Wheat-Seed Variety Recognition Based on the GC_DRNet Model

Round 1

Reviewer 1 Report

This paper discusses the modified GC_DRNet 2 model for recognizing wheat seeds. The model is based on the ResNet18 network and incorporates the dense network idea by changing its residual module to a dense residual module and introducing a global contextual module, reducing the network model's parameters and improving the network's recognition accuracy, the idea seems to be feasible.

Comments on the work:

1. It is advisable to refine the introduction. The problem of recognizing wheat seeds must be considered from the point of view of world experience, since hundreds of scientific works have been devoted to it. Describe the main disadvantages of existing methods and directions for their optimization and development. More specifically set the goals and objectives of the research being carried out, justify the work being carried out.

2. 3.4.1 states that “Comparative experiments with ResNet18, ResNet34, ResNet50, DenseNet121, and other network models have been performed on the public CIFAR-100 dataset...” which “and other network models” were applied?

3. In 3.4.3. GC_DRNet was compared with ResNet34, ResNet50 and DenseNet121, although 3.4.1 also had ResNet18.

4. Perhaps it is worth explaining in more detail why the ResNet-101 and ResNet-152 networks were not taken in the comparative experiments.

5. There was an error in the numbering of the figures; figure -6 appears twice.

6. I would like to note section “4. Conclusions", everything is clear and to the point, but it is possible to divide the first paragraph into several to improve perception.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The study presented a modified lighter version of ResNet18 model and compared it to ResNet18, ResNet34, ResNet50, DenseNet121 for wheat seed image classification. The results clearly show that the proposed GC_DRNet model is superior than the compared ones in terms of both accuracy and inference speed. However, I observed a few fundamental issues of the study.

The manuscript is purely an engineering report instead of a research article. There are no research questions being answered by the study. No research objectives. No knowledge gaps in current literature are being identified. The novel aspect of the study is its model modification, namely the incorporation of Dense Residual Block and the Global Context Module. Is it known that DRB and GCBlock are lightweight and can improve network performance in current literature? If the answer is already in the papers that proposed these modules, then what is the research purpose of the current study?

No proper literature review in introduction, as which should sharply focus on what is know and unknown about DRB and GCBlock in current literature, rather than application examples of CNN in agriculture. In the context of the study, model application is irrelevant, as the proposed GC_DRNet should outperform the original model for any dataset instead of only wheat kernel images. Hence, a literature review heavily focused on wheat is inappropriate.

Based on the image shown in Figure 1 and 2, I do not seem to see the point of training a model that recognizes images of neatly organized wheat kernels with the same variety in a blue background, let alone any practical application of the model. How will the model be able to recognize individual wheat seed variety when different seeds are mixed together? It seems that the model was proposed for the sake of proposing a model rather than solving an actual research question. In that sense, the study is not justified. By the way, there should be an example image showing each wheat variety in the manuscript.

In Table 6, the shallower network ResNet34 outperformed the deeper network ResNet50, which is unusual and can be questionable, as it does not make sense for a larger model to underperform a smaller model. This observation needs to be investigated in depth.

For papers that modify network architectures to be publishable, either a new module is invented and carefully evaluated, or a state-of-the-art model performance needs to be achieved utilizing existing modules, by comparing the proposed model to existing state-of-the-art models on common datasets, if not large-scale public datasets.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The article is devoted to the recognition of wheat seed varieties during its cultivation using neural networks, namely the GC_DRNet model. The article has scientific and practical interest. For publication in this journal, it is necessary to finalize the article according to the following comments:

1) At the end of the literature review, write what shortcomings the existing studies have, what research gap your research covers.

2) At the end of section 1, clearly formulate the purpose of the study and the objectives of the study in points: 1,2,3….

3) According to the objectives, write conclusions more clearly. Now the conclusions are more like a discussion.

4) Research methodology needs to be improved, use mind maps.

5) How were the images of the original data set marked? Manually or using a neural network? Must be described

6) The article clearly lacks examples. Provide screenshots of the implementation of your development: the learning process; image marking process; examples of recognizing different varieties of wheat (there should be several photos).

7) What is the practical significance of your research? Where exactly will grain recognition be used to identify wheat varieties and for what purpose? In the laboratory, at the elevator, at the bakery, in the field? What devices will take photographs - smartphone, stereo camera, etc.?

8) It would also be interesting to provide a photo of the process of obtaining the dataset, if you did it yourself. If you are using a ready-made data set from the library, you must provide a link

General summary:

- In general, taking into account the fact that the article is submitted to the journal “Agriculture”, it is necessary to pay more attention to the practical significance of the research for growing wheat, provide comparative graphs in relation to different varieties. How does your research improve upon existing research in the context of agronomy? What will be the benefit of using your method for identifying wheat varieties, what is its effectiveness?

- How is it practically planned to carry out classification in the field? Shoot with a smartphone and transmit data, process it on site, what mechanisms and devices will be used in the field, who will be the user of your system? If you have not studied this issue, then you can look in the literature and add to the discussion a few sentences on this matter with links to sources.

- The article pays a lot of attention to the indicators of neural networks, but it is necessary to formulate a practical meaning for growing wheat.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

A few issues remain unsolved in the revised manuscript. The authors have added literature review regarding GCBlock, but the literature review for DRB is still missing in introduction.

Line 152-173 needs to be rewritten. Specific research objectives are missing. The listed “contributions” are inappropriate, as they are still simply a description of what was done engineering-wise in the study, instead of what knowledge gaps in current literature are being filled by the study.

I understand the authors added Figure 1 because of my previous comment, which unfortunately has no practical meaning as no wheat kernel features can be observed. The point of Figure 1 is to show readers what type of objects (different wheat kernel varieties) that the neural network in the study is trying to differentiate. I suggest replacing Figure 1 with close-up images of a single wheat kernel for each variety.

I am not satisfied with the authors’ response to my comment on the practicality of the model, which was trained using neatly organized wheat kernels with the same variety in a blue background as I mentioned. There needs to be discussion on, how, in future studies when different wheat kernel varieties are mixed together and present in the same image, the network in the current study can be utilized to do object detection/instance segmentation type of work for individual wheat kernel identification in images.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Notes in the attached file

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf