Spatial Feature Enhancement and Attention-Guided Bidirectional Sequential Spectral Feature Extraction for Hyperspectral Image Classification
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript proposes a spatial feature enhancement and spectral attention method for HSI classification. Overall, it is an interesting idea in terms of spatial feature extraction. However, the algorithm is not fully explained, hindering the results' reproducibility. Also, the manuscript has many unsolved problems. Given this, I would advise a major reformulation. Many issues need to be solved:
1. The introduction section needs to be improved, the question to be solved is not sufficiently described and the main contributions are unclear.
2. There is a lack of sufficient literature research in the related work section. For example, the attention mechanism has been widely used in various hyperspectral classification methods, but there is a lack of literature survey on the attention mechanism in the hyperspectral field. In addition, the number of cited papers is insufficient, and there are only 7 papers published in the last three years in the cited papers.
3. To the best of my knowledge, spectral attention has been introduced in hyperspectral image classification for many years. Besides, a large number of attention mechanisms have been proposed in recent years. Thus, I think that the novelty of the proposed method is limited.
4. In Section 4.5, I do not agree that the use of 1% of the samples for training can reflect the actual situation of insufficient labeled samples. For a clearer definition and solution of the insufficient samples problem, please refer to the survey and SOTA methods in the field of hyperspectral image few-shot classification. In such methods, there are only 5 labeled samples of each category used for training. In addition, it is not uncommon for hyperspectral classification methods to use 1% or less of the samples for training.
5. There is a discrepancy between Equation 13 and the description in lines 327-328.
6. The description of the experimental setup could be improved. It is suggested to add the information including batch size and running platform in Section 4.1.
7. In the spatial branch, you use n1 and n2 to indicate the spatial dimensions of the input data. Since you use a 9×9 convolution kernel, as well as provide enhancement of the entire hyperspectral image, I assume that your input is the entire hyperspectral image. However, in the spectral branch, you mention in line 325 that the input is imge patch, but the manuscript lacks a detailed description of the input patch. It also conflicts with the description in the spatial branch. My question is, is the input to both branches the image patch? What is the size of the patch? In addition, a larger patch size would provide more spatial information. Are the image patch sizes of all the methods used in the comparison experiments the same? Please describe your input data in more detail, and maintain consistency of related description in the manuscript
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis manuscript focuses on hyperspectral image classification by developing spatial and spectral branches, with the spatial branch comprising guided filtering, image enhancement, and MMFE module. To enhance the quality of the manuscript, please consider the following comments:
1. The introduction lacks a comprehensive review of current classifiers in the rapidly evolving hyperspectral classification task.
2. The manuscript claims that hyperspectral images exhibit a scattered feature distribution, motivating the design of the spatial branch. However, no evidence supports this claim within the manuscript. It is recommended to add details to substantiate the assertion.
3. Clarify the contribution of the proposed IFEE method. Please elucidate the novelty and specific contributions of the manuscript.
4. The method section, particularly Section 3.1, is challenging to read. Simplify this section and ensure that each component is logically explained.
5. Ensure that each symbol is introduced at its first use. For example, clarify that "E" in Eq. 7 is the same as "E" in Eq. 10. Define symbols such as A, E, and D used in the hyperspectral denoising task.
6. Include corresponding ground-truth in the classification maps in Figs. 7, 8, and 9 to provide context and evaluation benchmarks.
7. Explain why the MMFE is not introduced in the method section and verified in the ablation experiments for thorough validation.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authorssee attachment.
Comments for author File: Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsNo further comments
