNet for classification, adapting to new HSIC tasks with limited samples and effectively selecting discriminative features from hyperspectral data. Detailed comments are as follows:

1. This paper replace common fully connected layer to orthogonal softmax layer. Why is that works on HSI datasets?

2. All of the used datasets seem to be relatively easy ones since most methods achieve more than 90% OA in your setting. More commonly used public datasets are recommended for experimental comparison, such as Indian Pines, Pavia, Houston, etc.

3. Traditonal deep learning methods are suggested to be used in perfomance comparison for better understanding how good the proposed method is.

4. Details of the used datasets should be described, such as the spectral/spatial resolution, size, diversity.

Author Response

All responses are provided in the attached file.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

In the second round, there are still some typographical errors. For example, the line 24 of the Algorithm 1, the lines 524 and 543.

Additionally, the authors should compare the differences between the proposed capsule network and Capsule Attention Convolutional Neural Network (CACNN) where a capsule attention layer is proposed, as the reviewer found that there is the similarity between the proposed network and CACNN. Please describe it in the Introduction section. 

Author Response

We are grateful for the reviewers comments to help improve our manuscript. The response to the comment are provided in the attached file.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

No comments

Author Response

We are grateful to the reviewer for his satisfaction on the responses to the comments.

Back to TopTop