Multianalyzer Spectroscopic Data Fusion for Soil Characterization

Round 1

Reviewer 1 Report

The manuscript “Multisensor Spectroscopic Data Fusion for Soil Characterization” by Richard R. Hark et al discusses the application of multimodal approach (combining Raman spectroscopy, X-ray fluorescence spectroscopy and laser-induced breakdown spectroscopy) for identification and classification of soil samples. Authors found that when used alone, LIBS technique provides the highest accuracy for soil classification. Nevertheless, the combined usage of these three techniques provides slight improvement in the classification accuracy.

I believe the revision of the paper is needed, however, to make it clearer.

1) Considering that in the focus of the manuscript is the development of methodology for classification based on inputs from multiple sensors, the actual algorithm employed in this study is presented insufficiently clearly, in  my opinion. Despite the significant length of the paper, all the procedure is described in six ambiguous sentences:

 “In order to combine data from the multiple sensors, a hierarchical classification approach was used (Figure 1). For each sample, a set of Ns x Ds spectra were collected for each sensor where Ns refers to the number of spectra for sensor s and Ds refers to the dimensions of a spectrum collected by sensors. The sensor-specific classifiers were applied to the spectra generating Ns x C classifier confidences that the sample’s spectra belong to each of C classes. In order to generate a feature vector for the sensor fusion classifier, the classifier confidences for each sensor were averaged to produce a single 1 x C vector. These vectors were then concatenated to create a single feature vector for the sample. This feature vector was classified by the sensor fusion classifier and a class label assigned to the sample based on the class that generated the highest fusion-classifier confidence.”

From this description it is not clear to me, first, why each sensor for each sample produces Ns spectra – what is the difference between these Ns spectra? Next, from this description, it seems, that the information from different sensors is processed separately, i.e., is used independently to classify the sample, and then the results of this classification are somehow merged together. Authors call the merged result a “feature vector”, which, I think, is misleading, since this vector, as far as I understand, does not contain features extracted from the spectra, but rather already the result of spectra classification. Figure 1 is also not informative and not helpful in resolving this confusion. I suggest authors provide more detailed explanation of their approach, perhaps, including concrete values for Ns, C, Ds parameters used, and the dimensions for the final “feature vector”.

2) Similarly, in Section 4.1 authors discuss the results of PCA analysis, but, again, it is not clear, what are the actual input vectors used for PCA.

3) Also, Figure 5 and Figure 6 are missing in the manuscript, at least in the PDF version available to me.

4) I also recommend to add in the introductory paragraphs a short, more specific description of the strengths and limitations of each of the employed experimental techniques (Raman spectroscopy, X-ray fluorescence spectroscopy and laser-induced breakdown spectroscopy). What unique insight each of them can provide for sample classification?

 5) Lines 80-81: authors wrote that among the advantages of multimodal approach is “increased dimensionality of the measurement”. It is not obvious to me, why the increased dimensionality by itself is an advantage.

Author Response

The text of the paper has been revised in response to this review as follows: 

1 & 2. The text of Section 3.4 (lines 305-329) has been revised for clarification in terms of the first two review comments and Figure 1 has been revised for clarity. Also Line 359 of Section 4.1 was modified in response to the second comment.

2. It was stated that Figures 5 & 6 were missing from the pdf version of the paper received for review, which we do not understand. These figures are in the original MS Word document submitted and also appear in the document when converted to a pdf, so it appears that the pdf prepared and sent out from the editorial office did not contain these figures. The two figures are present in the revision.

4. An short, introductory paragraph with a more specific description of the strengths and limitations of each of the employed experimental techniques was requested. This has been added (lines 100-108).

5. The use of complementary analyzers, which as described in lines 109-129, provides additional unique information and therefore added dimensionality for the chemometric analysis.

Reviewer 2 Report

In the manuscript Figure 5 and Figure 6 are missing, so that the content of the paper cannot be sufficiently judged at the present stage.

In my opinion some aspects are not sufficiently clearly described. Overall the manuscript needs to be refurbished, and in the present stage it cannot be accepted for a reviewing process.

Comments for author File: Comments.pdf

Author Response

1. The text of the paper has been revised throughout for clarification.

2. It was stated that Figures 5 & 6 were missing from the pdf version of the paper received for review, which we do not understand. These figures are in the original MS Word document submitted and also appear in the document when converted to a pdf, so it appears that the pdf prepared and sent out from the editorial office did not contain these figures. The two figures are present in the revision.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

During the revision authors have addressed my most critical comments, and the manuscript got much clearer, in my opinion. It can be published now. However, I encourage authors to rethink the importance of their section about PCA. As is clear now from the revised version of the manuscript, the input for PCA are the vectors that are just concatenated classification results from three different sensors. The latter, I imagine, should be already quite sparse vectors (if it is not the case, then the corresponding sensor has low sensitivity for discriminating samples belonging to different classes). Thus the input for PCA seems to be already quite low-dimensional, and it is not immediately obvious, what additional dimensionality reduction can contribute for simplifying the data representation. Perhaps, the role of PCA in the context of the present study can be better explained. Alternatively, this section can be removed at all, which would help also to reduce the total length of the paper.

Author Response

We have addressed the reviewer's comment about the PCA discussion by modifying the text for clarification. 

Reviewer 2 Report

The manuscript describes an interesting application of complementary spectroscopic analyzers and their data fusion for trustworthy soil characterization.

Some improvements on the content of the manuscript have to be done, as suggested in the attached pdf-file of the manuscript version 2.

Note that reference 35 has a wrong DOI number. 

In my opinion quite important is the link given in Table 1 for having access to supplementary information (https://osf.io/z3wgk/) in order to judge the extent and soundness of the results obtained, since in the manuscript only a small part is displayed as Figures.

Comments for author File: Comments.pdf

Author Response

We thank the reviewer for the detailed and thorough reading of the text and the suggestions made for its improvement. These suggestions have been implemented.