A Non-Destructive Method for Identification of Tea Plant Cultivars Based on Deep Learning

Round 1

Reviewer 1 Report

The authors introduce only two parameters to compare between the 5 models more statistical indexes need to be considered which give better and fair comparison 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This article examined transfer learning of five pre-trained models namely EfficientNet-B0, MobileNetV2, MobileNetV3, 13 MobileViT-S, and ShuffleNetV2 for tea plant cultivation.  Authors have discussed about the traditional methods of tea plant growing and the deep learning methods application for tea plant cultivations. Discussed in details about the various research reviews on the CNN techniques adopted for tea plant buds and diseases.  The manuscript is prepared well and suitable for publication in the Forest Journal of MDPI. It is well organised and may be accepted with the following changes incorporated in the manuscript.

·       Required all short forms abbreviations (Example LDA, SVM etc. are not available) at the first instant used in the manuscript

·       Justify the statement at least 10 individual plants were sampled for each cultivar, enough for the analysis. Describe total sample size used in this research

·       Why only one bud and two leaves taken as case. As all the plants in aa area may not grow one bud and two leaves. Why not number days after seeding? After two leaves further leaves may use it or not?

·       How author correlated the taste of the tea with phenol to ammonia ratio. That is any other test conducted (with reference to line number 202)

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper present tea plant cultivar identification using Deep Learning classifiers such as EfficientNet-B0, MobileNetV2, MobileNetV3, MobileViT-S, and ShuffleNetV2. Authors created their own dataset and performed experiments using multiple classifiers with different image resolutions to obtain testing accuracy of up to 99%. I have a few questions/suggestions for the authors

 

1.  for the keywords, please increase the keywords to five. Also, merge the first two keywords into one.

2.  On line 108, please mention what technique was used for the resize?

3.  Why images were rescaled from [0,255] (it's 255, not 256) to [-1,1]. 

4. Random crop is not enough for the augmentation, not it is recommended. Why authors haven't used other augmentation techniques that changes the geometry or color information?

5.  Why authors haven't used he validation dataset?

6.   Authors performed augmentation by randomly cropping to produce five images out of each single image sample. Then they used the centrally cropped for the testing whereas the other images for the training. Authors, more or less are using the same image for testing that they used for training This may be one of the reason to get high testing accuracy. It is recommended to separate the testing images before performing augmentation on the remaining for the training. It may have even caused the unnoticed overfitting. Could you please justify how your models are mitigating overfitting?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

I am still not satisfied with the response. I am concerned about the claim of getting high accuracy. Please perform the following experiment.

 

1.   Out of those 600 images, isolate 100 of them to be used for testing. Do not perform any kind of data augmentation on those 100, do not use them or any of their variant/augmented version for the training.

2.  Apply augmentation to the remaining 500 images so that they can be used for training.

3. Provide the accuracy on those 100 images after the training is completed. 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx