Maize Small Leaf Spot Classification Based on Improved Deep Convolutional Neural Networks with a Multi-Scale Attention Mechanism

Round 1

Reviewer 1 Report

This study proposes the deep learning-based disease quantification method for maize leaves. This is an interesting study, where the authors designed a new deep learning framework called DISE-Net. There are couple of suggestions to further improve the quality of manuscript.

 The detailed comments are as follows:

The title is ambiguous as it represents two types of deep learning networks (Image Classification & Object detection). Also, there should be some indication of new model development in title to highlight the novelty of this study.  

Abstract:

L1-19: Please rephrase this sentence to improve the language.

L25: Comma is misplaced.

The introduction is nicely written where the authors discussed detailed background of the subject.

Table 1: How K is calculated for each image? Is it calculated manually or with some computer program?

L168-177: Disease stage is an important factor to consider here. When were most images captured? During early crop growth stage or in later disease/crop growth stages?

Table 2: Please use consistent uppercase table header.

2.4.1: How the authors come up with stated design of this network? Why were the different layers arranged in this specific order? Did the author use hit and trial method to come up with this design as there is no justification is provided in selection of this specific architecture. The architecture design is the most novel component in this study and required detailed information for the scientific community.

Figure 5 needs detailed caption.

Figure 8 needs proper caption.

The method part is missing evaluation metrics. How the accuracy, Fscore and other indices were calculated for evaluation purposes. A separate section may be added to define all the evaluation metrics.

Table 4: It is unclear if these parameters are used for all the tested deep learning models or DISE Net only? Also, how the authors come up with these parameters? Why was SGD optimizer used for VGG-16? Why not Adam optimizer was used?

Figure 11-a: It is not clear why val_loss is increasing? Please double check.

Were there any underfitting/overfitting models in training/validation? The hyperparameters were not properly selected in this study and there is high probability to observe overfitting or underfitting effects. The authors should present the training and validation losses / accuracies is same graph.

Author Response

请参阅附件。

Author Response File: Author Response.pdf

Reviewer 2 Report

I read with interest and special attention the manuscript "Classification and Detection of Maize Small Spot Disease Based on Deep Learning Network "

Congratulations to the authors for the idea and approach!

But I have some minor questions / recommendations

- Latin name of all phytopathogenic agent / disease mentioned in article

- more detailed description of biologic material used and the environment conditions of occurrence and manifestation of the disease (duration of typhoon period)

Congratulations to the authors for their work!

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors justified all the comments, therefore, I think it is OK.