BerryNet-Lite: A Lightweight Convolutional Neural Network for Strawberry Disease Identification

Round 1

Reviewer 1 Report (New Reviewer)

This study presents a significant advancement in the application of deep learning for agricultural purposes. The innovative approach of BerryNet-Lite, emphasizing a lightweight design while achieving high accuracy in disease identification, is both commendable and promising for future agricultural technology applications. However, to further clarify the scope, methodology, and implications of your research, I have the following questions:

Why did you opt for a lightweight model like BerryNet-Lite for strawberry disease identification over more complex or deeper neural network architectures? Given the complexity of plant disease identification, which often involves subtle visual cues and variations, a detailed explanation of the decision to prioritize model efficiency and speed over depth and complexity could provide insight into the trade-offs considered during the design phase of BerryNet-Lite.

Why was the MobileNetV3 architecture specifically chosen as the foundational architecture for BerryNet-Lite? MobileNetV3 is known for its efficiency in mobile applications, but the rationale behind its selection and its adaptation for the specific task of strawberry disease identification would be enlightening. This includes any modifications made to the architecture to tailor it to the nuances of agricultural imaging and disease detection accuracy.

Why is the integration of Efficient Channel Attention (ECA) and Multi-Layer Perceptron (MLP) modules emphasized in your novel classification head design? The manuscript mentions the advantages of these integrations in terms of improved generalization capability and enhanced feature extraction. However, a deeper explanation of how these components specifically address the challenges of strawberry disease identification would be beneficial. For instance, how do these modules contribute to the model's ability to distinguish between diseases with visually similar symptoms or progress stages?.

The manuscript "BerryNet-Lite: A Lightweight Convolutional Neural Network for Strawberry Disease Identification" is well-composed with a clear and concise use of the English language, effectively communicating the research's objectives, methods, results, and conclusions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (New Reviewer)

Title: BerryNet-Lite: A Lightweight Convolutional Neural Network
       for Strawberry Disease Identification
Authors: Jianping Wang, Zhiyu Li, Guohong Gao, Yan Wang, Chenping Zhao,
         Haofan Bai, Yingying Lv, Xueyan Zhang, and Qian Li

Summary:

This paper introduces a lightweight network, that the authors call
BerryNet-Lite. The network is designed to identify strawberry diseases
from images. The authors curated a synthetic dataset of images of 4 different
classes (Heathy, Powdery Mildew, Anthracnose, Gray Mold) of strawberry diseases
at different maturity levels of berries.
They have used MobileNetV3 as a backbone of their system.
The designed Experiments indicate that BerryNet-Lite performs better
than the peer network architectures such as ResNet34, VGG16, and AlexNet,
but has fewer parameters.

The article is well motivated and should be of interest to the readers
of MDPI Agronomy.

Comments:

1) Is the synthetic dataset publicly available? Other researchers who work
on automating image-based classification of strawberry diseases may benefit
from it. The dataset was created via inversion and rotation.

In the data availability statement, the authors state that "the data used to
support this study are available from the corresponding author upon request."

The dataset should be made publicly available prior to the publication of
this article. It is an open access standard.

2) On lines 170-172, the authors write:
"Specifically, we utilize professional-grade equipment including the Canon camera
and the DJIMini3 drone to capture high-resolution images within a local strawberry
orchard setting."

How big was the orchard? Where is it located? Why did you use the drone?

3) Lines 138-140: The authors state that "in traditional machine learning methods,
feature extraction relies on manually designed features, requiring expertise to
identify and define characteristics related to strawberry diseases."

Many methods mentioned by the authors in lines 82-135 do not rely on manually
designed methods. This is especially true about deep learning methods. I would
consider removing this limitation or making it more specific and backing it
up with citations.

4) Lines 19-20:  Followed by, we introduce expansion convolution into the receptive
field expansion, promoting more robust feature extraction and ensuring accurate
recognition.

"Followed by" is not grammatically correct here. Did the authors mean
"We subsequently introduced..."

5) Line 65: The main contributions lists are as follows:

Rephrase this sentence as follows: Our main contributions are
as follows:

6) Overall, this a sound technical paper. All standard comparison metrics
have been used. The architecture of the deep network is clearly presented.

7) The algorithms are clearly described in pp. 11 and 13.

8) Which version of BerryNet-Lite from Table 4 was used in Table 6?
How are these two tables related?

9) The relevant references appear to have been cited and the references
section is well organized.

The quality of the English language is adequate. I pointed out a few sentences that should be paraphrased in my comments to the authors.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The author developed a BerryNet-Lite, a lightweight CNN, for detecting strawberry diseases. The author used three types of strawberry diseases along with healthy strawberries in their study. The authors also tried to compare their results with existing standard CNN models. However, the manuscript possesses too many flaws. Therefore, it cannot be accepted in its present form. Please rewrite the manuscript scientifically and academically and submit it again. Here are some major comments.

1. Some word selections and the structure of sentences throughout the manuscript are not suitable, so they should improve their manuscript massively. 
2. Citations style should follow a standard style.
3. Literature reviews are not done systematically to find the research gaps. The authors just put the title of the previous papers rather than a comprehensive analysis of their findings, strengths, and limitations. Completely rewrite the literature review section.
4. The presentation of knowledge gaps in the manuscript is not clear and convincing.
5. The data is crucial; however, the authors overlook the detailed presentation of the dataset, its sources and collection technique.
6. The network design is not clear as per their objective. Figures 3 and 4 contradict itself.
7. Some figure and table captions are not written properly and descriptively.
8. The author used numerous image enhancement techniques as image preprocessing, which should be avoided as far as possible since it just increases the detection process.
9. The authors mentioned that they have used the transfer learning technique and how they use it in the customized model. It's not mentioned in the manuscript.
10. How did the author set the parameter, as they mentioned they had conducted standardization of model parameters, whats that mean?
11. The discussion section was written with too generically and theoretical rather than logical and scientific evidence. They completely failed to discuss the results and findings of the study.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The study proposed BerryNet‐Lite for a four-class image-wise strawberry disease classification task. The manuscript is concise and complete, presenting relevant literature review, dataset explanation, network architecture, experiment design, and final results sufficiently. The ablation test suggested the utility of the proposed network components, and the comparison between BerryNet‐Lite and alternative CNNs clearly showed the superiority of BerryNet‐Lite.

While I do not have major comments, first I would like to urge the authors to open source the network code as well as the dataset, which are necessary for other researchers to replicate the study results.

For section 3.1.2, I do not think the title is appropriate as these techniques are standard data augmentation approaches and they are not necessarily for “image enhancement”. Image reversal and rotation simply do not change image quality, while adjustments of contrast, chroma and brightness can even reduce image quality.

I do not see any features of BerryNet‐Lite are tailored to strawberry imagery. In that sense, the results in table 6 should be replicable based on other public datasets. The outperformance of BerryNet‐Lite over MobileNetV3 in the study is not a trivial result and deserves further investigation, which is also why I urged the authors to open source the code. I suggest the authors to do another comparison experiment involving at least BerryNet‐Lite and MobileNetV3 based on the ImageNet dataset to investigate whether the discoveries from the strawberry dataset still hold true.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The author improved the manuscript as per the comments. However, the manuscript cannot fulfil the standard requirements of academic writing. Please carefully revise your manuscript. Take enough time to improve your manuscript rather than submitting the revised version quickly.

1. The English writing and structure need to be checked carefully and improved. 
2. Rewrite your abstract.
3. Please mention the number of images collected from the local Strawberry orchard and web-collected images and present samples from both sources in Figure 1.
4. What is the meaning of split data in Table 2? If it is the split dataset for training, testing, and validation, the table does not reflect this. If it is the data after augmentation, why is it not multiple of the whole number from the original image number (it is around 5.8), which does not seem genuine?
5. Data augmentation methods in the 3.1.2 heading and description in the paragraph are not linked to each other. Data augmentation methods and image enhancement methods are different things. Please separate them. 
6. Figure 2 is not displaying.
7. It seems the author used only horizontal and vertical flipping and rotation techniques for data augmentation, but how the split data came in Table 2. Please clearly mention those things. You have to present your work in such a way that if somebody wants to repeat your work, he/she should be able to conduct your research.
8. Since the ResNet has many variants, please mention which variant did you use (ResNet18, ResNet34, ResNet50, ResNet101, …)
9. Table 4 and Table 6 results are average ones or what? Please provide the class-wise accuracy, precision, and f1-score for at least your best model (BerryNet-Lite).

English must be improved.

Author Response

please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I am disappointed in the authors’ response to my simply comments and the quality of the manuscript revisions. I urged the authors to open source the dataset and code, the authors refused both and provided no valid reasons. I asked the authors to benchmark their network on a widely-used large public dataset such as ImageNet, the authors highlighted the entire manuscript as if the entire manuscript has been revised, while adding zero new experiment result as I requested.

Author Response

please see the attachment.

Author Response File: Author Response.pdf