Using Deep Neural Networks to Evaluate Leafminer Fly Attacks on Tomato Plants
Round 1
Reviewer 1 Report
Please read the attached review
Comments for author File: Comments.pdf
Author Response
Please, see attached file.
Author Response File: Author Response.pdf
Reviewer 2 Report
In this paper, authors present a comparison of three models (U-Net, FPN and LinkNet) coupled with backbones (VGG16, Resnet34, DenseNet121 and InceptionV3) for the automatic detection and estimation of symptom severity in tomato leaves due to leafminer attacks. They tested on a dataset consisting of 90 images that were categorized into 3 classes: background, leaf area and symptoms caused by leafminer flies. Their results show that the combination that provided the highest values of accuracy, precision, recall and IoU was the U-Net model with the Inceptionv3 backbone for the segmentation task, while for symptom severity estimation it was FPN with the ResNet34 and DenseNet121 backbones with a low value of mean square error. I think this manuscript is very interesting and well organized, however, I consider that has some drawbacks that must be addressed before a possible publication:
1. I consider that the authors should add more information on the models used in order to obtain a better understanding of them.
2. In dataset section, is it appropriate to display sensitive data such as crop coordinates? this to ensure the protection of data such as the location.
3. Why were the photos taken with a cell phone and not with a professional camera?
4. In Figure 3, I recommend including a notation of which model and which backbone is producing each of the results to better understand the description below the image.
5. Figures 7-9 are not enough comprehensive in terms of what you want to show in a linear regression, you should include a brief description of the values.
6. I recommend the authors to include codes of developed algorithms for a proper replication of the experiments shown in this proposal.
Author Response
Please, see attached file.
Author Response File: Author Response.pdf
Reviewer 3 Report
The work aims to automate the decision-making processes taking place during the production of tomatoes in Brazil. It concerns an indirect method of identifying feeding pests (moss moths) on crops grown in 2 selected locations. The identification was made on the basis of information encoded graphically in the form of leaf damage, presented in the form of 90 acquired digital images. The tests were carried out in field conditions. As a classification instrument, 3 known artificial intelligence models were used in the form of neural convolutional networks, optimized with deep learning techniques.
The work is a modern approach to the analysis of a digital image and has a distinguished application and utilitarian thread. In order to better understand the purposefulness of the applied scientific and research methodology (quite commonly used in image analysis techniques), it is worth supplementing the work with the following explanations:
- why neural convolutional networks were used. After all, there are "classic" neural classification models, e.g. Kohonen, MLP, RBF, etc.,
- what simulation techniques were used,
- what effect the lighting of the photographed scene in field conditions had on the acquisition of images,
- whether the number of acquired 90 images was not too small, especially in the context of dividing the training set into training, validation and test,
- what parameters of digital images (e.g. colour, shape, texture) were considered to be characteristics.
Author Response
Please, see attached file.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
I am very thankful to authors, who took into account the reviewers' comments; thus, I consider that the manuscript should be published in the present form.