Tree Species Classification in UAV Remote Sensing Images Based on Super-Resolution Reconstruction and Deep Learning

Round 1

Reviewer 1 Report

Tree Species Classification in UAV Remote Sensing Images based on Super-Resolution Reconstruction and Deep Learning

In this paper a method for forest tree species classification using UAVs and deep-learning is proposed. The proposed method compares several types of CNNs, such as ResNet, ConvNeXt, Swin Transformer and Vision Transformer. Also a Real-ESRGAN network for super-resolution reconstruction, denoising, and deblurring of the images is used. Finally a simple data augmentation is also included.

I think that this paper should be of interest to experts in the remote sensing field and also in ecology, since it deals with a common task in monitoring forests. The proposed method for detection of tree species in images captured by UAVs is well posed and solved. Without being very novel, effective techniques that are state of the art in the deep-learning field are used and evaluated. Specifically, this work compares very recent deep learning networks (resnet, transformers, etc) for classification. Also, a pre-processing stage that takes into account reconstruction, denoising, and deblurring of the images is included. For reconstruction and denoising, a recent generative network, Real-ESRGAN, is considered.

The work is well written and organized. The introduction comments on relevant papers and, in the methods section, each of the networks used is briefly and clearly described. The results of the experiments are also correctly presented.

Items to improve:

- Equations (3-1), (3-2) and (3-3) are outside the text. Please integrate these equations in the paragraph above, just when they are quoted.

- In the instructions for authors of this journal it is specified that a discussion section has to be included, which is missing in the current manuscript.

https://www.mdpi.com/journal/remotesensing/instructions

- The drawings are pixelmaps that have low resolution in the PDF. A vector graphics format should be used for best visualization.

Author Response

We really appreciate your time and effort and thanks for the encouragement. 
In the revised version, we use yellow highlighting to emphasize the areas we have modified.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript presents the importance of the subject and clearly outlines its goals and contributions. In this regard, the introduction of the manuscript is well-crafted. However, the study area's general characteristics have not been adequately explained, and the methodology used is superficial. The figures used to illustrate the methodology are not informative, and it is difficult for readers to establish a connection between the figures and the main text. Additionally, the text contains numerous repetitive and brief explanations about the practices carried out under different sub-sections. The results of the experimental tests have not been sufficiently discussed. 

Regarding specific concerns:

Line 65: Simply use "3D-CNN" as CNN has already been explained above.

Line 111: What does "three representative plots" mean? Does this term refer to specific areas? Please provide additional information about these plots and explain why they are being used.

Figure 2.1 is not informative. You have provided geographic coordinates in seconds, but the seconds are all zero! The image on the right is of poor quality and does not clearly depict the forested area. Please improve the map design.

Line 114: "Data Collection and Preprocessing": How many times was the preprocessing step used in the article? Pre-processing is a step that is only applied before the actual processes begin. However, from the manuscript, it seems that there is more than one. Please make the text more consistent.

Line 115: "Four tree species": Are there only four tree species observed in the study area? If there are additional tree species, please specify them and state that these three were selected for testing. If there are only three tree species in total, you can indicate that the test was performed on all three species found in the study area.

Line 122: Does Figure 2.1 depict three samples?

Line 126: It is unclear how the "augmentation" process was carried out. I believe you consider this step as a preprocessing step. Unfortunately, you did not adequately explain how you did it ("random rotations by 90°, 180°, and 270°"). Please clearly present what you did, supported by visuals or figures.

Why is Figure 2.2 provided? Why did you include two separate images for each tree type? What message were you trying to convey with these images?

Does Figure 2.3 depict the general methodology used in the study? Or do I have the wrong idea? If it is as I thought, why did you use the "Methodology" heading on line 148 and provide the information after giving this figure?

What are "Gaussian filtering," "isotropic," "anisotropic," "bicubic," "bilinear," "possion noise," etc., in Figure 3.1? When these terms are presented as figure text, and their relationships with each other are not explained in the main text, the explanations you provide in the manuscript do not make sense!

I have the same criticism for Figure 3.2! What do "conv" and "RRDB Block" mean?

Lines 194-215: All of these evaluations appear to be part of the discussion section of a completed application. However, at this stage, you have not carried out any experimental test results. If you can write these without making comparisons with the methods promised in your manuscript, why did you write this paper? Are these ideas from another study? If so, why not cite those studies? I recommend removing this part from the methodology's primary heading. After obtaining your results, you can discuss them in your own way, referencing the written materials.

Line 222: Why did you use 50? Is this standard? Can another parameter value be used? If you chose to use the number 50, please explain why!

Line 224-241: I could not establish a correlation between Figure 3.6 and the explanations in this section. If there is no compatibility with this issue, please explain it to me. If the issue is true, please correct the text or make Figure 3.6 more illustrative.

Line 246-315: The criticism I mentioned above also applies to all the algorithms mentioned in these subsections. Are there any problems with the figures used? Or are the explanations inconsistent?

Test results are not discussed in this paper. The discussions are presented in the conclusion section. The authors should add a discussion section to the paper, and the test results should be rearranged in a way that reveals the scientific contribution, taking into account previous studies.

The language used in the text is decent, but with further review and revision, it can be improved and refined for greater clarity and impact.

Author Response

We really appreciate your time and effort and thanks for the encouragement. 
In the revised version, we use yellow highlighting to emphasize the areas we have modified.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Figure 2.1 still contains deficiencies. I couldn't understand why the general map showing the study region was deformed by stretching from left to right. Please review Figure 1 of the article I linked below as an example. Then re-evaluate Figure 2.1 yourself.

https://link.springer.com/article/10.1007/s13369-022-07583-x

You use Data Collection and Preprocessing on line 121, and then you explain these processes in detail. Until the end of this subsection, I understood the purpose of the study and what you did in the preprocessing step without any problems. However, the subheading “Experimental procedure”, explains again the purpuse of the paper and reconsiders preproceses (between the lines 156 and 162). I would like to ask why you need to rewrite your aim and prosedures again in experimental prosedure? Or is it a new aim in experimental prosedure? It sounds like you have actually explained the operations you have shown in Figure 2.4. Am I right?

"The workflow of the models is illustrated in Figure 2-4 and comprises the following modules: “preparation and input of crown tree species image data”, “data preprocessing”, and “model training and validation”." If Figure 2.4 includes the basic operations relating to the methodology developed in this study, all of them should be covered in the “methodology” section.

Thank you for explaining the parameter descriptions regarding Resnet. However, writing these statements only to me means assuming that all other potential readers of this paper are experts on Resnet. The reason for all the questions I ask the authors is to make the manuscript easier to understand by enabling them to make the necessary explanations in the paper.

Maybe it needs minor revisions.

Author Response

We really appreciate your time and effort and thanks for the encouragement again.

In the revised version, we use yellow highlighting to emphasize the areas we have modified.

Author Response File: Author Response.docx