Partitioned Relief-F Method for Dimensionality Reduction of Hyperspectral Images
Round 1
Reviewer 1 Report
Dear Authors,
thank you for your paper. You will find my comments as follows:
Line 11: don’t use your abbreviations in your abstract. “is performed on three test sites”
Abstract: please add accuracy measures / information about your mentioned significance to the abstract.
Line 20-21: They did not shift. Multispectral is still very important and in focus of research. Please highlight the growing importance of HIS, but do not try to explain this in combination with a decrease of multispectral research (which actually means “shift”)
Line 21: spectra instead of spectrum
Line 27: you do not need to write “such as” together with “etc.” – one of them is enough
Line 40: Don’t start a sentence with AND
Line 58: write “Within the methods…” to improve the structure of the sentence.
Line 60: wrong comma
Line 61-63: one cannot understand the meaning of band interval and sub-interval in these terms
Line 64: Please clarify which “groups” are you talking about. This is not mentioned before
Line 68 – 73: please improve readability. Too many “ands” in this paragraph, especially at the end
Line 75: “hundreds of spectral bands” is not a good definition of HIS – already contiguous 60 bands with very small steps in-between would be hyperspectral.
Line 77: literature?
Line 79 - 80: you already explained this in chapter 1
Line 84: functions instead of function?
Line 90: “the proposed”? Was this explained beforehand? Or is it “a proposed”?
Line 138: please improve sentence structure
Line 145: Please no sentence start with AND. Maybe use here Furthermore?
Line 169: do you really mean “bands of spectra”? b bands would be enough as an explanation. Or “Every spectrum consists of b measured values.”
Line 170: You do not need to repeat Hughes and the need for dimensionality reduction
Line 173: Is it really “classification model of HSIs” or is it more “model for HSIs”?
Line 175: Please correct sentences for: No AND at the beginning and not so many semicolons.
Line 180-206: Please also add (wj) for your score to the text
Line 217-218: doubled information in brackets
Line 265: why do the authors select Lambda for the threshold? It is strongly related to the wavelength in this field of research?!?
Line 300: THE training set
Line 308: Please clarify what you mean with processed
Chapter 4.1) the authors do not provide any information about the processing stage of the three datasets – are they comparable? Are they corrected for atmosphere? If yes – which algorithm was used? Following the chapters before, I assume the authors used reflectance. Did they?
Furthermore the classes used in each of the three datasets are very different. Salinas has got spectrally very similar classes and PaviaU has got less and spectrally better separable classes. How can you interpret possible differences in your results with so many different basic information? Classes are different, sensor is different, spectral resolution and number of bands totally different? This is a very tough task to compare the three data with each other.
Table 4: The table would show a lot more information, if you add the importance score to the band indices. Furthermore a graphical illustration would show the information much better than this table. It is hard to extract important information from this “wall of numbers”. Suggestion: “Create a graphic with the wavelength on the x axis and the three sensors on the y axis. Include small bars for each band of the sensors. Now colorize the bands (color transit, lookup table) according to their importance score. Here is a link to a graphic that fits a little bit to what I mean. Just change your colorization according to your score:
https://www.researchgate.net/profile/Aurelie_Shapiro/publication/324537528/figure/fig7/AS:631598743568434@1527596283448/Comparison-of-band-widths-of-spaceborne-multispectral-sensors-and-one-hyperspectral.png
This graphic would clearly show the mentioned clustering without concentrating on only 30 bands.
Table 7 and Table 8: You don’t need these tables, as they don’t bring real new information to the topic. The variations of the OA are so small, that difference would strongly be linked to the data quality, training fields and validation fields as well as the selection of semantic classes. Table 6 is enough to show the increase of your accuracy due to adjusting the threshold.
Line 390: “can achieves” – wrong sentence construction
Please explain: How many dimensions did your threshold of 0.9999 deliver for the three test sites?
Line 399: what is a “very level”?
Line 400: No AND at the beginning
Line 419: AND
Line 431: improve English
Finally:
Please improve the written English. Some sentences are unfinished. Sometimes words are missing or even there were two verbs in a sentence.
VERY IMPORTANT: Please provide more information about the processing of the three HIS datasets. It is also unclear how field information was achieved. There ist no information given about the quality of the reference data. The method described is fine, but the basic data of the experimental part is still not described enough.
Author Response
Dear Reviewer,
Thank you for your comments and suggestions concerning our manuscript entitled “Partitioned Relief-F Method for Dimensionality Reduction of Hyperspectral Images” (Manuscript ID: remotesensing-732044). Those comments and suggestions are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.
According to your comments, we have revised the manuscript carefully. Our specific responses to your comments are given in the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The article "Partitioned Relief-F Method for Dimensionality Reduction of Hyperspectral Images" presents a modification of the Relief-F dimensionality reduction method for use with Hyperspectral data. The authors first describe ORF method of calculating band importance, and then describe an algorithm for dividing the bands into partitions based on their correlation. The authors present a nice way of determining the redunancy a sub-interval in equation (17). Two assumptions, made to motivate the new method, are tested, but it is not clear what the assumptions are. It is finally shown that the PRF method works effectively on several example data sets, and the effect of the lambda parameter is tested. Finally the PRF is compared to a few other dimensionality reduction algorithms, including ORF and PCA. K-means and BIRCH are also included in the comparison, though it is not clelar how as they are implemented in the test. Overall, this is an interesting paper, but a few aspects, particularly the typos, need to be addressed before publication.
Major:
1) Because the article has so many typographical and grammatical errors, it is hard to read. See for example:
(i) the listed page numbers are wrong
(ii) pg 2 ln 66 should read "SMVs perform"
(iii) pg 3 ln 89 should have space after ")"
(iv) "prohection" is not a word. Presumably it means projection
(v) the phrase "(which can represent the degree of redundancy)" is repeated twice in in one sentence
(vi) 3.3 ln 237 don't begin sentence with "And.."
(vii) "the spatial resolution is 18 meter/pixels" should be re-written as "the spatial resolution is 18 meters/pixel"
(viii) and many more...
2) The variables in Table 6 need to be explained
3) mu_D is mentioned a few times, but never explicitly defined
4) The assumption "there are quantities of adjacent bands in HSIs" is unclear.
5) Table 6 should list how many subintervals are calculated for each value of lambda
6) The number of dimensions is changing in PRF, even if only implicitly. In other words, the number of dimensions is a function of lambda. Therefore, the PRF line in Figure 6 should vary as the number of dimensions changes.
7) It is not clear how the SVM performs classification after K-means and BIRCH because they themselves are classification algorithms.
Minor:
a) Why is arg min used in equation (10) when the variable being calculated is called the "maximum average correlation vector"?
b) The derivation of (17) seems to hold, and it works for the case of identical bands, but it is hard to understand what is happening. Maybe a image of an example "maximum average correlation vector" and the bands used to compute it could be shown in a figure.
c) Figures 3-5 would only be relevant if they were compared to the classification as performed after the PRF dimensionality reduction (and preferably the other methods as well). I recommend adding those subfigures.
d) The two assumptions verified in section 4.2 should also be stated there
e) Is there a bias from the ordering of the bands? In other words, if the algorithm started with the longest wavelength and then decreased, would it end up with a different partition that your implementation? Does it matter?
f) The lambdas calculated in Tables 6-8 are all much larger than I would have expected. Is there a reason for this?
Author Response
Dear Reviewer,
Thank you for your comments and suggestions concerning our manuscript entitled “Partitioned Relief-F Method for Dimensionality Reduction of Hyperspectral Images” (Manuscript ID: remotesensing-732044). Those comments and suggestions are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches.
According to your comments, we have revised the manuscript carefully. Our specific responses to your comments are given in the attachment.
Author Response File: Author Response.pdf
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
Dear authors,
I really enjoyed to read your paper and the results of the presented method are promising.
I only have few comments on your paper. They are mainly formal comments. But I have one single major question regarding the NWS performance. I am wondering why you didnt finally compare the NWS to other spectral and spatial feature integration methods. It feels like this part is really missing at the end of your results.
Please use your PRF results and compare one or two other methods to the NWS. This would make your good paper a great paper!
And now my comments in detail:
Line 5: "resulting in inaccurate of the classification. “ à please correct spelling.
Line 13: Can the authors already provide an accuracy measure here?
Line 27: which method do the authors mean? Can you really state this for "all" methods?
Line 28 - 31: this is not clearly understandable for people who have not worked with data with high dimensionality. Also the tense is not right in this sentence.
Line 73: numbers instead of number
Line 76: feature instead of features
Line 213 - 215: improve grammar
Line 302: The beginning of this subchapter (4.2) belongs to the methods chapter up to line 322. I also suggest to add chapter 4.1 (data) to chapter 3 and name it methods and data.
I have a question regarding the chapter Experimental Results and Discussion: The authors introduce the NWS as an option for spectral and spatial feature integration. But why don’t they compare their method to one of the other spectral and spatial feature integration methods mentioned in chapter 2.2.2, like GDF or EMP? After the comparison of PCA, ORF and PRF in chapter 4, it is clearly shown, that PRF delivers the best results. Why didn’t the authors use different spectral and spatial feature integration methods on this PRF and compared them to the performance of the NWS?
Line 287: please specify the processing level of the data from Salinas. For example: is the data corrected for atmospheric influences?
Figure 4, 5, and 6: please provide the maps with coordinate information. The reader should be able to find the study sites in for example Google earth, to get a better overview of the test site used.
Line 293: it is not clear why and how the authors eliminate samples because they do not contain information. Please give an explanation.
Line 291: please specify the processing level of the data from PaviaU. For example: is the data corrected for atmospheric influences?
Line 291: please specify the processing level of the data from KSC. For example: is the data corrected for atmospheric influences?
Table 6: You do not necessarily need to present the Kappa as it does not give real additional information compared to the OA. The result of the interpretation will always be the same.
Line 427: Did you really mean RRF?
kind regards.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
The article "Partitioned Relief-F and Nested Sliding Window Method for Spectral-Spatial Classification of Hyperspectral Images" presents two techniques to assist the classification of hyperspectral data. The first, Partitioned Relief-F uses the original Relief-F technique to compute scores for each band in the hyperspectral spectrum, and then selects bands to use for classification from sub-partitions of the spectrum, unlike the original Relief-F which selects bands from the entire spectrum at once. The second, the nested sliding window, is a way to incorporate spatial information into the analysis of hyperspectral data. The Partitioned Relief-F method produces minor improvements over the baselines, while the nested sliding window produces notable improvements. However, there are significant flaws with the paper (the two presented methods are completely distinct, the nested sliding window is not well-defined mathematically--no information about choosing the width 'a'--, the nested sliding window is not compared to any other spatial-spectral methods, the discussion of other spatial-spectral methods is lacking, etc.) The most significant issue is that the two methods are completely distinct, which seems to suggest that this would be better as two separate papers rather than one. Both new techniques are interesting on their own, but since they have to share space in the same paper, the ideas behind them are not developed fully. Some parts of the paper, such as the investigation of accuracy versus number of dimensions and the simple mathematics are quite nice.
Major concerns:
1) There is no reason for the Partitioned Relief-F and the Nested Sliding Window to be discussed in the same paper.
2) The Nested Sliding Window method is compared to no other spatial-spectral methods, despite the sentence at the end which states "compared to the use of a fixed window..."
3) The introduction is very unbalanced. There are ~30 references discussing feature selection methods[31-60], but only 9 discussing spatial-spectral processing [61-69].
4) The very simple investigation of the experimental results (just OA, AA, Kappa) does not allow for much discussion about why the two new techniques improve upon the older techniques used in the comparison.
Minor concerns:
i) ORF, which seems to mean 'Original Relief-F', is not defined.
ii) There is no information about how to choose the width of the partitions in Partitioned Relief-F
iii) The portion of the introduction which discusses feature selection does not discuss any trends or categories, but only presents a list of methods.
iv) The algorithm described in Figure 2 is sloppy and has one more for-loop than necessary.
v) There is no information about the time required to run the various methods.
vi) Since the details of the SVM classification are not mentioned in the article, it might be worth pointing to the relevant papers in the references (e.g. 24, 25, 62).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The revised version of "Partitioned Relief-F and Nested Sliding Window Method for Spectral-Spatial Classification of Hyperspectral Images" is mostly the same as the original version. The authors have made some improvements, such as including the definition of ORF and adding a few references to the introduction to spatial-spectral integration. The discussion of the methods is more complete than before, but the writing in those sections is less clear. The paper is better than before, but I still cannot recommend its publication, because the NSW method, though it produces nice results, is not clearly described in the paper.
Major concerns:
1) As in the first version, there is still no reason for the Partitioned Relief-F and Nested Sliding Window to be presented in the same paper.
2) The definition of the NSW technique is not clear, e.g. what is the difference between neighbors and sub-neighbors? Also, is equation (11) defining a set of sets, one for each n,m in [0,a] or is it defining a set of pixels for a particular n,m which could be any value in the range [0,a]? It is also not clear what is meant by size: is it area or length of the side of a window? Some of the language in the description, e.g. "The size of H_whb is amplified" is colloquial and unclear. Maybe a picture would help here? Maybe coloring the pixels in Figure three to show whether they are a neighbor or a subneighbor would help to make the concept more clear.
3) Even if the method is not fully automated, guidelines about the choice of partition for PRF presumably exist. For example, you wouldn't partition into individual bands, because then PRF would do nothing. It should be possible to give potential users of the method some guidelines. If nothing else, state a few partitions that are definitely not good.
