Non-Destructive Quality Estimation Using a Machine Learning-Based Spectroscopic Approach in Kiwifruits

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript “A Machine-Learning-Based Spectroscopic Approach for 2 Non-Destructive Quality Assessment in Kiwifruits” with the objective “ The current study explores the application of non-destructive spectral and hyperspectral 15 imaging techniques for assessing the internal quality of kiwifruit cultivar "Hayward", focusing on 16 physiological parameters like soluble solids concentration (SSC), dry matter (DM), firmness, and 17 tannins. Particularly, the current study employs regression models including Principal Component 18 Analysis (PCA), Partial Least Squares Regression (PLSR), and Three-Layered Neural Network 19 (TLNN) to predict these quality parameters.”

The subject is current and relevant from a scientific point of view, but in terms of methodology it has serious flaws:

- The number of samples is low, it has 20 orchards, but it says that "From each orchard, a total of 10 individual fruits", which is a very low value for internal validation of results, which is a pity because they have a very good value of orchards analysed (external validation).

- Currently the S.I unit for firmness is Newtons and not kg.

-Samples were taken from 20 orchards and we only have a table with average values. It would be important to see these results for each orchard and only at the end check the average for all.

- We don't have spectral information at harvest, so it would be important to understand what happened during the 40 days of storage.

- What are the parameters of the kiwis at harvest? Considering the variability of the orchards, they are rarely the same!

- Why is there such a low number of different "n"s in Table 2?

The proposed spectral analyses are very good and well designed, but the number of samples is low, as is the processing of the analysed data. These two parameters need to be rethought and improved. If what is described in the materials and methods is that 10 fruits were analysed for each orchard, we should have n=200, how do you explain this difference?

After correcting the methodology and analysing the data, the conclusions should be reviewed.

 

A large number of self-citations, for example the author Molassiotis appears 8 times.

Author Response

Reviewer #1 comments:

The manuscript “A Machine-Learning-Based Spectroscopic Approach for 2 Non-Destructive Quality Assessment in Kiwifruits” with the objective “ The current study explores the application of non-destructive spectral and hyperspectral imaging techniques for assessing the internal quality of kiwifruit cultivar "Hayward", focusing on physiological parameters like soluble solids concentration (SSC), dry matter (DM), firmness, and tannins. Particularly, the current study employs regression models including Principal Component Analysis (PCA), Partial Least Squares Regression (PLSR), and Three-Layered Neural Network (TLNN) to predict these quality parameters.”

 

The subject is current and relevant from a scientific point of view, but in terms of methodology it has serious flaws:

 

1. The number of samples is low, it has 20 orchards, but it says that "From each orchard, a total of 10 individual fruits", which is a very low value for internal validation of results, which is a pity because they have a very good value of orchards analysed (external validation).

Response: We would like to thank the Reviewer for his/her comment. We should clarify that 20 orchards of 10 fruit were selected to ensure the variation in the fruit material tested.  We aimed to characterize these individual 200 kiwifruits after cold storage to obtain data regarding their sorting and not to correlate the 20 orchards concerning their fruit quality traits. This point was clarified in the revised manuscript (lines: 90-95).

 

2. Currently the S.I unit for firmness is Newtons and not kg.

Response: Thank you for the constructive comment. Firmness units were corrected in the revised manuscript (Table 1, Figure 3, lines: 41, 43, 136, 234, 235)

 

3. Samples were taken from 20 orchards and we only have a table with average values. It would be important to see these results for each orchard and only at the end check the average for all.

Response: We would like to thank the Reviewer for his/her comment. According to Reviewer #1 comment, a Supplementary Table (Supplementary Table 1) containing the kiwifruit quality traits from each orchard is now provided (lines 233-234; 507-508).

 

3. We don't have spectral information at harvest, so it would be important to understand what happened during the 40 days of storage.

Response: Our purpose was not to predict firmness based on spectral information at harvest but to examine the possibility of sorting kiwifruit following a cold storage period of 40 days based on four key quality traits, namely firmness, SSC, DM and tannin content.

 

4. What are the parameters of the kiwis at harvest? Considering the variability of the orchards, they are rarely the same!

Response: Indeed, the variability of the orchards depends on environmental conditions, the age of the orchard, kiwifruit training systems, plant density, irrigation systems, fertilization inputs, etc. Thus, the phenotype of kiwifruits seems identical, but storability and quality significantly differ. In the current work, we use AI approaches to predict the firmness, SSC, and DM after postharvest cold storage (sorting) and for the first time to sort fruit based on tannin content that is associated with the aftertaste intensity of kiwifruits (i.e. consumers acceptance).

 

5. Why is there such a low number of different "n"s in Table 2?

Response: Thank you for the constructive comment. The ‘n’ indicates the number of PLS components. This point was corrected in the revised manuscript (please see Table 2).

 

6. The proposed spectral analyses are very good and well designed, but the number of samples is low, as is the processing of the analysed data. These two parameters need to be rethought and improved. If what is described in the materials and methods is that 10 fruits were analysed for each orchard, we should have n=200, how do you explain this difference?

Response: Thank you for the constructive comment. Indeed, n=200 fruit was used for the analysis. To clarify this issue, an indicator of fruit analyzed (n) was added in Table 2.

 

7. After correcting the methodology and analysing the data, the conclusions should be reviewed.

Response: Thank you for the constructive comment. The conclusion section was revised accordingly (lines: 477-478).

 

8. A large number of self-citations, for example the author Molassiotis appears 8 times.

Response: We consider that these studies were necessary to discuss our results (4 of them) and carry out the measurements of firmness, SSC, DM, and tannin content (the other 4 studies). However, we understand the Reviewer's concern by reducing some self-citations (lines 138, 140).

Reviewer 2 Report

Comments and Suggestions for Authors

1. Both NIRS and hyperspectral imaging (HSI) are applied for assessment the quality of kiwifruit. The advantages of HSI to NIRS should be discussed and addressed the importance of HSI to kiwifruit quality assessment in the introduction section.

2. The split of the calibration and validation sets was not clear. After a model development, it was applied to predict the new samples. The predictive ability is very important to the practical application for asessment the performance of the technologies e.g. HSI.  What validation method was adopted e.g. cross validation, external validation?

3. The data acquisition time should be provided because of the scan mode (sample move or camero move) needed a long time. 

4. What is the reason for the unsmooth plots about 120-150 points in Fig. 2?

5. The tannins is a relative low level. However, the performance of the model (R2 and RMSE) is OK. The discussions and comparisons should be extended for explanation the reason and challenge for the low level target detection using HSI technology.

Author Response

Reviewer #2 comments:

1. Both NIRS and hyperspectral imaging (HSI) are applied for assessment the quality of kiwifruit. The advantages of HSI to NIRS should be discussed and addressed the importance of HSI to kiwifruit quality assessment in the introduction section.

Response: Thank you for the useful comment, which improves the structure and scientific consistency of the text. This point was corrected in the revised manuscript (lines: 59-61).

 

2. The split of the calibration and validation sets was not clear. After a model development, it was applied to predict the new samples. The predictive ability is very important to the practical application for assessment the performance of the technologies e.g. HSI. What validation method was adopted e.g. cross validation, external validation?

Response: A clarification for dataset splits and cross-validation has been added on lines 181-183 as follows:

A 10-fold cross-validation was performed by splitting the dataset into 10 subsets, with each subset being further divided into training-validation-testing sets at 70 %, 20% and 10% respectively.

 

3. The data acquisition time should be provided because of the scan mode (sample move or camera move) needed a long time.

Response:  A clarification has been added in lines 99-100 given as follows:

Each shot required 3 mins to be acquired, due to lower artificial light intensity compared to natural sunlight.

Regarding the setting of the scanning, there was no conveyor belt included so as to move the kiwis’ crates. Regarding the camera move, the spectral camera has been mounted on a static position, following the Specim IQ standards.

To clarify the above information into the manuscript the following text gas been added in lines 111-112¨

In the current study the camera has been placed on a tripod at a static position, following the Specim IQ standards for spectral data acquisition and following the Specim’s calibration procedure.

 

4. What is the reason for the unsmooth plots about 120-150 points in Fig. 2?

Response: A clarification has been added into the label of Figure line 186 as follows:

Figure 2. Raw reflectance values per kiwifruit at each spectral point between 400-1000nm

Figure 2 depicts raw reflectance data before any normalization or smoothing performed from the software. Such fluctuations in the NIR region are common phenomenon that has have been also confirmed by other relative studies that have been previously contacted and don’t affect the analysis negatively after the application of pre-processing methods.

 

5. The tannins is a relative low level. However, the performance of the model (R2 and RMSE) is OK. The discussions and comparisons should be extended for explanation the reason and challenge for the low level target detection using HSI technology.

Response: Thank you for the constructive comment. Tannin content units were corrected in the revised manuscript into mg/kg instead of mg/g (Table 1, Figure 6, lines: 147; 238-239).

 

 

Reviewer 3 Report

Comments and Suggestions for Authors

The safety and quality of fresh fruit deserves the utmost attention and is a priority for both producers and consumers. The authors present in the current study the use of a non-destructive hyperspectral imaging approach for the evaluation of kiwifruit and they give special attention to soluble solids concentration, dry matter, firmness, and tannins using  Partial Least Squares Regression, Bagged Trees Regression, and Three-Layered Neural Network.

Indeed, the learning method is a powerful tool, deep learning can realize highly abstract data to complete artificial intelligence tasks.

In the Materials and Method section, the authors give information about spectral data preprocessing and sample set split, but also about the three different regression analysis methods they have used. The statistical analysis is well described, and the tables and figures are complete. The interpretation of results and study conclusions are supported by the data. The discussions are well conducted and the results are well explained.

 

The only observation I have for the authors is that in the future they should provide more details about the hyperspectral analysis and data processing.

Reviewer 4 Report

Comments and Suggestions for Authors

In the present work the authors investigated the use of non-destructive hyperspectral imaginig approach and effectiveness of specific regression analysis methods  in predicting the firmness, soluble solids concentration, dry matter and tannins of kiwifruit.  The paper is well written, well structured, satisfies the form and contains a lot of relevant information that will certainly contribute to the scientific development of this field of knowledge. The aim, methodology and findings of this paper are concise and clearly describe.

The only complaint I have is that the obtained parameters of kiwifruit quality were not more compared with the previously published results of some other authors, and that the results of the application of these regression models were not more compared with the regression models available in the literature.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

Despite confirming the improvement in the manuscript, the problems in the design of the experience continue.

Probably the message you're trying to send is not being transmitted correctly, because I keep seeing the same problems I've already mentioned.

I suggest that the article be extensively revised and the writing revised, because this way the problems detected continue.

Another problem is the title, which is quite ambitious in relation to the results obtained and the methodologies analysed. I suggest changing it, because after reading the article I regard these as preliminary results, which are the starting point for a larger, improved and better conducted study.

Author Response

Round 2

Dear authors,

 

Despite confirming the improvement in the manuscript, the problems in the design of the experience continue.

 

Probably the message you're trying to send is not being transmitted correctly, because I keep seeing the same problems I've already mentioned.

 

I suggest that the article be extensively revised and the writing revised, because this way the problems detected continue.

 

Another problem is the title, which is quite ambitious in relation to the results obtained and the methodologies analysed. I suggest changing it, because after reading the article I regard these as preliminary results, which are the starting point for a larger, improved and better conducted study.

Response: Thank you for your comments. The manuscript was extensively revised, and the title was changed.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I have no questions about this version.

Author Response

Round 2

I have no questions about this version.

Response: We thank once again the Reviewer 2 for his / her constructive comments that improve the manuscript.