Rapeseed Variety Recognition Based on Hyperspectral Feature Fusion
Round 1
Reviewer 1 Report
Dear ref. N. 1888030
Title: Rapeseed Variety Recognition Based on Hyperspectral Feature
Fusion
Abstract: Change the symbols of the features to a respective name or something that indicates what they are. The reader has no idea what is “MF-h(0), IMF-Si/I1, and TRIP-Drmin” at this point.
Introduction:
Expand paragraph 1 a bit, explain a bit more some ideas here and add references.
Traditional methods of rapeseed identification are well defined and also spectral methods.
This section is ok and only require minor changes in the first paragrahp for me.
2. Materials and Methods
2.1. Materials
I am not familiar with these varieties, do they have a scientific name os something that make them easier to identify? A figure of the test site (with geographical location) could be interesting to the reader as well.
2.2. Data collection
Section is ok, if possible, could you add a picture (it can be figure 1b, for example) of the laboratory setup used? This is very interesting to see and just a few papers of spectrometry show them.
2.3. Methods
The features shown on sections 2.1-2.3 are well described and with formulas presented (as well references).
Section 3 – Results and discussion
I was expecting a little bit more on the methodology (section 2.3) on how the rapeseed identification would be done. Which is basically this section. However, it cannot de fully transferred to the methodology since it is a mix of the methodology with the results (e.g., mention that you used the Io index is methodology, but calculating it fot the 11 varieties is a result.
In this context, I would suggest adding a section after 2.3. (maybe 2.4 or 2.3.4) showing a flowchart of your methodology from data collection to the final results assessments. You can briefly comment the topics that were not covered in the methodology with the references (e.g., Io, classifiers used – SVM, validation steps – K-fold cross-v., etc..). This will let the results part to be easier to understand, since the reader has an overall idea of what to expect in the section.
The discussion of your results is fine, the confusion matrix, what features were better, model validation and so on. However, there is no discussion to other literature, and this is a major flaw. You need to add a couple of paragraphs here comparing and contrasting your results with other papers from the same area. I imagine that rapeseed identification can be somewhat hard to find references, but what for other oilseeds, maybe crops or weeds. Discuss with other papers using hyperspectral sensors for same use and different methodologies, for example.
I expect to see this on this section, and you can also consider splitting it into sections, one for results only and other for discussion. This will also affect a bit of your conclusion, since more scientific information can be summarized there.
The paper is good and needs to be improved in the section of discussion. This is a major flaw, but since it is easily corrected, I will address minor changes
Best regards,
Author Response
Reply to Reviewer’s Comments
We are very grateful to the anonymous reviewers for their valuable comments and suggestions. Based on their comments and suggestions, we have revised our manuscript. In the following we list comments (labeled as C#) from reviewers and give our responses (labeled as A#) right after each comment. We also highlight all the revision in red in our revised version.
Reply to Reviewer #1
C1. Abstract: Change the symbols of the features to a respective name or something that indicates what they are. The reader has no idea what is “MF-h(0), IMF-Si/I1, and TRIP-Drmin” at this point.
A1. Thanks for this constructive suggestion. I have revised the symbols as text description in
abstract.
C2. Introduction: Expand paragraph 1 a bit, explain a bit more some ideas here and add references. Traditional methods of rapeseed identification are well defined and also spectral methods. This section is ok and only require minor changes in the first paragrahp for me.
A2. We have expanded the content of the first paragraph and added some references in the revised version.
C3. Materials: I am not familiar with these varieties, do they have a scientific name os something that make them easier to identify? A figure of the test site (with geographical location) could be interesting to the reader as well.
A3. A good suggestion. We added pictures of the experimental area and environment in Figure 1(a) and (b) of revision. In addition, the rapeseed varieties are named in Chinese. Hence, we use the corresponding Chinese Pinyin for description of the varieties.
C4. Data collection:Section is ok, if possible, could you add a picture (it can be figure 1b, for example) of the laboratory setup used? This is very interesting to see and just a few papers of spectrometry show them.
A4. As shown in the Figure 1(c).
C5. Methods:The features shown on sections 2.1-2.3 are well described and with formulas presented (as well references).Results and discussion:I was expecting a little bit more on the methodology (section 2.3) on how the rapeseed identification would be done. Which is basically this section. However, it cannot de fully transferred to the methodology since it is a mix of the methodology with the results (e.g., mention that you used the Io index is methodology, but calculating it fot the 11 varieties is a result. In this context, I would suggest adding a section after 2.3. (maybe 2.4 or 2.3.4) showing a flowchart of your methodology from data collection to the final results assessments. You can briefly comment the topics that were not covered in the methodology with the references (e.g., Io, classifiers used – SVM, validation steps – K-fold cross-v., etc..). This will let the results part to be easier to understand, since the reader has an overall idea of what to expect in the section.
A5. It is a very good suggestion. We have added a flow chart in subsection 2.4 (highlight in red) of our revised version.
C6. The discussion of your results is fine, the confusion matrix, what features were better, model validation and so on. However, there is no discussion to other literature, and this is a major flaw. You need to add a couple of paragraphs here comparing and contrasting your results with other papers from the same area. I imagine that rapeseed identification can be somewhat hard to find references, but what for other oilseeds, maybe crops or weeds. Discuss with other papers using hyperspectral sensors for same use and different methodologies, for example.I expect to see this on this section, and you can also consider splitting it into sections, one for results only and other for discussion. This will also affect a bit of your conclusion, since more scientific information can be summarized there.
A6. We thank the reviewer for bringing this notion to our consideration. We add some sentences to highlight the significance of this work in Conclusion and discussion. Please see lines 337-342.
Reviewer 2 Report
see attached file
Comments for author File: 
Comments.pdf
Author Response
Reply to Reviewer’s Comments
We are very grateful to the anonymous reviewers for their valuable comments and suggestions. Based on their comments and suggestions, we have revised our manuscript. In the following we list comments (labeled as C#) from reviewers and give our responses (labeled as A#) right after each comment. We also highlight all the revision in red in our revised version.
Reply to Reviewer #2
General comments:
C1. Fusion is mentioned in the title and in the text. However, the fusion procedure is not explicitely described. After selection of 3 best features, how are these features combined to be used as input into classification algorithm?
A1. We thank the reviewer’s question. The three best parameters are selected by I0-index. According to the definition of I0, they are the characteristics that have the best performance of distinguishing rape varieties among their respective characteristics, while they represent different types of characteristics. Since that the intelligent models are employed in our work, the three best characteristics didn’t need to be standardized. Therefore, they are used as a feature combination (3-dimensional feature vector) and input into the intelligent models directly to be expected to acquire the optimal result for purpose of distinguishing rape varieties. To make readers understand more clearly that how our model works, we have added a flow chart in subsection 2.4 (highlight in red) of our revised version.
C2. Measurements are made in the spectral domain 400 to 1000 nm (visible and near infrared). Why other wavelenths are not taken into consideration? Most sudies about seed recognition are using near-infrared spectroscopy: soybean seeds (Tan et al ., 2014 : 1000-2500nm) ; oat seeds (Wu et al., 2019 : 874-1734nm) ; barley seeds (Singh et al., 2021 : 900-1700 nm) ; rice seeds (Jin et al., 2022 : 1000-1650 nm)
A2. We very thank the reviewer for bringing this notion to our consideration. At present, in the fields of crop seed classification, lots of interesting research was conducted based on the visible/ near-infrared (NIR) spectroscopy (700~2500nm), as the reviewer said. However, some scholars pointed out that long wavelength near-infrared (LW-NIR, 1100~2500nm) spectroscopy is vulnerable to be interfered by external factors such as moisture, which generates noise in the spectrum and thus jams model performance. In addition, we applied the SOC710 portable hyperspectral imager to collect the hyperspectral data. The acquisition spectrum wavelength range of equipment is 380~1091nm with resolution 4.9458 nm. Consider the factors of noise, therefore, we use the hyperspectral reflectance with wavelength range 400~1000nm for rapeseed classification in this work.
C3. There is a need to explain the advantage of carrying out hyperspectral measurements versus the use of a more simple and affordable method. For example, in the study by Qadri et al. (2021)computer vision was used to develop an automated seed classification system to discriminate eight canola varieties. The input data were digital camera acquired images using a ‘NIKON’ COOLPIX model with 10.1-megapixel resolution. Results indicate an accuracy almost varying from 95% to 98% for the eight canola varieties.
A3. This is very important point. Indeed, the digital images collected by camera is a very simple and effective method for the field operation. While the technology of hyperspectral imaging is a different technique used for crop growth diagnosis, which has become increasingly popular in recent years. Compared with traditional spectral technology, hyperspectral imaging technology can not only obtain the spectral information of the object, but also acquire the image characteristics and spatial distribution information of the object. In this regard, it integrates the advantages of digital image and spectrum. In addition, it is important that the hyperspectral imaging is a near-earth remote sensing. The research on the hyperspectral imaging data processing and modelling is actually for the purpose of using remote sensing (for example, China’s Gaofen-X series satellite) in the future.
