Improved U-Net Remote Sensing Classification Algorithm Based on Multi-Feature Fusion Perception

Round 1

Reviewer 1 Report

The revision clarified most issues. Further comments:

1. The results still cannot be replicated by readers. The paper should contain a link to a webpage on which the source codes will be released.
2. It is still not shown that the improvement is statistically significant. It cannot be written that there is a significant improvement without appropriate hypotesis testing.

Author Response

Concern # 1: The results still cannot be replicated by readers. The paper should contain a link to a webpage on which the source codes will be released.

Author response:  Thanks for the valuable comment. We have carefully revised the manuscript and the reply as follows:The code and models are publicly available at https://github.com/Yancccccc/CAM-UNet-SVM.We provide information about the original image, the network structure, the trained model and instructions for using the corresponding code at this link.

Concern # 2: It is still not shown that the improvement is statistically significant. It cannot be written that there is a significant improvement without appropriate hypotesis testing.

Author response:  Thanks for the valuable comment. We have carefully revised the manuscript and the reply as follows:

The significant improvement written in this paper is wrong. To verify the validity of the model, we downloaded landsat 8 images of Heshan City near the study area in the same phase (October 2, 2019) with the same sample category as the study area, classified the images of Heshan City using the trained model of the CAM-UNet+SVM, and performed statistical analysis with a standard sample, and the classification results are shown in Figure 1. Table 1 shows the comparison of CAM-UNet+SVM and U-Net accuracy. As seen in Table 1, CAM-UNet+SVM has a 2.67% improvement in accuracy compared to U-Net, but a decrease in accuracy compared to the 92.8% with the study area (Xingbin area) in the paper. This is because the Xingbin district is the model to do training and validation, but the Heshan city is directly the model of Xingbin district for classification, and no training is done for the Heshan city. Although the accuracy of Heshan city is slightly lower, it shows that the model has some applicability. Although the accuracy of this model is low, it is also generalizable to other study areas that have not been trained on the data.

Figure 1. Classification results based on the algorithm proposed in this paper.

 

Table 1. CAM-UNet+SVM and U-Net accuracy comparison.

	CAM-UNet+SVM	U-Net
Accuracy (%)	87.08	84.41

 

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors have addressed my review comments in the manuscript.

Author Response

Thank you reviewer！

Reviewer 3 Report

Authors have addressed my previous feedback. I have no further comments. Authors should improve the quality of English writing.

Author Response

Thanks for the valuable comment. We have carefully revised the manuscript and the reply as follows: We have chosen an English editing service company provided by MDPI, and make changes to the full language based on MDPI's revisions.

Round 2

Reviewer 1 Report

Authors have addressed the identified issues.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The paper describes an approach for remote sensing image classification. Briefly, the proposed algorithm uses a model trained by U-Net to obtain three different levels of features, which are the input of a SVM classifier. The classification results are finally analyzed by majority voting, which provides the final results.

The approach is correctly described and the obtained results are compared with the ones provided by other approaches, a process that illustrate the good performance of the proposed scheme. There are however several recent algorithms that, applied in a similar scenario, should be considered as the basis for comparison. Specifically, the authors should evaluate proposals such as the Res-UNet model (Cao and Zhang, 2020). The Res-UNet has shown its advantage against the U-Net model in landscape segmentation.

Cao, K.; Zhang, X. An Improved Res-UNet Model for Tree Species Classification Using Airborne High-Resolution Images. Remote Sens. 2020, 12, 1128. https://doi.org/10.3390/rs12071128

 

 

Reviewer 2 Report

In this paper, an approach to remote sensing classification is presented. In the opinion of the reviewer, the method is very simple and often used in other fields of computer vision. Also, to experimental parts requires attention. Comments:

1.     It is not clear why the U-Net architecture is suitable for the task. Can a similar approach be performed with a different architecture? 
2.    The contributions of this work and the novelty should be explicitly enlisted at the end of Section I.
3.    The usage of features from different layers of a network and their further separate processing to provide a joint decision has been already shown in the literature. Such approaches should be referred to, highlighting the differences. Now, the contribution of this work is minimal or difficult to assess. 
4.    The experimental protocol is confusing. It is written (page 6, lines 192-193) that 80% of samples were used for training and 20% for validation. It means that there was no testing subset and the results are shown for the validation. This should not happen. Typically 60% goes for training, 20% for validation, and 20% for testing. It is also written that “the same validation samples are used for the classification” (line 220). This confirms my suspicions and should be explained as it seems now that most of the results are wrong.   
5.    Is the obtained improvement (91.33% to 92.76%) statistically significant?
6.    The method is not compared with the state-of0the-art techniques devoted to the classification of remote sensing images.
7.    The paper requires proofreading. There are problems with the lack of space between letters or brackets (e.g., lines: 48, 57, and many more on page 2 or 94 on page 3, etc.)
8.    The presented results cannot be replicated by a reader.

 

Reviewer 3 Report

The manuscript proposes an improved deep learning classification method for remote sensing images. The authors presented the deep learning methods used in the review and explained their motivation very well. Detailed different optimization methods for deep learning models have been laid out as well. In addition, experimentation setting and learning parameters of the deep learning method have been clearly presented.
Overall, this is an excellent article.

Some minor comments:
- Lines 33 - 35 are repeated in lines 36 - 38.

- Line 44 - 45. "However, the above-mentioned early remote sensing image
classification methods are not applicable to the remote sensing images with gradually increasing resolution"  why are the methods not applicable? any reference? if not, explanation?

- I suggest mentioning the full name of the abbreviation when first mentioning the abbreviation. For instance, line 41, first mention of SVM. I suggest writing Support Vector Machine (SVM) first time mentioning SVM.

- Are the compared deep network methods belonging to a specific reference? If not, then what is Deeplabv3plus? A reference/clarification is needed.

 

Reviewer 4 Report

The article reports an Improved U-Net remote sensing classification algorithm based on multi-feature fusion perception. The authors claimed that it can improve classification and segmentation accuracy. However, I have some comments on the article and suggested some improvements. (see comments).

1. Many acronyms were used in the manuscript; a list of acronyms should be provided before the ‘references’ section for easy readership. 
2. There should be a related work section, and more references must be included. 
3. There are many highly cited research papers that improved classification and segmentation tasks on remote sensing images. Authors must include those works.
4. Authors should explore other models that are specialized in classification and segmentation tasks in satellite images and use those models for comparison. 
5. Authors may consider making the code open to the public to strongly support the paper’s overall claim (if possible). 

 

Overall, the paper tried to propose an improved model and claimed that it has superior performance. However, more experiments are needed to support their claim.