Superpixel based Feature Specific Sparse Representation for Spectral-Spatial Classification of Hyperspectral Images

Round 1

Reviewer 1 Report

The authors designed an online Mahalanobis-based metric learning strategy to acquire better matching between the training samples and test samples for classification. LSC was used to derive superpixel regions, then a kernel SRC is introduced to the classification of HSI. The introduction provides sufficient background of HSI classifications and covers relevent references. the Method section presents a detailed therory about the proposed approach. In general, the manuscript is good to understand and the presentation is good.

figure 5 can be improved like resolution, line width, compactness. also improve figure 1

1, the introduction fully discussed the current bottlenecks for HSI classification like feature extraction and model developments. in line 78, it is vague to state to tackle the aforementioned problems, be more specific for which problems you try to solve.

2, missing line number between line 95-96

add a reference for this NP-hard problem is solved by greedy search algorithms such as orthogonal matching pursuit (OMP).

3, the JSRC [] ???

4, Figure 1,  what's Gabor used for ? and Why?, it would be better to provide a more detailed diagram 

5, Line190, what the percentage of training samples accounts for the PaviaU dataset? 

6, line 199,  give the Std values of the 10 times repeated calculations

7, add background information for linear spectral clustering (LSC)  and why you chose this one among other superpixel algorithms?

8, line 241 test samples

Author Response

The response is in the attached pdf. Thank you for your comments.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper proposes an approach for spectral-spatial classification of hyperspectral images (HSI) at superpixel level. The first step of the method is to divide HSI to adaptive spatial regions, both adaptive in size and shape forming superpixel. Next, for each superpixel both spatial and spectral features were generated. Next, kernel sparse representation classification is utilized for classification of HSI employing metric learning strategy to exploit the commonalities of different features.

 

Comment:

-        Method is now performed on raw HIS. Hence, the procedure for generation of adaptive superpixels is not clear (using of the LSC algorithm).

-        It is not clear how test samples and training samples are chosen in superpixel-based SRC.

-        I propose adding subsection prior to Results section summarizing the method’s concept. This would be helpful to reproduce the algorithm.

-        In addition, authors have to clearly point out paper’s contribution since the proposed approach is not the novel approach but rather hybrid technique composed of existing methods.

-        In order to have convening comparison of the proposed method to CK-SVM, JSRC, KSRC, MASR, MFASR, SPSRC and SPFS-SRC authors should provide parameters used in these competitive methods.

Minor comments:

-        In Abstract, use HIS instead hyperspectral images after defining the abbreviation. Correct the same in paper text.

-        Literature review is not categorized and it should be restructured.

-        Also, in the paper text I find many abbreviations which are not used once they are defined or authors use whole phrase even though they have defined abbreviations.

-        Correct “as the component analysis (PCA)”.

Decision:

-        The proposed method outperforms some of the recent methods. I suggest accepting paper after some clarifications and corrections.

Author Response

The response is in the attached paper, thank you for your comment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript proposes a spectral-spatial classification framework for hyperspectral images based on sparse representation and superpixel segmentation. 

The framework contains a lot of elements, some of which are not clearly described, making it difficult to understand the framework. For example, Figure 1 doesn't seem correct to me, as it seems that spectral information is extracted from the PCs. My main concern is that the experimental setup is not performed correctly when training and test data are randomly sampled from the whole reference image. Random sampling causes training and test data to be in the same superpixels, resulting in an overestimation of accuracy. A fair, meaningful, and comprehensive experimental setup is based on a spatially disjoint sampling as well as a report of standard deviations and confusion matrices in order to assess the significance of the difference between the methods. Besides that, the results are not very convincing, which makes the added value of the paper questionable.

The segmentation does not seem to offer a good basis for the framework, as some class boundaries are not captured in the image. This leads to the fact that, for example, small areas are completely incorrectly estimated. An exact statement is not possible because the authors do not provide confusion matrices, however, it is indicated by a low average accuracy. 

In addition, there are several spelling mistakes and grammatical errors, as well as missing references in the text.

Overall, I cannot recommend a publication in its current form, as the methods are not described clearly and the experimental setup is not adequate.

Author Response

The response is in the attached PDF, thank you for your review.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

the authors have improved the original manuscript based on the feedback from the reviewers.

I suggest accepting the paper.

Reviewer 2 Report

Authors have clearly answered to my questions. The paper is improved according to the suggestions.

Thus, I suggest accepting the paper.