A Discriminative Model for Early Detection of Anthracnose in Strawberry Plants Based on Hyperspectral Imaging Technology

Round 1

Reviewer 1 Report

This manuscript used hyperspectral imaging to detect the anthracnose of strawberry plant at early stage.

1. The experiment design: it seemed that the sampled were destroyed, and certain parts of the plant were used. Even if the models developed from this situation had good performances, the experiment lacked in validation, to apply the models to the un-cutted plants.

2. Can the authors used the established models to identify which part of a unknow sample was inoculated?

3. Early detection, indeed, I would be cautious to use these words. I can see nothing about early detection. The authors failed to conduct time series analysis, and the symptom was obvious.

4. Different wavelength selection algorithms could select different wavelengths, please explain why the three methods were used? Were there any special information in the wavelengths? This lacked in-depth discussion.

5. Some Figures were not necessary, for example, Figure 7, Figure 8.

6. An in-depth discussion should be added. Some research had used spectral profiles for the anthracnose detection, please made comparisons and declared the novelty of this study. The wavelengths and models should also be compared.

7. Time series studies from the inoculation to obvious symptoms were suggested to support early detection.

In all, this study followed the general pattern of studies related in this research filed. The authors have made good presentations in the introduction section, and they failed to support what they wanted to do as listed in the introduction. This study had little with early detection. Fusion of spectral features and image features had been studied in various studies, which the authors had to check carefully and presented it in the introduction section. No doubt that a binary classification could achieve good performances, time series analysis were encouraged. Re-organize of data analysis to validate the findings could also be accepted. 

Author Response

Dear Reviewer #1,

Thank you very much for peer-reviewed and assessed our manuscript entitled ”A distinguish model for the early detection of the anthracnose of strawberry plant based on hyperspectral imaging technology” .We really appreciate all your comments and suggestions and prudently revised the manuscript following the comments and suggestions. For the questions you raised, we have made corresponding modifications in the discussion and results to match some of the problems mentioned in the previous introduction. Meanwhile, for some of your questions about early detection, we have also added corresponding microscopic observation diagrams in Fig.6 to prove the effectiveness of our early detection. Some of the specific changes are detailed below

Comment 1. The experiment design: it seemed that the sampled were destroyed, and certain parts of the plant were used. Even if the models developed from this situation had good performances, the experiment lacked in validation, to apply the models to the un-cutted plants.

Response 1: Thank you for your suggestions. First of all, as for the problem of sampling damage you raised, hyperspectral information was collected in a lossless state after acupuncture inoculation, and there was no damage in fact. Fig. 2 and Fig. 5 were designed to confirm the presence of anthrax at the inoculation site, which had nothing to do with the collection of hyperspectral information and were independent processes. Secondly, the experiment you mentioned lacks verification. The model established in References 48,51 has also not been tested in the field. Our work is an exploratory and innovative effort to find out if spectroscopy combined with deep learning and machine learning can detect asymptomatic strawberry plants infected with anthrax at the roots and neck. On the basis of the good model effect obtained in this experiment, field verification will be carried out in the future.

Comment 2. Can the authors used the established models to identify which part of a unknow sample was inoculated?

Response 2: Thank you for asking the key forward-looking questions. The work of this paper is to realize the preliminary exploration of non-destructive detection of strawberry seedling anthrax. On the basis of the results of this study, we will develop the supporting equipment embedded in the model and optimize it, and finally realize the non-destructive detection of whether strawberry seedling is infected with anthrax in the field.

Comment 3. Early detection, indeed, I would be cautious to use these words. I can see nothing about early detection. The authors failed to conduct time series analysis, and the symptom was obvious.

Response 3: Thank you for your suggestions. Testing for early detection of disease is a vexing problem. After acupuncture inoculation of anthrax, anthrax symptoms appear at the inoculation site within a short time (24h) (Fig. 5). This experiment innovatively collects the specific hyperspectral information of asymptomatic areas around the disease site (different from both the healthy and the disease sites). Microscopic examination (three microscope images were added in Fig. 6) showed that pathogenic spores were indeed lurking on the surface of plant tissues. From the perspective of the presence of pathogenic spores but no symptoms of disease, the information of pathogen infection had been captured, which belonged to the category of early detection.

Comment 4. Different wavelength selection algorithms could select different wavelengths, please explain why the three methods were used? Were there any special information in the wavelengths? This lacked in-depth discussion.

Response 4: Thank you for your suggestions. For different methods, we want to compare these three classical wavelength extraction algorithms and observe which algorithm extracts more effective information and extracts the least number of wavelengths, which can greatly reduce the modeling time. The subsequent modeling with deep learning CNN convolutional neural network can reduce the development time of the model. This is the crux of the problem we are considering. The discussion has been added. Please check the discussion section in detail.

Comment 5. Some Figures were not necessary, for example, Figure 7, Figure 8.

Response 5: Agree. We have put Fig. 7 and Fig. 8 into appendix B, named as Fig. A1 and Fig. A2, respectively.

Comment 6. An in-depth discussion should be added. Some research had used spectral profiles for the anthracnose detection, please made comparisons and declared the novelty of this study. The wavelengths and models should also be compared.

Response 6: Accepted. We explain the novelty of this study in the discussion. Most studies use feature wavelength combined with machine learning or PCA image processing combined with machine learning for modeling analysis, while this experiment uses MNF transform combined with feature wavelength extraction and deep learning for modeling. At the same time, the existing method of anthrax detection is in vitro leaf detection. In this study, we performed asymptomatic infection detection from the crown of living plants and combined new image processing techniques with deep learning models, which is our innovation.

Comment 7. Time series studies from the inoculation to obvious symptoms were suggested to support early detection.

Response 7: Thank you for your comments. The key to our study is the identification of asymptomatic infection, so the key to sampling is to sample asymptomatic epidemic areas. In this experiment, we innovatively collected specific hyperspectral information of asymptomatic areas around the disease site, and subsequent microscopic detection confirmed the presence of pathogenic spores lurking on the surface of plant tissues. From the perspective of disease detection, the purpose of early detection had been achieved.

Regards

Chao Liu

Jiangsu Key Laboratory for Horticultural Crop Genetic Improvement, Nanjing Agricultural University Nanjing, Jiangsu, China

Email:[email protected]

Reviewer 2 Report

The article is generally well-written. Authors come up with clear objectives and novel approach to HSI analysis. Authors present interesting post-processing of acquired images of strawberry anthrax infection. The research was generally well designed, however there are some minor issues in the text that need to be double-checked and corrected.

I recommend also some more scientific-sounding bibliography.

Figures could be more legible, especially:

Fig. 6. Text should be corrected - Wavelength [nm] instead of Wavelength\nm. Should be applied to all figures. I also recommend changing values to round sum (ex. 600, 800, etc.)

Fig. 7. Illegible lower axis description

FIG. 11 – Bigger text on axis recommended

Tab 1. In discussion there should be more about Wavelengths selection by algorithms placed in the table

Also, there are some minor issues in the text to be improved

Verse 273 – I would use „peaks” instead of troughs

Verse 370 - TIme should be explained. What are the factors that affects time in those models? Should be also a part of disscussion.

Verse 436 - Discussion should be expanded. More information about methods used, past studies, possible applications to different type of data and other plants and diseases is recommended.

Verse 532 – Should be moved to the discussion

General comment to the language – could be more scientific. I recommend using spelling-check through the language editiorial office. There are some fragments in a text that need to be corrected.
(Ex. Verse 423 – Not necessary sentence, Verse 485 – Unclear sentence)

Author Response

Dear Reviewer #2,

Thank you very much for peer-reviewed and assessed our manuscript entitled ”A distinguish model for the early detection of the anthracnose of strawberry plant based on hyperspectral imaging technology” .We really appreciate all your comments and suggestions and prudently revised the manuscript following the comments and suggestions. We have revised our results and discussion part to correspond to our introduction part. At the same time, we asked professional people to review our articles again and again to revise certain nouns, such as using 'anthracnose' instead of 'anthrax', and some specific changes in the wording are as follows.

Comment 1. [Figures could be more legible, especially: Fig. 6. Text should be corrected - Wavelength [nm] instead of Wavelength\nm. Should be applied to all figures. I also recommend changing values to round sum (ex. 600, 800, etc.) Fig. 7. Illegible lower axis description Fig. 11 – Bigger text on axis recommended. Tab 1.In discussion there should be more about Wavelengths selection by algorithms placed in the table.]

Response 1: Thanks to the reviewers for carefully reviewing the draft and raising many specific questions, we have revised them one by one all the figure /nm in the paper have been modified to [nm], because the selected band is 241 band corresponding to 474-965nm, rounding and summing will cause misjudgment for readers. We accept your suggestion, Fig.7 has been set to supplement Fig.A1 and Fig.A2 at the same time modified according to your suggestion.

We accept your suggestion. Fig. 11 has been transformed into Fig.9 and modified as you suggested. We accept your suggestion. In the discussion section, the discussion of wavelength information selection algorithm is added, specifically modified as ‘After obtaining raw hyperspectral data, feature extraction is crucial. The raw data volume of the captured hyperspectral images is large (up to 2-3G per image), which has the typical high data volume characteristics. Secondly, hyperspectral images have a high spectral resolution and contain a large number of wavelengths, resulting in high feature dimensionality. In addition, hyperspectral images have a strong correlation between various wavelengths, and the inter-spectral correlation coefficients of the images are large, which can easily cause redundant hyperspectral information stacking. To solve the problems of data correlation, redundancy, and covariance brought about by a large number of hyperspectral data and large data volume to reduce the complexity of the model and improve the modeling accuracy and operation speed. Using the characteristic wavelength extraction algorithm can significantly improve the accuracy and speed of the model

Comment 2. Verse 273 – I would use „peaks” instead of troughs.Verse 370 - TIme should be explained. What are the factors that affects time in those models? Should be also a part of disscussion.Verse 436 - Discussion should be expanded. More information about methods used, past studies, possible applications to different type of data and other plants and diseases is recommended.Verse 532 – Should be moved to the discussion

Response 2:We agree use „peaks” instead of troughs . We accept your suggestion. We explained in the discussion that the modeling time is affected by two aspects: 1. The number of layers of the model 2. The amount of input data and the use of feature wavelength extraction algorithm to reduce the amount of data and improve the modeling speed are explained. We accept your suggestion and add new citations to the discussion to compare the differences and novelties between this study and previous studies. Please refer to Verse432-434 for details. We accept your suggestion, Add this description to the model development in the last paragraph of the discussion. Please refer to Verse529-533 for details.

Regards

Chao Liu

Jiangsu Key Laboratory for Horticultural Crop Genetic Improvement, Nanjing Agricultural University Nanjing, Jiangsu, China

Email:[email protected]

Reviewer 3 Report

In this manuscript, the authors report “A distinguish model for the early detection of the anthracnose of strawberry plant based on hyperspectral imaging technology”. By acquiring strawberry hyperspectral images, three spectral data preprocessing methods such as successive projection algorithm (SPA), competitive adaptive reweighted sampling (CARS), and interval random frog (IRF) were used to extract feature wavelengths, and the grayscale co-occurrence matrix (GLCM) was used to extract 12-dimensional texture features of the images, and a monitoring model based on multivariate data was established. It is found that IRF+TF+BP model has the highest prediction accuracy. This study provided a theoretical basis for the nondestructive detection of early strawberry anthracnose.

This manuscript is good in basic grammar and professional term.

Author Response

Dear reviewer#3:

Thank you very much for peer-reviewed and assessed our manuscript entitled ”A distinguish model for the early detection of the anthracnose of strawberry plant based on hyperspectral imaging technology” .

Regards

Chao Liu

Jiangsu Key Laboratory for Horticultural Crop Genetic Improvement, Nanjing Agricultural University Nanjing, Jiangsu, China

Email:[email protected]

Round 2

Reviewer 1 Report

The authors have improved the manuscript, this manuscript can be accepted for publication.