Image Segmentation Method for Sweetgum Leaf Spots Based on an Improved DeeplabV3+ Network

Round 1

Reviewer 1 Report

Dear Authors,

I have finished my review on the paper. While the topic of the paper could be suitable for Forests, the format, content and the general way of writing it prohibits me to recommend it for publication. I am sorry for this outcome! Perhaps a carefully checked and improved version could be suitable for publication.

Best regards,

R.

Structure: Fails to adhere to the requirements of the journal. For instance, the Introduction starts at point 0. Then at point 1, it should be the Materials and Methods, and so on up to the Conclusions. I’ve failed to see a proper organization of the paper to include the typical sections.

Title:

In my opinion it is not OK since the Authors reconfigure existing methods so as to get to their outcomes.

Abstract:

Far from a good one. The Authors did not follow the instructions given by the journal.

Keywords:

Far from a good choice.

Introduction:

I’ve learned too less from the Intro section on why this study is important. Some sentences are not supported by references while being important steps in building a compelling case. Why leaves are more prone to plant diseases? I think that other parts are prone too. The first paragraph fails largely to convince about the utility of the study, then the following ones are enumerations of what has been done so far without a clear connection to what could be improved. In my opinion, terminology used is also questionable. An important part of the Introduction is rather materials and methods. No gaps in knowledge are identified so as to better formulate the goal and objectives of the study, which are missing in the manuscript.

Materials and methods:

Improperly formatted. Not only the titles of sections and subsections but also the general way of phrasing. For instance: 1.1. Get Sweetgum leaf Spot Images (by the way, who should get it?) >>> the text is rather indicating what one should do to take the image samples. Same in 1.2, where I also questioning the terminology and phrasing, as well as sharp/sudden introduction of new, unexplained terms. English phrasing is poor. In addition, figure quality is poor. Section 2.1 – no references provided to the concepts. Section 2.2 – grammar errors in the text. Also, part of it is purely something known and it should be included in introduction, not here.

Section 3: Based…. (What is based?). Poor phrasing. Part of the text does not belong here.

Table 1: tables need to be self-explanatory. Table 1 does not adhere to this rule. Operators are not explained in the text. The text following the table introduces new concepts which are not explained.

Text following Figure 6: not too academically written. Equation 1: not properly referred and the terms are not properly explained.

3.3.1. In general, it is not built in a suitable way to make sense of the steps used. Same for the rest of the text up to section 4.

4.1. It makes no sense to build a section from one sentence.

4.2. The text is written in a non-academic way.

4.3. Terminology is poor and there are errors in words. Same for the following text up to Figure 12. Figure 12 seems to have some problems in the text below images.

Conclusions:

Difficult to evaluate the implications of the study since the paper is very poorly written. I’ve failed to learn from it.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript presents an interesting work on Sweetgum Leaf Spot Image Segmentation method based on improved DeeplabV3+ network.

 

The manuscript is well written and extensive experimentation is carried out.

 

In Introduction section, more concentration is given with respect to the handcrafted and deep learning for Leaf defect detection. Since more recent literature is available with varying deep learning framework the literature survey section should be elaborated. There can be a separate Literature survey section.

 

“Yuan et al. (2022), An improved Deeplab v3+ Deep Learning Network Applied to the Segmentation of Grape Leaf Black Rot Spots, Front. Plant Sci. 13:795410” the authors too modified Deeplab V3+ for the same application. The author can compare and correlate the modifications with the proposed.

 

In Introduction Section Paragraph 2 what does graphic segmentation means?

 

The main contributions of the manuscript should be highlighted in the introduction section.

 

The author clearly pointed out the advantages of using Mobilenet as backbone network. Why not other light weight networks? Did the authors tried with other networks for performance comparison?

 

Detailed visualization of intermediate feature map for understanding the proposed architecture is needed.

 

In Figure 12, the groudtruth image can also be displayed.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Paper describes the use of neural network-based segmentation methods to detect spots on the leaves of a plant corresponding to its disease. According to the degree of leaf spot coverage, the leaves are then statistically classified into four levels of coverage.

The paper describes the different methods used, which are described sufficiently and have a logical structure. The methodology used corresponds to the standards followed in the scientific community.

The text is written in a clear manner and there are few typos or errors (e.g. "epich" instead of epochs in section 4.3). In addition, the style in the figures of the flowcharts is not consistent throughout the paper (alignment in text blocks, text background, etc.) In Figure 12, the title is not given in English for the input image. Based on the above, I recommend proofreading and correction.

For training purposes, the dataset was augmented from the original 160 to 7680 images, which were used for training, validation and testing of the neural networks. Due to this enhancement, there is a partial overlap of key features of the original images with the artificially generated images. In testing, it is useful to keep the test set completely separate from the training set. Has it been verified that there is no partial overlap between these sets?

Through experiments with added modules, it was shown that the modules in place bring improvements in each of the observed metrics. In addition to the evaluation provided, it would be useful to further describe the parameters of the topologies themselves in terms of the number of trainable parameters, the speed of execution of the segmentation itself, and the size that the networks occupy in memory. It is recommended to extend the paper with these parameters to improve the idea of the effect of the added modules on the size of the topology and the execution speed of the segmentation task.

Finally, the plant leaf disease was divided into 4 groups. Do these groups have any significance in relation to the health of the plant, e.g. for the possibility of curing it, the probability of curing at some level, etc.? If such conclusions can be assigned to the groups, their addition is encouraged.

On the basis of this evaluation, I recommend the paper for acceptance for publication after appropriate revision.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Authors,

I have finished my revision on this new version of your manuscript.

While it was better than the previous one, it still needs your careful attention to produce a better paper. In particular, the way of phrasing, organization of text, avoiding long and unnecessary sentences, format of the paper, quality of the figures and so on, should be in your main focus when attempting to improve the manuscript.

For now, I will recommend minor revisions but you should read again carefully the text and improve it as it still lacks clarity in many of its parts.

Best regards,

R.

Author Response

Please see the attachment.

Author Response File: Author Response.docx