The Potential for Hyperspectral Imaging and Machine Learning to Classify Internal Quality Defects in Macadamia Nuts

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The overall quality of the article is good. I suggest revising and then publishing.

1.       Introduction: Add some information about the importance of internal quality defects in nuts.

2.       2.2. The hyperspectral imaging system, calibration and image acquisition：The description of the specific configuration, parameter settings, and data. preprocessing steps of the HSI system (such as image correction, selection of regions of interest, etc.) is not detailed enough.

3.       2.2. The hyperspectral imaging system, calibration and image acquisition：Introduce the principle of MRMR algorithm and the process of selecting characteristic wavelengths.

4.       The results section lacks discussion on accuracy, recall, F1 score, and other results.

5.       The results section：the explanation of the chemical and physical mechanisms behind the spectral features and classification results of HSI images is not sufficient.

Comments on the Quality of English Language

Minor editing of English language required.

Author Response

Dear Reviewer 1,

Please see the attachment for detailed responses to all you comments.

Regards,

Michael

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study explored the potential of VNIR hyperspectral imaging combined with machine learning techniques to classify internal brown center defects in macadamia nuts, using two main hyperspectral imaging modalities, face-up (showing the internal pulp) and face-down (showing the external surface), and comparing the effects of multiple machine learning algorithms and different feature ranking algorithms on the learning effectiveness of classification models. The authors have made many contributions to this study and the workload is more than sufficient. However, there are some issues that need to be improved before publication.

1) In the abstract, the authors mention that the techniques in this study can be evaluated quickly, but there are no time-related experiments throughout the paper, so how can this conclusion be reached?

2) Why was VNIR hyperspectroscopy chosen instead of some other hyperspectroscopy such as Raman spectroscopy.

3) The author mentions four levels of spectral pre-processing in subsection 2.4, but does not explain why these four methods were chosen, and suggests that a relevant theoretical explanation be given.

4) The lack of theoretical description of each machine learning model and feature ranking algorithm in the Materials and Methods section of Section 2 lacks readability and is suggested to be added.

5) Figures a and b in Fig. 3 do not give the meanings of the two colored curves, and it is suggested that a legend be given.

6) In subsection 3.1, it can be seen from Fig. 3d that the samples of the two categories are not well grouped, why the conclusion that a linear model would be more suitable for classification? Give an explanation.

7) In lines 338 and 339, it is suggested that BC and PK either both use abbreviations or full names to improve readability and standardization.

8) There are some formatting errors in the paper that the authors should proofread carefully and make corrections. For example, in line 410, the title serial number is incorrect and the serial number 4.2 of the subheading should be changed to 4.3.

Comments on the Quality of English Language

There are some formatting errors in the paper that the authors should proofread carefully and make corrections. For example, in line 410, the title serial number is incorrect and the serial number 4.2 of the subheading should be changed to 4.3.

Author Response

Dear Reviewer 2,

Please see the attachment for detailed responses to all you comments.

Regards,

Michael

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper makes a valuable contribution to the field of non-destructive food quality assessment, specifically targeting macadamia nuts. The integration of hyperspectral imaging (HSI) with machine learning offers promising advances in defect detection during post-harvest processing. However, the manuscript would be greatly enhanced by addressing several key issues, particularly regarding validation, sample size, real-world applicability, and technical complexity. I suggest a major revision to address the following points:

 

Sample size and representativeness:

The current sample size of 248 macadamia nuts is relatively small and may not sufficiently reflect the variability across different nut varieties, growing conditions, or processing techniques. It is crucial to expand the sample size to ensure a more representative dataset. Including a broader variety of nuts from different regions, seasons, and varieties would improve the robustness and generalizability of the findings. A larger sample would also allow for a deeper exploration of outliers and potential biases that could affect model performance.

 

Lack of external validation:

The models were validated using the same dataset from which they were trained (through cross-validation), which can lead to overly optimistic results. It is essential to perform external validation on an independent dataset to assess the true generalizability of the models. This would provide stronger evidence that the models can perform reliably on unseen data, which is a critical factor for real-world applications in the food industry.

 

Image orientation and practical limitations:

The study's accuracy relies heavily on the orientation of the macadamia kernels (face-up and face-down), but in a commercial processing environment, the orientation of nuts on a conveyor belt is likely to be random. Address how the model could be adapted to handle this variability, such as developing real-time adjustment mechanisms to account for nut orientation or testing the model under more varied conditions to simulate real processing scenarios. This would strengthen the practical applicability of the proposed system.

 

Complexity of hyperspectral imaging implementation:

While HSI is a powerful tool, it is expensive and may require complex operational setups (e.g., precise lighting and orientation controls), which could hinder its adoption in commercial processing. A cost-benefit analysis of implementing HSI in industrial settings would be beneficial, highlighting any operational challenges and potential solutions. Additionally, discussing alternative methods, such as multispectral imaging, which could be more cost-effective and still offer acceptable accuracy, would provide a more balanced view of the technology’s commercial feasibility.

 

Insufficient discussion of model assumptions and limitations:

The manuscript does not sufficiently address the assumptions made during the study, such as uniform lighting or controlled nut handling conditions. Expanding on how varying environmental factors (e.g., humidity, lighting conditions) or kernel conditions (e.g., damage or cracking) might affect the performance of the models would provide a more comprehensive discussion. Additionally, the assumptions underlying the machine learning models (e.g., why linear models outperformed non-linear ones) should be clarified, including their limitations in the context of hyperspectral imaging.

 

Limited exploration of advanced machine learning techniques:

Although several machine learning techniques were tested, there is limited discussion of more advanced models that could enhance performance with the dataset, such as convolutional neural networks (CNNs), which are particularly well-suited to image classification tasks. Exploring advanced models like CNNs or ensemble methods would offer insights into whether these techniques could outperform the models currently in use. The reasoning behind selecting specific machine learning techniques should be clearly articulated, particularly when working with high-dimensional data such as hyperspectral images.

 

Clarity of writing and technical jargon:

Some sections of the manuscript, particularly in the methodology and results, contain dense technical jargon that may not be accessible to all readers, especially industry professionals. Simplifying the language and providing brief explanations of key concepts (e.g., hyperspectral imaging, machine learning models) would make the paper more approachable for a broader audience. This is especially important for readers unfamiliar with the technical aspects of the research.

 

Figures and visualizations:

While the figures and tables are informative, the accompanying legends and descriptions could be more detailed to enhance reader comprehension. Consider adding more visual aids that summarize key findings, such as flowcharts illustrating the model selection process or comparative performance plots of the different models. This would make the results section easier to navigate and reinforce the main conclusions of the study.

 

By addressing these issues, the paper will not only enhance its scientific rigor but also increase its relevance and potential for practical application in the food industry. These revisions will strengthen the manuscript, making it more impactful and accessible to both researchers and practitioners.

Author Response

Dear Reviewer 3,

Please see the attachment for detailed responses to all you comments.

Regards,

Michael

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have solved all the problems I raised.

Reviewer 3 Report

Comments and Suggestions for Authors

Thank the authors for their kind responses.