Review Reports - Computer Vision and Transfer Learning for Grading of Egyptian Cotton Fibres

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript investigated some CNNs with transfer learning for classification of cotton fibre. This manuscript investigated some CNNs with transfer learning for classification of cotton fibre. Although the study has practical significance, it is tough to accept in this presentation and some issues should be considered.

1. The novelty is not satisfactory. Although transfer learning is considered to be used in the classification network, the improvement structure or module for these networks have not been proposed.

2. Because the manuscript considered the transfer learning, the experiment results should compare the performance between with and without transfer learning. It is a suggestion that these comparisons should be considered.

3. Although the CNNs used in this manuscript are classic, they have been proposed for a long time. And some new deep learning networks should be considered, such as Transformer. Furthermore, for the classification of cotton fibre, some new research should be considered and compared. Such as the following studies, not limit to.

[1]Jeyaraj, Pandia Rajan, and Edward Rajan Samuel Nadar. "Computer vision for automatic detection and classification of fabric defect employing deep learning algorithm." International Journal of Clothing Science and Technology 31.4 (2019): 510-521.

[2] Huang J, He H, Lv R, et al. Non-destructive detection and classification of textile fibres based on hyperspectral imaging and 1D-CNN[J]. Analytica Chimica Acta, 2022, 1224: 340238.

[3]Xu, Weicheng, et al. "Cotton fiber quality estimation based on machine learning using time series UAV remote sensing data." Remote Sensing 15.3 (2023): 586.

4. The hard point of the classification of cotton fibre should be solved is not clear. More words are on the requirements of vision processing method. It should be noted that many vision methods could be used in this task but maybe not efficient. The reason for the failure of other methods should be given and explained, so as to reflect the research value of this manuscript. Otherwise, simple combination of some networks and methods could not be accepted.

5. The number of samples is questionable. On the one hand, from table 1, the amount of samples for some grade of cotton is only about one hundred. On the other hand, according to the "All images from all cultivars were divided into two groups; training and validation 297 (70%), and testing (30%).", some training samples are only about 70. So, is this amount of samples enough for CNNs networks? Perhaps other machine learning methods and few-shot networks are suitable for this study. Furthermore, limited testing samples (30 samples) also affect the performances.

Comments on the Quality of English Language

English writing can be understood.

Author Response

Reviewer no. 1

This manuscript investigated some CNNs with transfer learning for classification of cotton fibre. Although the study has practical significance, it is tough to accept in this presentation and some issues should be considered.

Comment 1: The novelty is not satisfactory. Although transfer learning is considered to be used in the classification network, the improvement structure or module for these networks have not been proposed.

Response: The author would like to state the following notes in response to this comment:

- The study investigates a significant challenge that faces the agricultural sector in Egypt and particularly reduces the competency of one of the world-reputed products, which is the Egyptian cotton fibres. The study provides a foundation for a practical, cost-effective, easy to use colour system that provides a robust offline grading of cotton fibres in early stages of handling.

- Applying colour imaging to develop grading models for Egyptian cotton fibres through Transfer Learning (TL) was not studied before especially with the wide range of cultivars listed in the study. The studied cultivars included long staple (Giza 86, Giza 90, and Giza 94) and extra-long staple cultivars (Giza 87 and Giza 96), which represented the premium Egyptian cotton fibres. This indeed adds a comprehensive understanding of the classification models and feasibility of the study to be upscaled into a commercial system.

- The study included a section for model fusion to optimise the classification decision obtained. Fusing CNNs’ models has not been studied before for classifying of cotton fibres and more specifically for Egyptian cotton fibres, considering the number of grades and cultivars in the study.

- Fisher, O. J., Rady, A., El-Banna, A. A., Watson, N. J., & Emaish, H. H. (2023). An image processing and machine learning solution to automate Egyptian cotton lint grading. Textile Research Journal, 93(11-12), 2558-2575.

- Fisher, O. J., Rady, A., El-Banna, A. A., Emaish, H. H., & Watson, N. J. (2023). AI-assisted cotton grading: active and semi-supervised learning to reduce the image-labelling burden. Sensors, 23(21), 8671.

Comment 2: Because the manuscript considered the transfer learning, the experiment results should compare the performance between with and without transfer learning. It is a suggestion that these comparisons should be considered.

Response: The authors would like to state that this manuscript is the third in the same project that aims to study the feasibility to utilise colour imaging and machine learning for grading Egyptian cotton fibres. As mentioned before, there were two previous studies in the same project that focused on using random forest, artificial neural networks, and support vector machines; as well as active learning. Please note the results obtained from the aforementioned two studies were stated in the manuscript (Lines: 440-447) as follows:

“Random Forest (RF) was applied on the same data set in this study and classification accuracy was up to 82.1-90.2% (Fisher, Rady et al. 2023). Additionally, active learning was implemented and the with resulting accuracy of 82.9-85.3% compared with semi-supervised learning which achieved 81.4-85.3% accuracy (Fisher, Rady et al. 2023). Comparing results for individual cultivars reveals that RF outperformed TL for Give 86, Giza 87, and Giza 94. This lack of generalisation may be due to the more complexity associated with CNN models leading to high variance (i.e., overfitting) (Mehta, Bukov et al. 2019). “

Comment 3: Although the CNNs used in this manuscript are classic, they have been proposed for a long time. And some new deep learning networks should be considered, such as Transformer. Furthermore, for the classification of cotton fibres, some new research should be considered and compared. Such as the following studies, not limit to:

[2] Huang J, He H, Lv R, et al. Non-destructive detection and classification of textile fibres based on hyperspectral imaging and 1D-CNN[J]. Analytica Chimica Acta, 2022, 1224: 340238.

[3]Xu, Weicheng, et al. "Cotton fiber quality estimation based on machine learning using time series UAV remote sensing data." Remote Sensing 15.3 (2023): 586.

Response: The authors would like to state that the long computational time was due to the configurations of the laptop used (Dell Precision m4800 laptop that has Intel® Core I7 4th Gen. processor, 16 GM DDRam, NVIDIA Quadro K1100M with 2GB GDDR5 memory). Considering the low GPU ram, it is reasonably common to obtain the reported processing times listed in the manuscript. The computational time should significantly decrease with more sophisticated and higher GPU ram, processor speed, and/or available ram.

Regarding the application of other CNNs such as Vision Transformers, please note that we implemented several CNNs during the data analysis study including GoogleNet, AlexNet, EfficientNet, SqueezeNet, ShuffleNet, ResNet50, ResNet101, VGG16, and VGG19. We only reported the optimal CNNs in the manuscript. We appreciate the reviewer’s suggestion and we agree that there are other possible CNNs that can be investigated in future work. Please refer to the updated manuscript in the conclusions section (Lines: 515-517) where the following sentence was added:

“However, the comprehensiveness of the work will be enhanced by including other commercial cotton cultivars with all standard grades, and applying modern pretrained CNNs such as Vision Transformer, YOLOV9-v11.”.

The authors appreciate the reviewer’s suggestion and assure that the referred manuscripts were cited in the Introduction section as follows:

Lines 155-156:

“In another study, a multi-scaling CNN structure was used to detect six defects in fabrics and the average accuracy and sensitivity were 96.6% and 96.4, respectively (Jeyaraj and Samuel Nadar 2019).”

Lines 173-179:

“Visible/NIR hyperspectral imaging accompanied by 1D CNN were also applied for classifying textile fibres with a resulting classification accuracy of 98.6% (Huang, He et al. 2022). Additionally, UAV-collected and combined RGB and multispectral temporal data were investigated to assess cotton fibres quality in the field. ENVI-CNN was used for segmenting cotton and a Back Propagation (BP) network was implemented for predicting fibre quality and the coefficient of determination (R2) was as high as 80.6% (Xu, Yang et al. 2023).”

Comment 4: The hard point of the classification of cotton fibre should be solved is not clear. More words are on the requirements of vision processing method. It should be noted that many vision methods could be used in this task but maybe not efficient. The reason for the failure of other methods should be given and explained, so as to reflect the research value of this manuscript. Otherwise, simple combination of some networks and methods could not be accepted.

Response: The authors would like to state that no single study handled the classification of Egyptian cotton fibres using transfer learning considering the large number of cultivars and grades used in the study.

Please refer to the updated manuscript (Lines:105-109) where the problem of the Egyptian cotton grading was listed as follows:

“Fibre classing through human experts is significantly affected by a sample’s conditions (i.e., homogeneity, size, sampling procedure, assessor’s skills, experience, performance degradation with time, and bias, in addition to the assessor’s awareness of growing, and postharvest handling of the raw cotton fibres; and finally, the grading conditions especially lighting (Ministry of Agriculture and Land Reclamation 2022). ”

Additionally, please refer to the updated manuscript (Lines: 201-203) where the following sentence was added:

“. Despite the promising results obtained from the previous studies, there is no study investigated the utilisation of TL as a deep learning technique for classifying Egyptian cotton fibres into their standard grades.”

Comment 5: The number of samples is questionable. On the one hand, from table 1, the amount of samples for some grade of cotton is only about one hundred. On the other hand, according to the "All images from all cultivars were divided into two groups; training and validation 297 (70%), and testing (30%).", some training samples are only about 70. So, is this amount of samples enough for CNNs networks? Perhaps other machine learning methods and few-shot networks are suitable for this study. Furthermore, limited testing samples (30 samples) also affect the performances.

Response:

The authors agree that the number of samples does not help in developing a convolutional neural network from scratch. Thus, this study focused on transfer learning as a domain adaptation technique providing a practical solution for datasets with significantly limited number of samples compared with the ImageNet (14,197,122 images). Additionally, the main focus here is classify samples into different grades rather than detecting different impurities like studies listed in the Introduction section (Li, Yang et al. 2010, Jeyaraj and Samuel Nadar 2019). In our study, the problem is to obtain the features where the level of impurity (i.e., stain, stems, and dust) is correlated with the grade. The requirement here is to classify each samples into the appropriate standard grade, which is more challenging than detecting foreign materials. This objective was achieved through transfer learning as shown in the Results and Discussion section.

Please note that as mentioned earlier, this study is the third in the same project where the previous two studies (Fisher, Rady et al. 2023, Fisher, Rady et al. 2023) used shallow machine learning and active learning techniques.

References:

Huang, J., et al. (2022). "Non-destructive detection and classification of textile fibres based on hyperspectral imaging and 1D-CNN." Analytica chimica acta 1224: 340238.

Jeyaraj, P. R. and E. R. Samuel Nadar (2019). "Computer vision for automatic detection and classification of fabric defect employing deep learning algorithm." International Journal of Clothing Science and Technology 31(4): 510-521.

Li, D., et al. (2010). "Classification of foreign fibers in cotton lint using machine vision and multi-class support vector machine." Computers and electronics in agriculture 74(2): 274-279.

Mehta, P., et al. (2019). "A high-bias, low-variance introduction to machine learning for physicists." Physics reports 810: 1-124.

Xu, W., et al. (2023). "Cotton fiber quality estimation based on machine learning using time series UAV remote sensing data." Remote Sensing 15(3): 586.

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript is devoted to the comparison of different computer vision algorithms for cotton classification. The manuscript is very interesting but some points should be clarified.

1. What is the objective of the study? The objective of the work should be described more clearly in the Introduction section.

2. The authors used many CNNs, but all of them are quite outdated. Is there any reason not to use SOTA architectures like Yolo or visual transformers?

3. It is necessary to describe in more detail the conditions under which the classification process takes place. How quickly should it happen, how many cotton fiber samples will be on the conveyor belt, what are the requirements for computer vision systems for rapid action.

Author Response

Reviewer no. 2

The manuscript is devoted to the comparison of different computer vision algorithms for cotton classification. The manuscript is very interesting but some points should be clarified.

Comment 1: What is the objective of the study? The objective of the work should be described more clearly in the Introduction section.

Response: The authors would like to state that the objective of the study was to develop classification models of Egyptian cotton fibres using RGB colour images and transfer learning. Please refer to the updated manuscript (Lines: 203-205) as follows:

“Therefore, the objective of this study is to investigate the potential for developing classification models for Egyptian cotton fibres from various cultivars using computer vision coupled with Transfer Learning( TL).“.

Comment 2: The authors used many CNNs, but all of them are quite outdated. Is there any reason not to use SOTA architectures like Yolo or visual transformers?

Response:
The authors appreciated the reviewer’s advice. Regarding the application of SOTA CNNs such as Transformers and YOLO, please note that we implemented several CNNs during the data analysis study including GoogleNet, AlexNet, EfficientNet, SqueezeNet, ShuffleNet, ResNet50, ResNet101, VGG16, and VGG19. We only reported the optimal CNNs in the manuscript. These CNNs are still in considering the large data set (i.e., ImageNet) that comprises 14,197,122 images. We appreciate the reviewer’s suggestion, and we agree that there are other possible CNNs that can be investigated in future work especially with the availability of larger datasets in terms of the number of cultivars and inclusivity of all standard grades. Please refer to the updated manuscript in the conclusions section (Lines: 515-517) where the following sentence was added:

Comment 3: It is necessary to describe in more detail the conditions under which the classification process takes place. How quickly should it happen, how many cotton fiber samples will be on the conveyor belt, what are the requirements for computer vision systems for rapid action.

Response: The authors would like to state the following in response to the comment:

The classification models were developed through pretraining several CNNs including AlexNet, GoogleNet, SqueezeNet, VGG16, and VGG19.The data was divided into training and validation (70%) and testing (30%) sets. Tuning each pretrained CNN was conducted for each CNN to obtain the optimal classification model based on the loss function. Please refer to the updated manuscript in the following sections: data processing (Lines: 314-324), Fusion of pretrained CNNs’ models (Lines: 326-348), and evaluation of classification models (Lines: 350-361).
Regarding the elapsed time for training each classification model, please refer to Table 3 in the updated manuscript where “Training time (s)” was reported.
The authors would like to state that the study was conducted based on acquiring images under stationary mode as this is more realistic for assessing the grade of the cotton fibre in the early handing stages and can serve as an effective, accurate, and highly reliable offline alternative for human experts.

Reviewer 3 Report

Comments and Suggestions for Authors

This study investigated a computer vision and transfer learning method for grading of Egyptian cotton fibres. The topic is interesting and fits with the scope of the journal. Before the final recommendation, some suggestions are listed for further consideration:

(1) Generally, the reported methodology was insufficiently novel because the network structure and transfer learning process were standard, and no original computational contributions and new functional modules were designed.

(2) How did the author choose the selected CNN models for Egyptian cotton fibre classification? Would the utilized network structure be further optimized? This point should be strengthened. Otherwise, it might read like a course report investigating several classical CNN models on a specific dataset on cotton fibre images.

(3) Comparison fairness is a crucial issue for comparative studies of different deep learning models, which is recommended to clarify further, especially avoiding the local optimum for each investigated model.

(4) Some loss curve and metric figures seemed like MATLAB screenshots. It was recommended to obtain the original data and generate scientific plots.

Author Response

Reviewer no. 3

Comment 1: Generally, the reported methodology was insufficiently novel because the network structure and transfer learning process were standard, and no original computational contributions and new functional modules were designed.

Response: The authors would like to state the following notes to address the degree of novelty in the study:

- The study addresses a challenge that faces the agricultural sector in Egypt and significantly reduces the competency of one of the world-reputed products which it the Egyptian cotton fibres. Such a problem is about to adopt digitalisation techniques in early stages of postharvest process. The study provides a robust foundation for a practical, cost-effective, easy to use colour system that provides a robust grading of cotton fibres in early stages of trading.

- Applying colour imaging to develop grading models for Egyptian cotton fibres through Transfer learning (TL) was not studied before considering the various, high-value cultivars studied including long staple (Giza 86, Giza 90, and Giza 94) and extra-long staple cultivars (Giza 87 and Giza 96). This side of the work adds a comprehensive understanding of quality evaluation of Egyptian cotton fibres and a robust feasibility of the study to be upscaled into a commercial system.

- The study included a section for model fusion to optimise the classification decision obtained. Fusing CNNs’ models has not been studied before for classification of cotton fibres and more specifically for Egyptian cotton fibres, considering the number of grades and cultivars in the study.

Comment 2: How did the author choose the selected CNN models for Egyptian cotton fibre classification? Would the utilized network structure be further optimized? This point should be strengthened. Otherwise, it might read like a course report investigating several classical CNN models on a specific dataset on cotton fibre images.

Response: Please note this study was used to investigate the transfer learning as a domain adaptation technique in deep learning, which is suitable for hard-to-obtain data sets in terms of cost, or practicality which is a common problem facing researchers in the agri-food domain. The authors admit the fact that other modern CNNs are available such as Vision Transformers and YOLOV10 and beyond. Vision-dedicated CNNs are still growing and there is no possibility to “apply all available CNNs” as this is not practical from the computation point of view.

Please note that the authors investigated various pretrained CNNs including GoogleNet, AlexNet, EfficientNet, SqueezeNet, ShuffleNet, ResNet50, ResNet101, VGG16, and VGG19. We only reported the optimal CNNs in the manuscript. These CNNs are already justified and distinguished in their performance on a significantly larger data set which is the ImageNet. The results showed that these CNNs were able to extract the low-level and high-level features in the images and yield relatively accurate classification models that can be used to produce more accurate classification models with the availability of larger data sets.

Regarding the optimisation of the training process, this was achieved through mainly minimising overfitting and improving the generalisation through choosing the hyperparameters as follows:

Using multiple solvers including “adam”, “sgdm”, and “rmsprop”.
of iterations were tried to be 500-5000.
Initial learning rate, several values were tried in the range of 1e-7 and 1e-2 to achieve a balance between computational time and performance.

The optimal pretrained models for each studied CNN was selected based on the loss function in the training phase.

Comment 3: Comparison fairness is a crucial issue for comparative studies of different deep learning models, which is recommended to clarify further, especially avoiding the local optimum for each investigated model.

Response: The authors would like to state the following points in response to this comment:

The samples were obtained from the Cotton Arbitration & Testing General Organisation (CATGO), Alexandria, Egypt, which is the only government authority responsible for officially grading cotton fibres in Egypt. The samples were already graded via licensed human experts. This ensures there is no bias towards choosing one grade or cultivar.
Every cultivar was brought to CATGO from the same demographic region, which ensures the individual fairness (Parraga, More et al. 2025). Please refer to the updated manuscript (Lines 213-214).
Every cultivar had approximately similar numbers of samples in the training and validation group. The number of samples per a single grade was 113-150 for Giza 86, 110-116 for Giza 87, 100-131 for Giza 90, 99-115 for Giza 94, and 97-120 for Giza 96. This ensures that every grade has the same opportunity to be represented in the output classification model, which guarantees the individual fairness (Du, Yang et al. 2020). Please refer to the updated manuscript (Liens 220-224), where this sentence was added.
Every cultivar had the same number of samples in the testing set, which was 10 samples for each grade.
We used the classification performance metrics to obtain a fair judgement of the models. These metrics included accuracy, precision, Recall, and F1-score.

Comment 4: Some loss curve and metric figures seemed like MATLAB screenshots. It was recommended to obtain the original data and generate scientific plots.

Response: The authors would like to state that MATLAB R2021a was uses for all computations implemented in this study (Lines 323-324 in the updated manuscript). In the R2021a version, the option for saving or exporting the training graph was not available. Thus, we had to use screenshots to save the training graphs. However, we worked on enhancing the resolution of the training graphs and the enhanced graphs are shown in the updated manuscript.

Additionally, we made a test on an updated version of MATLAB that offers exported training graphs. Please note that the processing time (3273 27 sec or 2.27 days) that was not sufficient to do all training graphs on the updated MATLAB version and obtain exported images of the training graphs. You can note that there is no difference in resolution or contrast between our enhanced image and the one that was exported.

Exported image for training graph of the Giza 86 using VGG16 (please refer to the attached file).

Enhanced image for training graph of the Giza 86 using VGG16 (please refer to the attached file).

References:

Du, M., et al. (2020). "Fairness in deep learning: A computational perspective." IEEE Intelligent Systems 36(4): 25-34.

Parraga, O., et al. (2025). "Fairness in deep learning: A survey on vision and language research." ACM Computing Surveys 57(6): 1-40.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Although the authors gave a detailed response, some points are not satisfied.
1. In the response of comment 3, why did the author discuss the "long computational time"? Maybe some misunderstandings of "a long time". which means many years. Such as GoogleNet, AlexNet and some proposed CNN have been proposed for many years. And many improvements have been considered in these methods. Therefore, the novelty of the manuscript is inadequate.

2. Furthermore, because in the response of comment 4 the explanation of hard point of the classification is still not clear, the improvement of the proposed methods is hard to correspond to the difficulty of classification. “there is no study investigated the utilisation of TL as a deep learning technique for classifying Egyptian cotton fibres into their standard grades.” is not an appropriate explanation of hard point in classification of Egyptian cotton fibres. Not using a method is not a expression of hard point, just an attempt to solve a problem.

3. The transfer learning using the features from ImageNet has been proposed for many years. Recently, in some professional fields, such as medical image processing, this processing has been discussed with some concerns because the features from nature pictures of ImageNet are inadequate.

In summary, this manuscript has some practical significance. If the authors could explain the hard points in classification of Egyptian cotton fibres clearly, the logic of the manuscript is more reasonable and it has more research value.

Comments on the Quality of English Language

The English writing is acceptable.

Author Response

Reviewer no. 1

Comment 1: The authors have revised the manuscript according to the previous comments. However, the key concern was that the novelty and scientific contribution of this study were unclear. It read like an application report that utilized several existing models and algorithms on a specific dataset. Therefore, no additional knowledge and phenomenon could be directly discovered through the proposed methodology. The reviewer felt difficult to recommend the final acceptance.

Response:

The authors respect the reviewer’s comment. However, we disagree with their conclusion about the novelty of the study. Apart from the detailed response that was submitted in the first revision round, the authors would like to state the following points:

Using pretrained CNNs is a handy technique for fields where data are challenging to collect either from the cost, time-consuming, or practicality of providing enough data points to develop robust machine learning modes. This study utilised the pretrained CNNs that were already proven in the domain of transfer learning.
Please consider looking at the study from a wider field of view where the design of the imaging system, selection of cultivars, selection of the CNNs, and optimising such CNNs. These layers of hard work led to the results discussed in the submitted manuscript.
Please refer to the following Springer page, https://link.springer.com/referenceworkentry/10.1007/978-3-319-55065-7_398

Where novelty was identified as, according to Cambridge Dictionary “The quality of being new or unusual”. As stated in our first round of responses, this study was not investigated before either for Egyptian cotton fibres or developing the models for such a large group of cultivars.

Please refer to the following examples of recent peer-reviewed manuscripts (2023-2025) that used one or more of the CNNs utilized in this study:
VGG16, InceptionV3, MobileNet and DenseNet:

Camgözlü, Y., & Kutlu, Y. (2023). Leaf image classification based on pre-trained convolutional neural network models. Natural and Engineering Sciences, 8(3), 214-232.

DenseNet201, ShuffleNet, ResNet101, VGG16, VGG19, DarkNet19, and Xception.

Matarneh, S., Elghaish, F., Rahimian, F. P., Abdellatef, E., & Abrishami, S. (2024). Evaluation and optimisation of pre-trained CNN models for asphalt pavement crack detection and classification. Automation in Construction, 160, 105297.

Efficientb0, DenseNet201, Resnet101, Resnet50, Inceptionresnetv2, Xception, MobileVnet2, ShuffleNet, Darknet19, NasnetLarge, and AlexNet

Tasci, B. (2023). Automated ischemic acute infarction detection using pre-trained CNN models’ deep features. Biomedical Signal Processing and Control, 82, 104603.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors responded to all my comments. I have no questions about the article.

Author Response

Comment: The authors responded to all my comments. I have no questions about the article.

Response:

The authors appreciate the effort and time the reviewers take to review the manuscript and provide significantly important points to improve the quality of the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have revised the manuscript according to the previous comments. However, the key concern was that the novelty and scientific contribution of this study were unclear. It read like an application report that utilized several existing models and algorithms on a specific dataset. Therefore, no additional knowledge and phenomenon could be directly discovered through the proposed methodology. The reviewer felt difficult to recommend the final acceptance.

Comments on the Quality of English Language

No typical English issues.

Author Response

Reviewer no. 3

Comment 1: Although the authors gave a detailed response, some points are not satisfied.
1. In the response of comment 3, why did the author discuss the "long computational time"? Maybe some misunderstandings of "a long time". which means many years. Such as GoogleNet, AlexNet and some proposed CNN have been proposed for many years. And many improvements have been considered in these methods. Therefore, the novelty of the manuscript is inadequate.

Response:

The authors apologise for the misunderstanding of the meaning of “long time”.

While we agree that the listed pretrained CNNs have been used since 2014, it is scientifically obvious that these CNNs were trained in millions of “real-world”, non-vanishing objects. Consequently, using these CNNs as transfer learning tools will not be “old” by the meaning of” old”. The main criteria here are the results and conclusions advised. The studied CNNs in our work provided robust classification results either individually or when the individual CNNs’ models were fused.

The authors also agree that other modern CNNs were brought to the machine vision domain, and this will be an opportunity to use such CNNs for future work with the availability of larger data sets.

Please note that this study aims to investigate whether the concept of transfer learning is applicable in the problem of classifying Egyptian cotton fibres. The results included in the study showed a robust likelihood to achieve such a target and more optimisation is indeed needed to obtain up-scalable or deployable systems to serve the farmers and traders at the appropriate cost.

Additionally, the authors would like to provide the following recent studies (2023- 2025) that investigated using one or more of the CNNs used in this study:

Surface detect detection, CNNs used: VGG-19, GoogLeNet, ResNet-50, EfficientNet-b0:

Singh, S. A., Kumar, A. S., & Desai, K. A. (2023). Comparative assessment of common pre-trained CNNs for vision-based surface defect detection of machined components. Expert Systems with Applications, 218, 119623.

Detecting COVID‑19 in chest CT images, CNNs used: ResNet (50), VGG (19), VGG (16), and Inception V3:

Hassan, E., Shams, M. Y., Hikal, N. A., & Elmougy, S. (2024). Detecting COVID-19 in chest CT images based on several pre-trained models. Multimedia Tools and Applications, 83(24), 65267-65287.

Classification of MRI brain images, CNNs used: DCNN VGG-19, VGG-16, ResNet50, and Inception V3

Krishnapriya, S., & Karuna, Y. (2023). Pre-trained deep learning models for brain MRI image classification. Frontiers in Human Neuroscience, 17, 1150120.

Crop monitoring, CNNs used: VGG16, MobileNetV2, DenseNet121, and ResNet50

Peng, M., Liu, Y., Khan, A., Ahmed, B., Sarker, S. K., Ghadi, Y. Y., ... & Ali, Y. A. (2024). Crop monitoring using remote sensing land use and land change data: Comparative analysis of deep learning methods using pre-trained CNN models. Big Data Research, 36, 100448.

Detection of lung tumours, CNNs used: MobileNetV2

Gao, Z., Tian, Y., Lin, S. C., & Lin, J. (2025). A ct image classification network framework for lung tumors based on pre-trained mobilenetv2 model and transfer learning, and its application and market analysis in the medical field. arXiv preprint arXiv:2501.04996.

Recognition of pistachio species, CNNs used: AlexNet, VggNet, and ResNet

Patel, F., Mewada, S., Degadwala, S., & Vyas, D. (2023, October). Recognition of Pistachio Species with Transfer Learning Models. In 2023 International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS) (pp. 250-255). IEEE.

Skin cancer detection, CNNs used: MobileNetV2, AlexNet, VGG16, ResNet50, DenseNet201, DenseNet121, InceptionV3, ResNet50V2, InceptionResNetV2, and Xception.

Hossain, M. M., Hossain, M. M., Arefin, M. B., Akhtar, F., & Blake, J. (2023). Combining state-of-the-art pre-trained deep learning models: A noble approach for skin cancer detection using max voting ensemble. Diagnostics, 14(1), 89.

Please note there are numerous studies in various applications that used the same, some, or similar age pretrained CNNs.

Comment 2: Furthermore, because in the response of comment 4 the explanation of hard point of the classification is still not clear, the improvement of the proposed methods is hard to correspond to the difficulty of classification. “there is no study investigated the utilisation of TL as a deep learning technique for classifying Egyptian cotton fibres into their standard grades.” is not an appropriate explanation of hard point in classification of Egyptian cotton fibres. Not using a method is not a expression of hard point, just an attempt to solve a problem.

Response:

The authors would like to state that they referred to “there is no study investigated…” is valid as of the time this work was concluded. The main core of the work was to investigate the possibility of using transfer learning as a time, cost, and computational capability methodology to classify standard grades of Egyptian cotton fibres. The case of using machine vision specifically for classifying these products was not studied before this project. We stated the two peer-reviewed manuscripts that were published out of the same study.

Please note that, the main challenge in this task is to overcome the problem of the shortage of human expert at the early stages of postharvest followed by ginning processes specially in villages where transportation and other logistic means are not easily available.

Please refer to the updated manuscript (Lines: 93-94) where the following sentence was added:

“Additionally, there is a noticeable shortage of human experts due to the degradation of the cotton industry in Egypt in the last three decades (Ahmed and Delin 2019)”

Comment 3: The transfer learning using the features from ImageNet has been proposed for many years. Recently, in some professional fields, such as medical image processing, this processing has been discussed with some concerns because the features from nature pictures of ImageNet are inadequate.

Response:

The authors the reviewer’s comment and agree that the features extracted from pretrained CNNs might not be sufficient for developing generalised classification models in certain fields where the overlapping is not dependable between the problem of study and the ImageNet extracted features. The authors would like to state that we fine-tuned the pretrained CNNs by implementing data augmentation, changing initial learning rates and the number of iterations. The differences might not be obvious between consecutive grades for the non-expert personnel. The advantage of the CNNs is the ability to extract high-level features that significantly affect classification accuracy and that was clear in the confusion matrices as shown in Figure 4a-e of the updated manuscript, and Tables3-4.

Comment 4 In summary, this manuscript has some practical significance. If the authors could explain the hard points in classification of Egyptian cotton fibres clearly, the logic of the manuscript is more reasonable and it has more research value.

Response:

The authors appreciate the time and effort that the reviewer dedicated to providing such important feedback. These comments were important to make the manuscript more readable and beneficial to the scientific community.

Round 3

Reviewer 3 Report

Comments and Suggestions for Authors

The previous comments were not well addressed. The final recommendation of publication could not be approved.

Comments on the Quality of English Language

No Language issues.

Author Response

“Despite the promising results obtained from the previous studies, there is no study investigated the utilisation of TL as a deep learning technique for classifying Egyptian cotton fibres into their standard grades without the need for human-intensive feature extraction and optimisation associated with shallow machine learning algorithms, or the necessity of acquiring large imaging data sets. Therefore, the objective of this study is to investigate the potential for applying digital technologies in addressing a challenge that faces the agricultural systems in Egypt by developing classification models for Egyptian cotton fibres sourced from various cultivars using computer vision coupled with Transfer Learning ( TL) based on pretrained CNNs.

Therefore, it is important to state that due to the limited computational resources, and the availability of pretrained models in MATLAB Deep Learning ToolBox, it was not possible to either develop CNNs from scratch, nor to use the more modern pretrained CNNs. Please refer to the following source from Mathworks, that shows the relationship between the accuracy and relative prediction time using GPU:

https://uk.mathworks.com/help/deeplearning/ug/pretrained-convolutional-neural-networks.html

It is clear that NAS Net-Large, Dese Net-201, Inception-ResNet-v2, DarkNet-53, Xception, need longer training tome to achieve the global minima of the model error.

Furthermore, we stated examples of peer-reviewed articles published in 2023-presents, where each article includes one or more of the pretrained CNNs used in our study.

Please refer to the updated manuscript (Lines: 208-210), where examples of these recent studies were added as follows (Please note that these are just examples of numerous studies):

“The application of pretrained CNNs for image classification was studied in different domains including agriculture (Camgözlü and Kutlu 2023, Patel, Mewada et al. 2023, Peng, Liu et al. 2024), medicine (Krishnapriya and Karuna 2023, Tasci 2023, Hassan, Shams et al. 2024, Gao, Tian et al. 2025), and material science (Singh, Kumar et al. 2023, Matarneh, Elghaish et al. 2024), among others.“

Response: The authors deduce from this comment is to clarify “the hard point of classification”. The reviewer also claims that “the improvement of the proposed methods is hard to correspond to the difficulty of classification.“

The authors stated the objective of the study, which was to investigate the potential of using TL in cotton fibres grading. This means we test the validity of using TL as an alternative deep learning approach to shallow machine learning algorithms such as ANN, Knn, LDA, and decision trees. These shallow techniques were already studied in the previous two studies of the same project as follows:

Please note, that either one of these two studies is about utilising the need for human-intervention for extracting the appropriate features from images to reflect the variability between different grades (classes). This adds more time of choosing, extracting, and optimising the features associated with shallow machine learning algorithms. However, TL only needs fine-tuning which we already implemented in our study. Consequently, results obtained from our TL methodology already saved the time and inconsistency of human intervention when using shallow algorithms. TL is not a new approach and the applications of such an innovative technique will not end due to the fact that it is almost impossible to collect the number of images in most agri-food research areas to build a CNN with the same performance that can be obtained from training pretrained CNNs on millions of images. The objective was made clear in the updated manuscript. Please refer to the updated manuscript (Lines: 206-210 ) where the following sentence was added:

“It is worth stating that one of the key advantages of deep learning, including TL, is the less dependency on human intervention for feature engineering, which is a regular step in shallow machine learning techniques. The application of pretrained CNNs for image classification was studied in different domains including agriculture (Camgözlü and Kutlu 2023, Patel, Mewada et al. 2023, Peng, Liu et al. 2024) , medicine (Krishnapriya and Karuna 2023, Tasci 2023, Hassan, Shams et al. 2024, Gao, Tian et al. 2025), and material science (Singh, Kumar et al. 2023, Matarneh, Elghaish et al. 2024), among others.”.

Response: The reviewer raised the point of the insufficient features extracted through convolutional layers in the pretrained CNNs. The authors partially agree with this opinion for several reasons:

While the domains of training CNNs and applying them are different, fine-tuning is a key factor for acquiring the most representing features. We have added a sentence in the manuscript (Lines 201-203) where we explained the cpncet of using TL for learning low-level features from the source domain:

“In image classification tasks, TL helps use low-level features (i.e., corners, colour, edges) learned from the source domain to segment objects in the target domain (Hosna, Merry et al. 2022)”.

We also stated the concern of negative transfer learning that yields from the inadequate information deduced from the source domain. Please refer to the updated (Lines manuscript (Lines 203-206) where the following sentence was added:

“On the other hand, negative transfer learning is also a concern where the information learned from the source domain is not well-adapted in explaining the variability in the target domain, which obligates to study the transferability between the two domains before implementing TL.”.

Conducting model ensembling or fusion that aimed to using the decisions from different CNNs to enhance the classification performance for each cultivar. Generally, our fusion results showed enhancement in classification performance.
We also added a suggestion to use a modern transfer earning technique which is the deep transfer learning. Please refer to the updated manuscript (Lines 526-529) where the following sentence has been added:

“Moreover, deep transfer learning is a more sophisticated techniques that can apply the knwoledge learned from one domain or task in a another domain, that might differ in distribution, in an aim to reduce costs related to learning such as labeling (Iman, Arabnia et al. 2023).”.