A Spatial–Spectral Combination Method for Hyperspectral Band Selection
Round 1
Reviewer 1 Report
This study presents a spatial-spectral combination method for the hyperspectral band selection. In the first part, all the bands are divided into the required number of subsets through subspace division. Then, in each subset, the spatial and spectral combination features were integrated into a uniformed objective function by constructing weight coefficient.
Generally, the article is well written but there is no real problem contextualization. I would include in the introduction a short note on the applications of hyperspectral analysis in the various fields mentioned, to better contextualize the study.
In addition, the aim is confused: it is more of an activities list.
I will consider the work for a publication only if a thorough review will be done.
Minor comments
Not all acronyms are spliced. Report either an initial table with all the explanations acronyms or always specify them before writing.
The abstract is not very exhaustive. Improve it.
Do not use pronouns in the text (e.g., we).
Generally, figure’s captions are not very detailed and explanatory.
Major comments
Introduction
Include in the introduction a short note on a short note on the applications of hyperspectral analysis in the various fields mentioned (e.g., agriculture, geological exploration, environmental monitoring, etc.), to better contextualize the study.
L.41-47: Introduce references.
There is not a real aim but a list of activities that have been carried out. Insert it to make the activity more understandable.
Experiment
Enter the GPS coordinates of all the study areas.
L.418: insert version of MATLAB.
Results
Insert tables to better understand the obtained results.
Conclusions
This section is not well organized.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Considering the data sets presented for assessing the methodology, where two of them are combined by homogenous agricultural plots and the third is represented by small objects, the method presented cannot be regarded as of general value. It is an idea tested on almost trivial examples. Furthermore there is no real use of training and testing subsets.
The authors overload the article with simplified presentation of several alternative methods, where I am not sure their implementation and testing represent fairly their potential performance. Furthermore, the use of SVM and ANN is limited by certain selection of parameters.
The comparison of the classification results is based on very crude parameters: in real applications there would be deeper analysis carried out looking at differences between the confusion metrices.
The method presented does not compare to state of the art methods for reducing hyperspectral dimensionality.
Good selection of training and testing sites and implementation of SVM method with all bands may perform as good as the new method where differences in running times over small areas selected for this study would not be critical.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
This manuscript presents an original idea of dimensionality reduction for hyperspectral images but suffers from some inaccuracies and abuses of language that should be corrected.
1) An important question that is almost ignored in the current version of the manuscript is the selection of the value of M?
In practice, given a new dataset not seen yet, how to determine this value appropriately? What are the recommendations?
2) In section 2.1.2, point 3, are there systematically M-1 local minima in the sequence of correlation coefficients or M-1 local maxima in the sequence of Euclidean distances between adjacent bands, even after smoothing? This point needs to be further discussed and clarified
3) Figure 3 needs to be better presented and detailed. It should really serve as a support for the fine partitioning strategy algorithm described just above.
4) In the subsequent description of the fine partitioning strategy algorithm, point 6 should also be clarified and the proposed explanation completed with the explicit updating of the labels performed
5) A caption should be added to Figure 4, explaining the reasons for and interpretation of the colours used.
6) In (9), is theta(P,Q) of the same order of magnitude as d_{pq}? Besides, is the weighting ratio theta(P,Q)/d_{pq} homogeneous in terms of the underlying information carried by these two quantities? This is taken into account later to compare the spatial-spectral combination function and the information entropy in (10) and (11) but is not done here. Why?
7) K-Nearest Neighbor is part of the proposed Spatial-Spectral Combination Algorithm. It is also used for the evaluation of the proposed method with respect to the classification accuracy achievable after dimensionality reduction. Is there not a bias introduced in the analysis of the results that makes the comparison non-objective due to this? Please justify and convince with your answer.
8) I think it is necessary to consider another classification method in addition to SVM unless you can seriously justify that the criticism made in the previous question does not biaise the results
9) The standard deviation in Tables 1 and 2 (as a result of averaging five calculations to compensate for the instability of the final classification result) should be specified/commented.
10) There is a recurrent confusion between the qualifier "optimal" used several times by abuse of language and what in my opinion results from a better result obtained following one or several numerical experiments according to a somehow limited numerical protocol and a limited number of methods involved in the comparison).
11) The space partitioning method suggested in this paper, the SSCBS method, tends to be the best among the methods involved in the comparison, but it can not be claimed it is optimal, based on the current content of the manuscript.
12) For me, the following sentence makes no sense and contributes nothing "In particular, the advantages of the proposed algorithm are more evident when optimal bands are selected." if it does not contribute explicitely to the explainability of the proposed method.
13) Further comment (explanation) is needed on the ability of the method to produce stable results regardless of the scaling ratio used (namely 0.005), as shown in Figure 12.
Typo(s):
cluster.NC
(SSMRPE)method
we proposes
proposed, Through
comparison, The Means
.The obtained band subset having more discriminative bands simultaneously.
Salians
(f), The SSCBS method
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The manuscript in the present form can be published in Remote Sensing as the authors have been improved it. Parts that better contextualize the work have been added.
I don’t need to review another version because I accept the work in the present form.
Reviewer 3 Report
The authors have responded appropriately to the requests, comments and suggestions made in the previous round.
The flaws have been substantially corrected and the content of the manuscript is now in an acceptable form
