Advances in Hyperspectral Image Classification Methods with Small Samples: A Review

Round 1

Reviewer 1 Report

The paper discusses the accuracy of hyperspectral classification with samples of different sizes performed with different methods, aiming at checking in particular the behaviour with small samples. An extensive literature review is performed, which is significant. An experiment is performed with well-known datasets, with samples of different sizes to assess the accuracies. The presentation and discussion of the results can be improved. I suggest to make graphs showing the variation in the accuracy with different sample sizes. The discussion of the results should be extended, commenting the changes. The acronyms are a bit abused, it is difficult sometimes to follow the text. There are typos and incomplete sentences, the text should be revised.

The English language is clear. There typos and incomplete sentences that should be fixed.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript gives a review on HIS image classification with small samples. A taxonomy is proposed. But the importance of the proposed taxonomy is not clear. Whether is it necessary or essential? And is it useful for solve the small samples problem? The analysis and discussion are also too weak. Some details comments are listed as follow.

1. Lack of elaboration of the problems faced by hyperspectral image classification methods with small samples;

2. Lack of summary of non-deep learning methods in both IS and ESE-KT categories;

3. Taxonomy section should review more articles, the explanation of the confusing concepts of few shot learning and small sample learning is unclear.;

4. The interpretation of intra-domain and extra-domain is not clear in Fig.1;

5. There is an ambiguity in the description of Fig.2, intuitively the top image is ‘differentiability’ and the bottom image is ‘connection’;

6. In ISE-PG, there is a lack of article citations to support the idea of ‘leveraging certain knowledge to select samples and generate pseudo-labels’;

7. Meta-testing, mentioned in Fig.4, lacks conceptual explanation and does not differ from Meta-Learning in terms of the process of generating the model;

8. There is ambiguity in taxonomy. The article classifies the small sample hyperspectral classification into ‘Method based on intra-domain sample set’, ‘Method based on intra-domain sample set expansion and pseudo-label generation’, and ‘Method based on extra-domain sample set expansion and knowledge transfer’. However, the methods of using pseudo-labels and not using pseudo-labels in extra-domain classification are mixed up, and the article does not sort out the classification priorities of the two concepts of cross-domain and using pseudo-labels;

9. The ISE-PG experimental summary selected methods from 2016 and 2019. A collation of recent experiments with new methods is lacking.

none

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

“In the past fifteen years,…………………….……….are from the last three years” This statement also needs the specification for the scope of the search. What were the criteria for choosing the journals etc?

The paper presents a set of fresh taxonomy challenging the earlier and organising the entire literature survey on this basis. The proposal looks interesting and needs a very strong justification for the same.

Figure 2 tries to represent points being grouped based on the focus on connections, and the lower one shows the connections. The diagrams are shaped to reassure the proposal. The shapes used need stronger justification with an explanation of the points being incorrectly placed.

The language needs minor English corrections

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Hyperspectral image classification is one of the hotspots in remote sensing. However Without a sufficient amount of labeled data, even the best model may difficult to present full performance. In such situation, many researchers turn their attention to the study of hyperspectral imaging classification with small samples.  In the manuscript “Advances in hyperspectral image classification methods with small samples: A review” authors present a new overview of the research progress of hyperspectral image classification with small samples.

General comments.

It is not clear from the article which methods are currently the most promising for classifying hyperspectral images.

The conclusions are too short and looks like an abstract. There are no specifics at all.

Specific comments.

Line 16. Keywords should not contain terms from the title of the article.

Line 65. “In the past studies, researchers have proposed various models with different ideas.” This statement is superfluous since it does not introduce any specifics. It can be applied to anything.

Lines 63-64. It is not clear on what basis this statement is made. Reference needed.

Line 67. “… some feature extraction and feature selection methods were born as a result.” It is necessary to explicitly specify which methods were born.

Lines 67-69. It is necessary to explicitly state which methods were used and how they are better than those discussed in the previous sentence.

Line 115. Without examples, it is difficult to understand what the three types of methods are. For example, where will CNN, RF or SVM end up?

Line 361. Table 1. Which of these is a method "Based on intra-domain sample set (IS)" or "Support vector machine (SVM)"?

Line 327, 329. All transcripts of abbreviations must be given. DFSL, KNN, SVM

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 4 Report

Thank you very much for answering all my questions.