Classification of Wetland Vegetation Based on NDVI Time Series from the HLS Dataset

Round 1

Reviewer 1 Report

The authors made most of the changes that were suggested in the previous review. This time, I only found a slight typing error. I also suggested a change in the list of references (comments made next to the text of the article in the attached PDF file). 

Comments for author File: Comments.pdf

Author Response

We added the link to the github, which was recommended by the referee for the 'dzetsaka’ classification.

We corrected the typo in phenology (thanks for catching that.

 

Reviewer 2 Report

The paper by Ju and Bohrer has been much improved. Anyway, there are still a few errors that need to be corrected.

Author Response

We thank the reviewer for the supportive review.

We thoroughly edited the manuscript and corrected several typing and grammar errors. 

Reviewer 3 Report

Line 179,  it is not sufficient for ten samples for classification of Worldview.

Author Response

Comment- Line 179,  it is not sufficient for ten samples for classification of Worldview.

response- 10 was a typo. We had 40 training samples. It is still not a very high number, but the small size of the wetland combined with the intrinsically small size of individual patch do not leave us too many pure pixels to pick for the training. We believe that 40 was a sufficient number for our case. The goodness of fit of the result confirms that.

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

It is a very basic and bland manuscript, like a master student's coursework report. The manuscript attempts to develop a method to classify wetland vegetation. However, it is a very basic application of an existing method and I could not see any innovation.

The introduction is not specific enough in describing the background to the related research area and it is not clear how your work differs from others.

For the methods section, yes, high-resolution images are good to be used for generating training and validation samples. But I don't understand why the authors use unsupervised classification, since only the very basic three vegetation classes needed to be classified? I think it would not be difficult at all to achieve more than at least 80% accuracy for WV-3 images in your study area.  I would recommend trying the supervised classification methods.

    In lines 145-149, I did not see where you used these data.

Lines 202-205, Only five pure pixels were averaged to get the characteristics of NDVI time series, which were not enough.

Lines 205-207, a simple introduction of DTW is needed and the advantages of DTW should be introduced to explain why you used it.

Table 2.  why this is important？ The results of the WV-3 are only used for generating validation samples and do not need to make up the majority of the conclusion section. It can only be used as a small, simple illustration. The focus should be on the analysis of the results of the NDVI- time series.

Figure 6. How many samples are used for the validation of the HLS classification result of 2020? Only 29?!!!!

 

Author Response

Please see you attached file with item by item response. 

Author Response File: Author Response.docx

Reviewer 2 Report

The structure of the article is clear and well organised and contain all of the components (Introduction, Materials and Methods, Study Area, Remote Sensing Data, NDVI Timeseries Construction, Classification Using Dynamic Time Warping), Results (Classification of WorldView-3 image for identification of “pure” pixels), Discussion, Conclusions, References.

It is also mentioned Author Contribution, Funding, Data Availability Statement, Acknowledgments and Conflicts of Interest.

The introduction covers all the aspects that will be developed later. However, I am not convinced that paragraph (46-57) about wetlands as reservoir of sequestered carbon and sources of methane emission is very appropriate of the topic.  Although the authors mention it as the last paragraph at conclusions this is a tangential idea of the paper.

Otherwise the literature is well synthesized and very connect to the data

Materials and Methods is very well developed with sections (Study Area, Remote Sensing Data, NDVI Timeseries Construction, Classification Using Dynamic Time Warping) that describe methodology in detail. The figure describing schematic flowchart of the classification process using NDVI time series computed form HLS dataset is useful in understanding the classification process.

Results are consistent with the evidence and arguments presented and describe well what the data show with tables or figures especially (Classification of WorldView-3 image for identification of “pure” pixels)

To be it easier to understand, the classification results of the WorldView-3 image are showing in Figure 3 (a-b), and a confusion matrix (Table 1) was calculated to evaluate the classification accuracy of the WorldView-3 imagery.

The both figures and tables added (figure 5, the fractional cover, table 2. Percentage of HLS classification pixels, and figure 6. HLS classification result of 2020) allow the release of very interesting disscutions.

In Discussion are clearly explained the results and the main idea that new classified maps clearly show the ecological changes

Anyway I am not understand why the figure 7 (a) Time sequence of patch location and extent in OWC, produced by classification of HLS data from 2016 to 2020  and (b) Rapid increases of Lake Erie water elevation over the last decade is at Discussion and not at Results?

In Conclusions the authors present  their new approach to classify land-cover types using repeated NDVI data from HLS in accordance with the stated purpose.

References

 Gives due recognition to the initial discoveries and related work that led to the work under assessment  

Not over-reliant on self-citation

46-57 this paragraph is not relevant to the paper 

84 the term NDVI is not explained (the explanation appears at 95)

150 the term QA band is not explained

All the species should have a unitary name scientific or popular

190 It is Nymphaea ssp. instead of waterlily 191 Typha ssp and Nelumbo ssp

214 This paraghaph should be at Disscution

224 Nelumbo ssp. insetead of Nelumbos

 227 'of each land-cover type' should be replaced (the phrase is repeated)

327 cattail should be Typha ssp.

336 vegetation instead of types

349 species of plants instead of types of plants

385 and 437 please complete references

Zhang, Xiao, et al. "GLC_FCS30: Global land-cover product with fine classification system at 30 m using time-series Landsat imagery." Earth System Science Data 13.6 (2021): 2753-2776.

Comments for author File: Comments.pdf

Author Response

Please see the attached file for our item-by-item response 

Author Response File: Author Response.docx

Reviewer 3 Report

The reviewed article is interesting and well prepared. I recommend that it be published after editorial corrections and responding to methodological comments. I have added my comments to the text of the article in the attached PDF file. Please also pay attention to errors and gaps in the list of references.

Comments for author File: Comments.pdf

Author Response

Please see our item-by-item response in the attached file

Author Response File: Author Response.docx

Reviewer 4 Report

Thank you for this article about classification based on NDVI time-series remote sensing images.

Summary

The document starts by explain knowledge come from wetland vegetation studies and the impact of nitrogen, carbon or methane on environment. Remote sensing images are useful in this context. Then location and sensors are describe. The classification process is drawn figure 2. 100 patches are selected (20 nelumbo, 26 open water, 29 typha, 25 water lily) and 20 in figure 6 (16 nelumbo 4, water lily). The result shows the accuracy of mixed patches are lower than pure patches. Some method details are presented at time (how to consider true positive rate L229). Then discussion section shows the limits of this approach. The conclusion remind the time-series could work around of missing data.

35 references form 1978 to 2020 (20% since 2018, 20% until 1997)

Strengths

Study based on well-known technics

Weakness

Methods and results are sometimes in inappropriate sections.

Comments

L181. The figure 2 presents a method that is not clearly develop in the text.

L226. The result section explain how calculate information. This details must be in the previous section.

L296 “… but will lead to larger errors in pixels of more complex mixture.” The reader may be excepted in discussion some track to raise this kind of limitation.

The goal presented in introduction and recalled in conclusion (impact of nitrogen, carbon or methane on environment) are not highlighted in result or discussion section.

Author Response

Please see our item-by-item response in the attached file

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The manuscript has improved a lot.  However, there are still many problems.

Abstract:  

1. Please add your results.
2. “The plant life form, functional type, …… and greenhouse-gases emission rates.” Does this sentence make any sense in the abstract?
3. Timeseries -> time series or time-series

Introduction:

1. There are already many studies used Sentinel-2 and Landsat based time- series for wetlands plants classification. How your method differs from theirs needs to be explained, or what are the problems within their research？

Materials and Methods

1. Sentinal -> Sentinel
2. “Gaussian Mixture Model” need a reference. Why did you choose Gaussian Mixture Model to do the classification? Why not use others (SVM, RF, KNN)?
3. How many training samples were used for fitting the Gaussian Mixture Model??? I still think that the overall accuracy of 78% is low given only four land cover classes need to be classified.
4. “We used the land-cover type classification result of OWC in 2020 by comparing HLS classification result with the data from a ground-based survey to evaluate the performance of the standards that were developed with 2017 data on classifying wetland vegetation in other years” ---- This is unclear for me, please rephrase it.

Results

Add a sentence to explain the results shown in Figure 3, otherwise it is meaningless to use this picture?

Discussion

1. “It is worth noting that the overall accuracy of WorldView-3 classification, which we use as the reference of true class to assess the HLS classification accuracy, is 78% (Table 1), and therefore, any coarser classification results, such as HLS, with accuracy of above 78% should be considered very accurate.” --- This is not right. High-resolution images do not always yield higher accuracy compare to lower resolution images. Moreover, it might be possible to improve the result by choosing other methods. So, this result shouldn’t be used as a standard.

Conclusion

1. Still miss the result produced by your proposed approach
2. “Land surface models that resolve biogeochemical processes in wetlands improve and can handle increasing resolution and complexity in their representation of ecological and vegetation function within wetlands. Annual classification of wetland patch types data can inform such advance land-surface and ecosystem models. Available remote-sensing derived input on patch type distribution will allow such models to fully leverage their ability to resolve the effects of vegetation types within the wetlands to reduce the uncertainty of the prediction of the ecological function of these wetlands, specifically, nutrient uptake, carbon sequestration, and methane emission rates, under climate change, rising water levels and changes to precipitation and hydrological cycles.” It doesn't make any sense to put it in the conclusion! ! !