The Promise of Hyperspectral Imaging for the Early Detection of Crown Rot in Wheat
Round 1
Reviewer 1 Report
My comments are as follows:
- The authors should write their contributions at the end of the introduction section in bullet form.
- I have not found any contributions/novelties in the work.
- No imaging techniques have been employed to detect diseases.
Author Response
Dear reviewer,
We are grateful to reviewer for these insightful comments on my paper. We have been able to incorporate changes to reflect most of the suggestions. We have highlighted the changes within the manuscript.
Here is a point-by-point response to your comments and concerns.
Point 1: The authors should write their contributions at the end of the introduction section in bullet form.
Response 1: Thank you for pointing this out. We agree with this comment. Therefore, we write our contributions at the end of the introduction in bullet form to point out our contributions, this change can be found on page 3 and line 111-119.
Point 2: I have not found any contributions/novelties in the work.
Response 2: To clarity the contributions, we have provided a list of contributions at the end of the introduction in bullet form. We also mention that at the conclusion to point out the novelties of our work, based on the initial experiment result. This change can be found on page 14 line 520-529.
Point 3: No imaging techniques have been employed to detect diseases.
Response 3: Thank you for this suggestion. Because this is a review paper to point out the direction of crown rot detection and not focus on imaging techniques. However, we conducted an initial experiment to support our conclusion. This change can be found on page 11 line 440 – page 13 line 512.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors discuss in their article the application of hyperspectral imaging for early detection and disease screening of crown rot. They use hyperspectral imaging in the range between 450 nm, and 1050 nm. This article is neither a review article nor an original research paper. The authors did not, as far as I understood, carried out additional experiments. Instead, they have collected information from other articles. For instance, in section 3.3, they describe published studies devoted to the identification of wheat heads affected by F. pseudograminearum. I believe that a very similar HSI approach can be useful to identify other types of plant diseases, including crown rot disease. What is the real aim of this work, and where is the novelty element?
Below, I enclose several specific comments. Unfortunately, the present version of the manuscript lacks novelty, and therefore, I cannot recommend it for publication.
- Page 3, line 126 – please explain what R^2 is. In general, when classification/discrimination models are constructed, their prediction power is described, for instance, by correct classification rate, sensitivity, specificity …
- Figure 1: please verify the information provided in Fig. 1 – region 380-750 nm corresponds to the visible region and not near-infrared. Moreover, description of the region between 750 nm and 1050 is missing.
- The caption of Fig. 2 – I would say “reflectance spectrum” instead “curve”.
- Page 4, information about SNV – preprocessing of spectra is a wide topic, and in addition, some specific undesired effects are typical for remote measurements. Therefore, I would suggest extending the collection of preprocessing methods mentioned in the text.
- Bearing in mind the character of this article (commentary), the authors should pay more attention to explaining in detail what they mean by classifying images. I would expect more advice from the authors. For instance, please provide a few examples of typical classification/discrimination problems. When and why should one use certain preprocessing techniques and specific machine learning methods?
- The authors should clearly distinguish between classification and discrimination problems, taking into account the mechanism of assignment of samples into considered groups of samples. In particular, there is a large difference between ‘soft’ and ‘hard’ classification rules.
- There are many machine learning techniques available and implemented in different software packages. I suggest starting the story with the most simple ones, for instance, partial least-squares discriminant analysis, SIMCA. SVM is an example of a very complex pattern recognition technique. In some cases is used with very little concern, and much simpler methods can offer similar prediction power.
- Page 5: description of PCA – please also mention that principal components are constructed to maximize the description of data variance. PCA is mostly used as an exploratory method to visualize the structure of the data set. On the other hand, principal components summarize data structure, and thus they can serve as input variables to other techniques, e.g., SIMCA or PCR. It is also relevant to mention outlying samples and the large sensitivity of PCA to outliers. See, e.g.:
[1] I. Stanimirova et al., A comparison between two robust PCA algorithms, Chemometrics and Intelligent Laboratory Systems. 71 (2004) 83–95. https://doi.org/10.1016/j.chemolab.2003.12.011.
[2] M. Daszykowski et al., Robust SIMCA-bounding influence of outliers, Chemometrics and Intelligent Laboratory Systems. 87 (2007) 95–103. https://doi.org/10.1016/j.chemolab.2006.10.003.
[3] M. Daszykowski et al., Robust statistics in data analysis - A review: Basic concepts, Chemometrics and Intelligent Laboratory Systems. 85 (2007) 203–219. https://doi.org/10.1016/j.chemolab.2006.06.016. - 4 – it is recommended to add color scale and refer it to reflectance values.
- Page 10 – the authors mention that the differences between infected and healthy wheat are significant. However, this conclusion is based only on the visual inspection of images at one particular wavelength (800 nm). In my opinion, this conclusion and the selection of a particular wavelength require reasonable justification. Furthermore, it is intended to use HSI as a screening technique for the early detection of a disease of wheat. I wonder if the period of 42 days can be considered as an early detection possibility?
Author Response
Dear reviewer,
We are grateful to reviewer for reviewer’s insightful comments on my paper. We have been able to incorporate changes to reflect most of the suggestions. We have highlighted the changes within the manuscript.
Here is a point-by-point response to your comments and concerns.
Comment: in section 3.3, they describe published studies devoted to the identification of wheat heads affected by F. pseudograminearum. I believe that a very similar HSI approach can be useful to identify other types of plant diseases, including crown rot disease. What is the real aim of this work, and where is the novelty element?
Response: Thank you for your comment. Although there are many studies to use HSI to identify the plant disease, however, there are very limited studies have been reported to detect crown rot using HSI and the HSI approaches developed for detecting other diseases would not work for crown rot. Compared with most other plant diseases, crown rot as fungal and soil studdle disease does not cause clear symptoms in plant leaves and tissue and therefore it is difficult for people to detect the crown rot at the early infection stage. Based on the review, most papers mention using HSI can distinguish infected and healthy plants at maturity stage when is too late to make optimized management. In contrast, our study is to investigate how to use HSI to detect crown rot disease at the early stage when there are no visible symptoms on leaves.
Point 1: Page 3, line 126 – please explain what R^2 is. In general, when classification/discrimination models are constructed, their prediction power is described, for instance, by correct classification rate, sensitivity, specificity …
Response 1: Thank you for this suggestion, R2 is for regression and precision is for classification. We add a full name ‘coefficient of determination’ with the symbol at the first time, it will be clearer to describe the meaning of the result. This change can be found page 3 in line 136.
Point 2: Figure 1: please verify the information provided in Fig. 1 – region 380-750 nm corresponds to the visible region and not near-infrared. Moreover, description of the region between 750 nm and 1050 is missing.
Response 2: Thank you for pointing out this mistake, we have changed the Fig 1, the near infrared (NIR) light includes wavelengths between 750 and 1050 nanometers.
Point 3: The caption of Fig. 2 – I would say “reflectance spectrum” instead “curve”.
Response 3: Agree, we changed the caption of fig 2.
Point 4: Page 4, information about SNV – preprocessing of spectra is a wide topic, and in addition, some specific undesired effects are typical for remote measurements. Therefore, I would suggest extending the collection of preprocessing methods mentioned in the text.
Response 4: We added an introduction of image pre-processing. This change can be found on page 5 in line 177 – 191.
Point 5: Bearing in mind the character of this article (commentary), the authors should pay more attention to explaining in detail what they mean by classifying images. I would expect more advice from the authors. For instance, please provide a few examples of typical classification/discrimination problems. When and why should one use certain preprocessing techniques and specific machine learning methods?
Response 5: We described our initial experiment which is a good demonstration of binary classification problem. This change can be found page 11 line 444 – page 13 line 512.
Point 6: The authors should clearly distinguish between classification and discrimination problems, taking into account the mechanism of assignment of samples into considered groups of samples. In particular, there is a large difference between ‘soft’ and ‘hard’ classification rules.
Response 6: Hard classifier has more stringent conditions than soft classifier, because the principle of hard classifier is to determine the final result compared to a pixel can only have one and only one category. The soft classifier takes the average of the probabilities of all model prediction samples into a certain category as the standard, and the corresponding type with the highest probability is the final prediction result, so a pixel does not belong fully to one class, but it has different degrees of membership in several classes. Although there are obvious differences between these two methods, but this paper focused on detection of crown rot disease by HSI, not machine learning. So that we just reviewed the most often used machine learning method for plant disease detection and the avoid the detailed review of machine learning. Moreover, there are limited studies of using HSI for crown rot detection and which machine method is most suitable for crown rot detection is an open question.
Point 7: There are many machine learning techniques available and implemented in different software packages. I suggest starting the story with the most simple ones, for instance, partial least-squares discriminant analysis, SIMCA. SVM is an example of a very complex pattern recognition technique. In some cases is used with very little concern, and much simpler methods can offer similar prediction power.
Response 7: Thank you for this suggestion. It would have been interesting to explore this aspect. However, in our study, due to the limitation of the length of the paper, it is impossible to cover all of the machine techniques from the basics to the complex ones. We just reviewed the most widely used machine learning algorithms for HSI processing. And most of the described machine learning methods have proven to be successful in analyzing the water and nutrient content, photosynthesis rate measurement, and detecting the fusarium infected plant in a late growth stage.
Point 8: Page 5: description of PCA – please also mention that principal components are constructed to maximize the description of data variance. PCA is mostly used as an exploratory method to visualize the structure of the data set. On the other hand, principal components summarize data structure, and thus they can serve as input variables to other techniques, e.g., SIMCA or PCR. It is also relevant to mention outlying samples and the large sensitivity of PCA to outliers. See, e.g.:
[1] I. Stanimirova et al., A comparison between two robust PCA algorithms, Chemometrics and Intelligent Laboratory Systems. 71 (2004) 83–95. https://doi.org/10.1016/j.chemolab.2003.12.011.
[2] M. Daszykowski et al., Robust SIMCA-bounding influence of outliers, Chemometrics and Intelligent Laboratory Systems. 87 (2007) 95–103. https://doi.org/10.1016/j.chemolab.2006.10.003.
[3] M. Daszykowski et al., Robust statistics in data analysis - A review: Basic concepts, Chemometrics and Intelligent Laboratory Systems. 85 (2007) 203–219. https://doi.org/10.1016/j.chemolab.2006.06.016.
Response 8: PCA is an established and widely accepted technique for data description and dimension reduction. Instead of a detailed introduction of PCA, we would like to make a brief introduction of PCA related to HSI so that fit the aim of this paper, using HSI detecting crown rot at early stage.
Point 9: Fig 4 – it is recommended to add color scale and refer it to reflectance values.
Response 9: We change our Fig 4, please see the detail in response 10.
Point 10: Comment: Page 10 – the authors mention that the differences between infected and healthy wheat are significant. However, this conclusion is based only on the visual inspection of images at one particular wavelength (800 nm). In my opinion, this conclusion and the selection of a particular wavelength require reasonable justification. Furthermore, it is intended to use HSI as a screening technique for the early detection of a disease of wheat. I wonder if the period of 42 days can be considered as an early detection possibility?
Response 10: Thank you for your suggestion, we change the Fig 4 and brief describe our initial experiment result. Our initial not only provide our contribution in hypothesis that HSI can detect the change of photosynthesis, water and N before visible symptoms on upper stem and leaves to distinguish healthy and infected plant after plant infected with disease at 32 days, but also help readers to understand how HSI could support people to screen the crown rot disease in early stage. Compared with the current method of human inspection at the time of harvesting, this is a big improvement, since there are not visible symptoms at the upper stem and leaves tissue in plant after infection with crown rot. This change can be found on page 11 line 444 – page 13 line 511.
Author Response File: Author Response.docx
Reviewer 3 Report
This manuscript was about the application of hyperspectral imaging for detection of Fusarium psudograminearum in wheat. This is a well writen review. I suggest the authors provide the wheat images taken by normal camera side by side with the images in Fig 3 and 4, which could help the reader to better understand the hyperspectral imaging system.
Author Response
Dear reviewer,
We are grateful to reviewer for reviewer’s insightful comments on my paper. We have accepted most of the suggestions and we have highlighted the changes within the manuscript.
Here is the response to your comments and concerns.
Comment: This manuscript was about the application of hyperspectral imaging for detection of Fusarium psudograminearum in wheat. This is a well written review. I suggest the authors provide the wheat images taken by normal camera side by side with the images in Fig 3 and 4, which could help the reader to better understand the hyperspectral imaging system.
Response: Thank you for this suggestion. For the Fig 3, we cite this fig 3 from the previous study, so it is difficult to make a normal camera version for this figure. For the Fig 4, we have made a major revision by including an initial experiment of using HSI to detect crown rot infected plants. We also add Fig 4 showing our HSI system to help readers to understand how the hyperspectral image are taken by the different sensors. Our experiment not only support our hypothesis that HSI can distinguish healthy and infected plants by detecting the changes of photosynthesis, water and nitrogen before visible symptoms on upper stem and leaves, but also help readers to understand how HSI could support people to screen the crown rot disease in early stage. This change can be found on page 11 line 444 – page 13 line 512.
Author Response File: Author Response.docx
Reviewer 4 Report
This commentary briefly describes the symptoms of crown rot disease and traditional screening methods with their limitations. Then, a literature review is conducted on imaging technologies for disease detection, from colour to hyperspectral imaging. More specifically, the authors highlight the suitability of hyperspectral-based screening methods for crown rot disease.
Screening crown rot is challenging since there are no visible symptoms on leaves at earlier growth stages.
The paper is well-approached and, its structure is valid. Nonetheless, you should avoid extended elaboration. It would be better to organize your text into smaller paragraphs. Also, you should smoothly connect the sections to prepare the reader for what is going to be presented.
The title of the paper should be shorter. The "Fusarium pseudograminearum" can be omitted.
The introduction section presents the general framework of the topic. However, the structure and main contribution of the work is not mentioned at the end of this section.
In sections 2 to 3, the authors' analysis focuses on hyperspectral imaging for Plant Phenotyping and its combination with machine learning methods. The problem is described more abstractly and from a theoretical (without any mathematical models) viewpoint highlighting the importance and contribution of HSI and machine learning in pseudograminearum disease screening.
Moreover, the authors should present (maybe organized in tabular form) the research methods/approaches and outcomes of previous studies on the same topic or similar topics.
Conclusions summarize the main ideas investigated in the study. The content of this section, which concludes the paper, could be extended by referring to open issues and directions.
Finally, concerning the references, the authors have cited works that were published before 2016. Please replace such references with more recent ones.
Author Response
Dear reviewer,
We are grateful to you for your insightful comments on my paper. We have been able to incorporate changes to reflect most of the suggestions. We have highlighted the changes within the manuscript.
Here is a point-by-point response to your comments and concerns.
Point 1: The title of the paper should be shorter. The "Fusarium pseudograminearum" can be omitted.
Response 1: We change the title to ‘The promise of hyperspectral imaging for early detection of crown rot in wheat’.
Point 2: The introduction section presents the general framework of the topic. However, the structure and main contribution of the work is not mentioned at the end of this section.
Response 2: Thank you for pointing this out. We list the main contributions at the end of the introduction in bullet form to point out our contributions, this change can be found on page 3 and line 111-119. In addition, we also summarize our contributions based on the initial experimental result at the conclusion. This change can be found on page 14 line 520-528.
Point 3: In sections 2 to 3, the authors' analysis focuses on hyperspectral imaging for Plant Phenotyping and its combination with machine learning methods. The problem is described more abstractly and from a theoretical (without any mathematical models) viewpoint highlighting the importance and contribution of HSI and machine learning in pseudograminearum disease screening.
Response 3: We have made a major revision based on your suggestion. Based on the review and analysis in section 2 and 3, we made a hypothesis that HSI is a promising method for crown rot detection because HSI can detect the changes of photosynthesis and water and nutrient contents without any visible symptoms on upper stems and leaves. We also conducted an initial experiment to support our hypothesis. This change can be found on page 11 line 444 – page 13 line 512.
Point 4: Moreover, the authors should present (maybe organized in tabular form) the research methods/approaches and outcomes of previous studies on the same topic or similar topics.
Response 4: We created a table to summarize recent research about using His to predict water and nutrient and detect Fusarium infected plants. This change can be found on page 10 table 1.
Point 5: Conclusions summarize the main ideas investigated in the study. The content of this section, which concludes the paper, could be extended by referring to open issues and directions.
Response 5: We agree with your suggestion. We suggested further investigations at the end of the conclusion. This change can be found on page 14 line 530-550.
Point 6: Finally, concerning the references, the authors have cited works that were published before 2016. Please replace such references with more recent ones.
Response 6: Thank you for this suggestion. In the case of our study, there are very limited reports in this topic. Some good reports of HSI for plant phenotyping and machine learning published before 2016 are worth of a review in this paper.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
Dear Authors,
the manuscript has been modified. Unfortunately, some important comments of mine were not addressed.
In my opinion, the mechanism of classification is of crucial importance in such types of studies. It has a large impact on the model's usefulness (predictions) - and I assume it is the intention of the authors.
I am puzzled by response no. 7 provided by the authors. A reasonable approach is to select an optimal method for a given problem and not the one most often used by the others. Therefore, the authors should start with verifying the performance of basic classification/discrimination methods first, and if simple methods fail select more advanced ones. Then, it is sufficient to mention in the discussion that basic methods such as X, Y, and Z were also considered, but figures of merit describing their performance were worse compared to the ones obtained from methods A, B, and C.
Response 8 - in most applications PCA is considered as a dimensionality reduction technique. Thus it is used to explore and visualize data structure. On the other hand, PCA can support the construction of classification models to the great extent. Moreover, it is one of the simplest approaches enabling the construction of soft classification rules, dealing with outlying samples and missing data.
If the authors consider my suggestions I believe that the revised version of the manuscript eventually will be more comprehensive, versatile, and useful for diverse readers. Please reconsider addressing these remarks in the text. I believe that incorporating them will do no harm to the overall presentation.
Author Response
Dear reviewer,
We are grateful to reviewer for reviewer’s insightful comments on my paper. We have been able to incorporate changes to reflect most of the suggestions. We have highlighted the changes within the manuscript.
Here is a point-by-point response to your comments and concerns.
Point 1: I am puzzled by response no. 7 provided by the authors. A reasonable approach is to select an optimal method for a given problem and not the one most often used by the others. Therefore, the authors should start with verifying the performance of basic classification/discrimination methods first, and if simple methods fail select more advanced ones. Then, it is sufficient to mention in the discussion that basic methods such as X, Y, and Z were also considered, but figures of merit describing their performance were worse compared to the ones obtained from methods A, B, and C.
Response:
Thank the reviewer for this suggestion. The reviewer suggested conducting a thorough investigation of the classification algorithms from the simple ones to the advanced ones. However, as this is a commentary paper, as described in the introduction, our aim is to show that hyperspectral imaging technologies have the potential to detect crow rot disease before clear visible symptoms appear and open a new research topic in this area, not introducing a specific detection method we have developed. The literature review and the preliminary experiment already provide enough evidence to support our arguments. We used the well-known and widely used SVM classifier to demonstrate the feasibility of this approach. We agree that other simple or advanced algorithms might work better than SVM, we prefer to leave this as an open question for further study. Over description and discussion of the classification algorithm would not add more contributions to this paper and could make the aim of the paper unclear.
Using the combination of hyperspectral imaging and machine learning for plant disease detection is very complex. From the literature and our experience, a certain machine learning model developed in a specific project would not work well in other projects when the sensors and data collection and data processing methods are different. Thus, in this paper, instead of developing a certain method ready for real applications, we point out the research direction and provide evidence of the feasibility and leave the open questions of sensing methods, data processing techniques and machine learning algorithms for further intensive study.
Considering the suggestion of the review, we add the following in the conclusion:
‘We conducted a preliminary experiment to demonstrate the feasibility of using HSI for crown rot disease detection. However, the different types of sensors, data collection, data processing and machine learning algorithms need further intensive study’. This change can be found page 14 545 - 547.
Point 2: Response 8 - in most applications PCA is considered as a dimensionality reduction technique. Thus it is used to explore and visualize data structure. On the other hand, PCA can support the construction of classification models to the great extent. Moreover, it is one of the simplest approaches enabling the construction of soft classification rules, dealing with outlying samples and missing data.
Response 2: Thank you for this suggestion, we add more explanation for the PCA in the data analysis section. ‘As one of the simplest approaches in hyperspectral imaging analysis method, it has a strength of common method of eigen-decomposition of the covariance matrix for solving problems in construction of soft classification rules, dealing with outlying samples and missing data.’ This change can be found page 6 in line 239 - 242.
Author Response File: Author Response.docx
Round 3
Reviewer 2 Report
Thank you for providing your answers.
Author Response
Dear reviewer,
We are grateful to reviewer for reviewer’s insightful comments on my paper. Thank you for your comments.
