Analysis and Research on Rice Disease Identification Method Based on Deep Learning

Round 1

Reviewer 1 Report

The authors present a computer vision deep learning approach for the rice disease regression task. The problem is defined properly. Despite the deep learning method is already mature, the author apply them to their application and achieve good accuracy. This paper is not a machine learning method paper but could be use as an example paper for machine learning applications. I have the follow few questions for the authors to address:

1. The information fusion in the model architecture is good, fully connected layer is used to take the other features than image and CNN based architecture is used to analyze the images. I feel overall the discussion of the neural network model is too detailed and may cause the reader not interested in the results itself. So, I would recommend to shrink the model details a bit and high light the contributions.

2. ResNet and VGG are discussed. But those are techniques before 2018. These days the transformer based model is becoming popular, could you compare with vision transformer as well?

3. The machine learning methods are proposed and compared. While I am more interested how this problem was tackled before these methods? If there is any benchmarks that you can compare with just to show how useful these proposed methods are? Otherwise, the results might lack evidence since you are comparing machine learning with machine learning. 

4. Why resNet looks worse than VGG?

Please modify some typos and grammar, overall it looks fine. 

Author Response

The text in sections 2.1 and 2.2 has been revised and reduced.

In section 2.1 of the question, the author assumes that the condition uk>uj is not written in the text.

There are no other equations listed after equation 10 in this article, so there are no other numbers.

Author Response File: Author Response.pdf

Reviewer 2 Report

In this paper, the authors present the results of rice disease identification using artificial intelligence. These researches are important for rapidly detecting rice blast, false smut, and bacterial blight.  The authors specify that the recognition accuracy of the optimized model is 98.64 %.

 

I have a few comments: 

The title is too general. I suggest to att rice main diseases instead of rice disease.

 

The abstract is ok.

 

Introduction

Citations needed for rice diseases damages (rows 45-46).

Need citations for advantages of deep learning technology (rows 57-59)

 

Material and Methods is ok, but I have a question. What type of computers did you use for this research? Can you say some technical characteristics? 

The Results and Conclusions are ok. 

 

 

 

English is ok

Author Response

The language of this section has been reorganized in the revised manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I think the paper overall is good. Just I feel the authors should compare with the state of the art ML models such as transformer models. But even without that, they paper is in a descent position to be accepted.

Some minor grammar edits are needed. 

Author Response

Response to Reviewer 1 Comments

 

Point 1: I found that sections 2 & 3 should be re‐organized and be shortened. It may be easier for the readers if the authors define properly the mixture of regression model and the class‐ membership equation first before moving to the computation of the GINI and of the Polarization of subgroups. Sections 2.1 and 2.2 are too long and can be significantly reduced. In section 2.1 the authors assume the condition u_k > u_j, but this does not appear anywhere else in the calculation of the mixture of regression model. After equation (10) all the other equations are not numbered.

 

Response 1: Please provide your response for Point 1. (in red)

The text has been revised, please refer to the revised draft for details.

 

Point 2: The probability for a given country h to be in a class k should be the proportion of observations (households) in country h that belong to the income class k. On page 9, the first equation (it would be easier for the reader if the equation is numbered) is not exactly the proportion of people because the authors take the sum of the probability. The interpretation of the equation in not obvious. Normally, after estimating a mixture of regression model we have for each observation its estimated probabilities to be classified into the different classes identified. What is often done is to classify a given observation into the class where its estimated probability is higher. In many software this is also the method used that gives us the proportion of people in each of the classes. The authors should explain the equation on page 9 and how to interpret it. Alternatively, they may use the proportion approach which will make the interpretation easier.

 

Response 2: Please provide your response for Point 2. (in red)

Firstly, I did not clarify which equation "mentioned on page 9 of the text" refers to? The content listed on page 9 is all about comparing the results of the model and training iterations. I have not been able to clarify the reviewer's question regarding this issue.

 

Round 3

Reviewer 1 Report

I suggest the manuscript to be accepted for publication.

English quality is descent.