Space Target Material Identification Based on Graph Convolutional Neural Network

Round 1

Reviewer 1 Report

The paper presents a DL solution for the identification of space target materials from hyperspectral images. The proposed solution performs a segmentation of the hyper spectral image, and creates a materials identification dataset. It then combines GCN and 3D CNN in one model that learns from global and local spatial features and spectral features to identify the material of each pixel. Simulated data and lab measurement data are used to evaluate the model and compare with existing solutions. 

The problem tackled is clearly defined, the method is described to some good level of detail, the results are promising, the discussions and conclusions reinforce the motivation and accomplishment of the work.

The English language needs more attention. At times, sentence formulation get obscure and makes difficult the conveying of the message, especially in the Methodology section.

Related work is squeezed in the Introduction section, very quickly passing over what is missing in current research and what makes the contribution of this work unique. Related work section is instead background information to this work.  

Detailed comments

Page 6: "the connection relationships between two components" - what are connection relationships ?

Pg 6: "Nf is the length of node features" - what is length of node feature?

Pg 10: "The nodes of Sl and S0,i are connected if their spatial distance in the image are close" - what is close distance? 

Pg 26: The introduction of metrics for inter-, intra-class and spectral separability J could be better placed at the end of Methodology, e.g. an evaluation metrics subsection. The related table and graphs with results fits very well in the Results section. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposes an identification method for space target material under complex illumination conditions, which is based on an improved graph convolutional neural network identification algorithm. The following are specific comments.

 

1.The English language needs polishing, and there are obvious mistakes in the manuscript.

1) P28 L20 “are chose as” à”are chosen as”

2) P28 L22 “overall performances” à”overall performance”

 

INTRODUCTION

2. First paragraph: References are required.

3. Why graph convolution neural network is introduced is not explained clearly.

4. The second contribution is a common method, which cannot be the contribution of this paper.

5. Introduction: The Deep Learning Methods have not been well summarized.

 

RELATED WORK

6. A. Complex Illumination Conditions of the space target: This part is too fragmented.

7. RELATED WORK should be an introduction to this field, not a description of the methods used in the paper.

8. Related Work(A): Why is the spectral data distribution most sensitive to sunlight direction?

 

METHODOLOGY

9. In my opinion, a sentence should not become a paragraph.

10. Figures 5 and 6 are repeated to some extent.

11.P17 L18 The different identification conditions adopted for data T0, T1, and T2 need to be explained in detail.

12. The superpixel segmentation algorithms mentioned in Related Work(B) include ERS and SLIC. What motivates to choose SLIC finally?

13. P19 Table II, P22 Table IV: The proposed methods are-A, -M and -C, but the table is expressed as -A,-B,-C.

 

RESULTS

14. The % in the precision evaluation table should not be repeated.

15. How to set the parameters of the proposed network and comparison method, and how to ensure that these parameters are optimal.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors have proposed a graph capsule network for graph classification. 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. The motivations or remaining challenges are not so clear or what kinds of issues or difficulties are this task that is facing. Please give more details and discussion about the key problems solved in this paper, which is largely different from existing works.

2. A deep literature review should be given, particularly advanced and SOTA deep learning or AI models in hyperspectral image classification. Therefore, the reviewer suggests discussing some currently SOTA works by analyzing the following papers in the revised manuscript, 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, Deep hybrid: multi-graph neural network collaboration for hyperspectral image classification, Multi-feature Fusion: Graph Neural Network and CNN Combining.

3. How about the computational complexity?

4. The compared methods are not sufficient. Some SOTA compared methods should be involved.

5. It is well-known that the hyperspectral image usually tend to suffer from various degradation, noise effects, or variabilities in the process of imaging. Please give the discussion and analysis by referring to the paper titled by e.g., MultiReceptive Field: An Adaptive Path Aggregation Graph Neural Framework. The reviewer is wondering what will happen if the proposed method meets the various variabilities.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Thank the authors for considering all my suggestions. I think this paper is already in a state where it can be published.

Reviewer 3 Report

No more comments.