Multiscale Entropy-Based Surface Complexity Analysis for Land Cover Image Semantic Segmentation

Round 1

Reviewer 1 Report (Previous Reviewer 3)

This version improved by far the level of the paper. Most of my former comments have been addressed. As a matter of the fact just two comments remains to be considered:

The paper seems to be to much long, in my opinion there some description that can be cut or ri-formulated improving the readability of the paper.

It is not clear which are the inputs and the outputs of the U-net. Please define explicitly those quantities.

Additionally, be careful of the  spaces between sections and figures. (Examples ca be found in: page 16, table 3 is to close to the text below;  page 10, there is no space between Sections 3.4 and 4)

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report (New Reviewer)

The manuscript “Multiscale Entropy-Based Surface Complexity and Applicability for Land Cover Image Semantic Segmentation”, is a well planned and executed work that propose a quantitative definition of multiscale surface complexity using the entropy-based method of deep learning and evaluate its applicability for land cover image semantic segmentation.

 

The topic is not original, because similar research has been carried out by other researchers, but under different conditions. Therefore, I believe that the results of the conducted research provide a lot of new data, significantly enriching the knowledge in this area. I also believe that the research has been conducted correctly and the results are well documented. Conclusions resulting from the conducted research are consistent with the presented evidence and arguments and relate to the main research objective. The discussion of the obtained results against the results of other researchers was also correct. Tables and figures are complete. I believe that the article can be published after taking into account the comments in the table below.

 

In section 3.1. Dataset. “the large-scale land-cover dataset with the Gaofen-2 (GF-2).” About the Landsat Image, how to keep away from “Cloud cover”

 

In line 363-370, why bold text?

 

In line 370, “4. Results.” Maybe the next line.

 

In line 391,  “4.2. Learning of Local Surface Complexity.” Maybe the next line.

 

In line418, Figure 5. Quantification of multiscale entropy-based surface by varying kernel sizes, Please provide the “map elements” such as legend, scale of build-ups, forest, meadow and waters.

 

In line 450-452. “where complexity-informed sampling and constrained optimization improved JI by 3.734% and 2.905%, respectively, when compared to baseline UNet. Together, the two improved JI by 4.288%. ” Please Check the data of percentage.

 

In line 462, regular test PA was only 2.130%. Complexity constraint alone yielded 1.065-1.405%. Please Check the data of percentage.

 

In line 478, complexity constraint had an improvement of about 1.583% in regular test JI. Please Check the data of percentage. Is 1% or 1.583%?

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report (New Reviewer)

1. Why did the study images not use spatial geographic data containing latitude and longitude coordinates, which may result in the study results not being universal?

2. Line285-286, "Of the selected images in the target class, 80% were used to train the model and the remaining 20% were used to test the model", the validation results are missing from this study and would be better if they were available.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report (New Reviewer)

This paper proposes a deep learning method based on multiscale entropy for quantifying and learning the pixel-level complexity for land cover semantic segmentation. While the authors have illustrated their method in detail and demonstrated its superiority over other methods through extensive experiments, I think the main drawback of the method is the lack of innovation. Concerns that need to be addressed are in the following list.

1.     The method in this paper lacks innovation. The network structure comes from UNet. The proposed multi-scale is only expressed by using different convolution kernel sizes, and the entropy-based complexity analysis is also a common sense.

2.     The meaning of the title is not clear. Complexity, complexity modeling or analysis.

3.     The abstract and title should match. And this abstract lacks a generalization and summary of the technology, in short, it lacks a methodological description and an expression of innovative points.

4.     At the beginning of the introduction part, the task of this paper is not clearly stated. It is only broadly described that the processing of remote sensing scenes is very difficult, and the extraction of difficult information is very important. The task of this article is not specified, and the specific challenges for this task is also an empty.

5.     There is a lack of innovation on the network, and the comparison methods are all traditional machine learning methods, lacking the recent sota.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report (Previous Reviewer 3)

I really appreciate your work and I notice all the improvements and, therefore, in my opinion the paper can be accepted in present form. Good job!

Author Response

Thank you very much for agreeing to publish our manuscript in Remote Sensing. We greatly appreciate this opportunity and look forward to sharing our research with the wider scientific community.

Reviewer 4 Report (New Reviewer)

Some state-of-the-art remote sensing image segmentation methods should be included in the literature review, such as:

Chen K, Zou Z, Shi Z. Building extraction from remote sensing images with sparse token transformers[J]. Remote Sensing, 2021, 13(21): 4441.

Hu Y, Wang Z, Huang Z, et al. PolyBuilding: Polygon transformer for building extraction[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2023, 199: 15-27.

Author Response

Thank you so much for the constructive feedback. In the revision, we have expanded the literature review to include additional state-of-the-art remote sensing segmentation methods, including the two references recommended here (Lines 76-77).

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 topic is indeed very interesting and I do see potential of this to be a good publication. However, the writing style is not elegant since it is quite evident it has been extracted from a thesis or a report. Unnecessary and long explanations are seen throughout the manuscript. The authors have not indicated how the U-Net architechture was determined. This is very important since based on that your results may change again. The choice of the error metrics should be more detailed. The reviewer is not fond of seeing too many references in the discussion section since it may challenge the originality and also avoid stating the trivial details; always stay within the context. Table 2 in Section 4.3 already provides lots of details and an extensive explanation is not a necessity in this case. Look into that and revise accordingly. I have give a few suggestions to improve the quality of the manuscript, which can be found in teh attached marked up file. 

Comments for author File: Comments.pdf

Reviewer 2 Report

It is better to provide a more detailed description and definition of the surface complexity. Generally speaking, the surface complexity of remote sensing images is not only related to specific types of surface features, but also to all pixels in the local area. However, this article conducts separate experiments on the five categories. Is this design reasonable?

 

Too many references are too old such as ref. 2, 3, 4, 19, 20, 22, 28, 30, 31, 32, 35, 37, 38, 43, 45, and 56. In addition, the citation format of many references is not standardized and the content is incomplete.

 

The literature review is not sufficient. It’s advised to add a more detailed investigation of the related works. e.g. (1) 10.1109/TGRS.2020.3026051 (2) 10.1109/TGRS.2022.3147513

 

Usually, the training and test set are in different regions to ensure the reliability of the model performance verification. How are the training and test set split considered in this study? How do the authors avoid selection bias?

 

Generally, a performance improvement of 1% on the IoU metric is hard to gain. The precision preserved for the results which evaluated the performance of different methods in tables 1 and 2 is too low. It can not compare the performance difference for related methods with a less than 1% accuracy gap.

 

For experiment 4.2. “Learning of Local Surface Complexity”. The related experiment results should be provided to prove the conclusion such as “We used random forest, XGBoost and UNet to learn the models for predicting local surface complexity. Sensitivity analysis showed that UNet improved test performance (test R2 ) by 5-22% compared to random forest and XGBoost. Since the entropy-based local complexity scores were computed using the convolutional kernels, UNet based on convolutional operators is a natural way to learn the local complexity score. In contrast, random forest, XGBoost and support vector machine etc. can’t explicitly model neighborhoods unless such context information is encoded into the input using the nearest neighbor method. However, the convolution kernels in transposed convolutions are learnable and can be optimized, while in the nearest neighbor interpolation, it is fixed, which also explained the superior performance of UNet in learning the complexity scores than the other methods”.

Reviewer 3 Report

Please refer to the attached file.

Comments for author File: Comments.pdf