Combination of Hyperspectral and Quad-Polarization SAR Images to Classify Marsh Vegetation Using Stacking Ensemble Learning Algorithm

Round 1

Reviewer 1 Report

1. This study conducted a series of experiments to show the capacity of using the combined data in the classification of wetland vegetation. However, the authors have listed some works about integrating optical data and SAR data for classification, and the motivation of this study is not clear enough in the current version. Therefore, it would be better to clarify the contributions of this study to the international scientific community.

 

2. For the employed images, I am concerned whether the difference in acquisition time will result in errors in classification accuracy. It varies significantly in vegetation between summer and late autumn in a mid-latitude temperate zone.

 

3. Since the author did not show the statistical differences between the four classifiers, I do not know whether the Stacking algorithm does improve the accuracy of classification significantly. The Catboost also showed great performance. If there are no significant statistical differences between the four classifiers, I think the Stacking algorithm is not suggested in the algorithm, because it will take more computational time and increase the complexity of the workflow after all.

 

4. It varies greatly in the number of variables in each scheme. In my sense, more variables will bring more effective information for classification. This is why I think scheme 11 showed the best performance in the accuracy of classification. Therefore, I am not quite sure whether all schemes can be conducted with a similar level of variables, then to show the advantages of the combination of optical and SAR data than the single one. Moreover, it would be better to show why seven types of land cover are misclassified and what kinds of variables introduced by other sensors will help to improve the accuracy of classification.

 

5. In the discussion section, I would expect to see more information about electromagnetic mechanisms regarding why the combination of optical, C-band, and P-band shows better performance on vegetation classification than the single type of data. Please provide more information since the current version is a little bit superficial. Similarly, it would be better to clarify why the authors chose these three ML algorithms, explain the advantages and disadvantages of these three single models, and how Stacking improves the accuracy of classification.

 

Moreover, there are too many typos, grammar errors, and repetitive statements in the current version, please conduct a go-through check and fix them.

 

More specifical comments are shown below

 

1. quad-polarization or fully polarimetric, please be consistent

2. L42 Due to? too many typos to point out.

3. L58 to selection? too many grammar errors to point out.

4. L59 ability to vegetation, hard to understand.

5. though we all know the “SAR” is synthetic aperture radar, please define it at its first presence.

6. section 2.1, provide more details about vegetation in this study site.

7. L133 important international wetland or wetlands of international importance? Be consistent.

8. Headlines 3.3 and 3.3.1 use different words to describe optical images and SAR images, which will confuse the readers.

9. Fig.9, what is HF?

10. L537 to 538, This is because C-band is more advantageous in recognizing semisubmerged shallow-water marsh vegetation and paddy field. Why? More details.

11. L588 identifying wetland vegetation, what kind of vegetation, be specific.

12. L592 identification information. Too vague, more details.

13. L629. Pioneer? should use these words carefully, I do not believe this work is the first or pioneer in this field.

Author Response

请参阅附件。

Author Response File: Author Response.docx

Reviewer 2 Report

The aim of this study is focusses on demonstrate the feasibility of integrating hyperspectral and fully polarimetric SAR sensors to improve the classification accuracy wetland classification. Specifically the study proposed an approach of classifying marsh vegetation in a National Nature Reserve, northeast China, by combing backscattering coefficient and polarimetric decomposition parameters of C-band and L-band quad-polarization SAR data with hyperspectral images. Further show the results of developed an ensemble learning model by stacking Random Forest (RF), CatBoost and XGBoost algorithms for marsh vegetation mapping, and evaluated its classification performance of same vegetation between combinations of hyperspectral and full-polarization SAR data.

My general opinion is that the paper is to complex to read and to follow all the steps done in this research. I suggest simplifying the text and move some data in appendix.

However, there are minor issues to be addressed before the manuscript can be suitable for publication.

Please following my comments and suggestions in the attached file.

My review response of this paper is major revision.

All the best,

 

ZAP

 

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The article Combination of hyperspectral and quad-polarization SAR images to classify marsh vegetation using stacking ensemble learning algorithm is well written and well structured. Wetlands mapping and monitoring is of great interest in the remote sensing community. Because of the complex structure of wetlands, different data types have been investigated over the years. 

In this article, the authors hyperspectral and radar data. The authors also tried a new machine learning model and investigated several datasets. 

Before further consideration of the manuscript, several points need to be addressed.

1. The date of the satellite products are different. Multi-temporary satellite data are common in remote sensing; however, what is the pattern used in this paper? We have July, September, and October. What is the reason for this?

2. Field measurements were done in 2015, 2019, and 2021. But the satellite data is from 2020. Are there any changes in the field during the mentioned years? How is this affecting the results? It would be optimal if we have fielnd data and satellite data with a close date. 

3. The results section is long and hard to read. I suggest some restructuring.

4. The analysis and the results were done over a small study area. Different training and testing areas would result better.

5. I strongly recommend the following study to be included in the introduction and also to be compared withing the results section as it is similar to the conducted study.

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I am pleased with the revision and I have no more comments now.

Reviewer 2 Report

Dear Authors,

I have read  the manuscript reviewed, the authors have carefully considered the comments suggested and tried the best to address them.

My response is accepted the paper in present form.

All the best

Zap