QPWS Feature Selection and CAE Fusion of Visible/Near-Infrared Spectroscopy Data for the Identification of Salix psammophila Origin

Round 1

Reviewer 1 Report

FORESTS MDPI – Review of QPWS “Feature Selection and CAE Fusion of Visible/Near-Infrared Spectroscopy Data for the Identification of Salix psammophila Origin”

Title/Abstract/Keywords/Acronyms

Although rather long, this Section is well organized, conveying the problem to be solved, while also introducing (1) the algorithm to be used, and (2) the optimal means of intertwining the selected algorithms for accomplishing effective identification of Salix psammophila. Each keyword and acronym should be explained in depth at its very first appearance in the main text.

 Introduction

A critical exposition of both parameter extraction and parameter selection should be added at this point, highlighting the pros and cons of each “candidate” methodology, thus also affording a comparison of the two methodologies used in the present context, and justifying fully the authors’ choice.

The text in lines 80-90 should be presented in a critical – and quantitative - manner that justifies the potential of applying the proposed methodology. This is a most important point, as are quantitative results and numerical comparisons derived from cross-validation, all of which help to foolproof the choices made by the authors, as these must be discussed (also) in more depth in the following text.

Theory and Implementation

This Section is in need of clarifications at a number of levels. Initially, straightforward cross validation (CV)  in the form of leave-one-out, as well as 2-, 5-, and 10-fold should be implemented, with the results (I)  reported and compared  in a critical manner, in order to (II) provide  a quantitative in-depth exposition of comparisons between the four schemes, thus (III) providing critical information on the specific dataset.

Final Comments 

Quantitative information (e.g. how straightforward it would be to apply the proposed methodology to other – differing in their characteristics – locations of interest) would be instrumental for evaluating the proposed methodology. Furthermore, the choice of methodologies, as well as all the results, should be fully supported in more adequate critical detail.

Decision

Minor Revision

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this study, machine learning-based methods were used to identify the origins of Salix psammophila.  For the study to contribute better to the literature, the reasons for choosing machine learning approaches in different processing steps should be explained in detail.  For example, it appears that the data set used in the study is labeled.  However, in some intermediate processes, unsupervised approaches such as PCA are considered.  Since the data set is labeled, why have supervised approaches such as Linear Discriminant Analysis (LDA) not been used in the study?  Moreover, why was the Bayesian optimization method preferred to optimize the hyperparameters in the 1D-CNN model?  Here, can better optimization be achieved with genetic algorithm [1]?  Additionally, the 1D-CNN model consists of simple modules.  In this context, couldn't the proposed 1D-CNN model be further improved?  For example, couldn't the attention module be used in the model?  On the other hand, has any hyperparameter optimization been made in the proposed CAE model?  Why was the Bayesian optimization approach not used in the optimization of the CAE model?  Why was the dimensionality reduction process needed twice?  Couldn't this process be done in a single step?

[1] Alibrahim, Hussain, and Simone A. Ludwig.  "Hyperparameter optimization: Comparing genetic algorithm against grid search and Bayesian optimization."  2021 IEEE Congress on Evolutionary Computation (CEC).  IEEE, 2021.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear authors, tahnk you for your contribution.

Please, define all abreviations with a first use in the text (1D-CNN;, PCA are for exaple in the abstract without defining).

next CARS, ASD etc.

First sentence in your abstract - Salix psammophila, as a leading pioneer tree species... define, what is it (not all readers know it) for example as in your introduction: it is a deciduous, bushy, and upright shrub belonging to the willow)

Introduction

China faces a serious desertification issue...
start more generally, this is not just a China problem; there should be a few lines about desertification in the world with 1-2 citations, and then start with China.

Ad a figure, which shows China´s deserted parts, ssale bar an orintations (or coordinates) needed

Fig. 3 should be bigger

Fig.12-13 why have these images dark (black) background?

Conclusion

These findings demonstrate that the approach employed in this study has signifi antly improved the classification of Salix psammophila origins based on Vis-NIR, providing valuable insights and methods for addressing similar problems.

OK, but I have not found a comparison with the possibility of this certainly better identification on some remote sensing data, i.e. from airborne hyperspectral or satellite data. If you have found a way to find the types of that woody plant using spectra, that's good, but how do you apply it to practice? Do you have any hyperspectral scanner data to make it meaningful for a search in a country like China? That needs to be explained.

Add a part, which will be focused on practical using with (necessary) hyperspectral data! You work with spectral curves from field measurements, but if you will use it in practice, you need a data from alarge areas (remotely sensed data).

See for example https://doi.org/10.3390/rs15123130, these data can be used for your using in practice, or find next references like using hyperspectral data in forestry (https://www.zora.uzh.ch/id/eprint/234294/1/Thesis_Zehnder.pdf)

References

2/3 of all citations are from China, it is too much for a scientific paper with with global potential;  but that's mainly the editor's problem, the citation has to be worldwide, otherwise the journal gets into trouble.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Dear authors, thank you for improved text. I havent more question, all was solved. Only: - remote sensing and spectroscopy in forest is long time used, there should be more references worlwide. Your reference list is not entirely appropriate and complete.

Question:

1D-CNN QPWS versus PLS DA QPWS is 3 times better; why and it is really so? (tab.6)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf