Adaptive Weighting Feature Fusion Approach Based on Generative Adversarial Network for Hyperspectral Image Classification

Round 1

Reviewer 1 Report

The manuscript adapted weighting feature fusion approach based on generative adversarial network for hyperspectral image classification. Authors presented the issue of having a small number of samples and introduced unlabeled data to overcome this issue. I found that this paper is well written and presented. I don't have comments because I think the paper was written carefully and scientifically. I think this paper is suitable to publish in Remote Sensing.

Good Luck

Author Response

Thank you for your approval of this manuscript. 

Reviewer 2 Report

The authors established a new method of deep learning, Adaptive Weighted Feature Fusion-Generative adversarial network (AWF²-GAN), to use spatial and spectral information for hyperspectral image classification.  They tested this new method on the Indian Pines and University of Pavia datasets, and demonstrated its superior performance over traditional classification algorithms like Support Vector Machine, other GANs, and several types of feature fusion GANs with different methods of incorporating spatial information.  This paper was well written, with comprehensive literature review, comprehensible descriptions of GANs in general and AWF²-GAN in particular, and helpful figures to further illustrate the methodology.  Questions and suggestions follow.

1. Referring to L80, please clarify the phrase “HSIs actually contain much less land-cover objects.”
2. I am still unsure of where the input dimensions in Figure 5 come from.
3. Referring to L338, where is “*” in the above equations?
4. Referring to L342 and L427, what do “He_normal” and “He_kaiming” mean respectively?
5. Some additional horizontal lines in Table 2 would be useful.
6. Are all variables in Equations 21-23 defined?
7. Referring to L438, please clarify the phrase “partial spectral bands between these classes are approximate.”
8. In L444, change “start” to “state”
9. Please define and provide brief descriptions of SVM and EMAP, and the hyperparameter values selected for this study.
10. In Figure 8, what methods do c, d, e, and f demonstrate?
11. In the captions for Tables 3 and 5, please redefine OA, AA, and K, and explain what the bold-face indicates.
12. In the captions for Tables 4 and 6, please redefine OA.
13. The caption for Figure 9 is incorrect.

Well done!

Author Response

请参阅附件。

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors develop a novel GAN method that enhances the accuracy of HSI classification. Please refer to the comments below.

6: suffer in -> suffer from

15: three datasets -> I see only two (UP, IN) but is there anything else?

109: apply -> applying

151-154: explanations are a bit strange.
Eq (1) should be minimized for G to better deceive D such that it outputs true(=1)
Maximization of Eq (2) is correct, but its first term should equal Eq(1)
https://github.com/LantaoYu/SeqGAN/issues/33

417: define AVIRIS

421: define ROSIS

441: model collapse -> the model collapse

448: when number -> when the number

506: equal -> equally

526: the adversarial learning -> adversarial learning

572: to many -> too many

576: The F2-Con. model always yielded the highest accuracy
with no unlabeled samples. Why? Also, is it possible to
add unlabeled samples to other GAN methods to improve
accuracy also? Would be useful to know.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf