A New Automatic Extraction Method for Glaciers on the Tibetan Plateau under Clouds, Shadows and Snow Cover

Round 1

Reviewer 1 Report

Thank you for adding these last statistics! This really helps to improve the MS.

Author Response

Thank you again for your helpful comments. This is of great help to our manuscript. We checked the grammar of the full text again, hoping that these changes can promote the publication of the manuscript.

Reviewer 2 Report

The manuscript entitled “A new automatic extraction method for glaciers on the Tibetan Plateau under clouds, shadows and snow cover” submitted to Remote Sensing.

Please revise the manuscript for minor grammar mistakes. There are quite many and some of them below:                          

Line 17: a reflectivity different index

Line 22: the two classifications? The new text was placed in a wrong location?

Line 23: Kappa coefficient were 0.92 and 0.83 respectively?

Line 25: and

Line 26: remove “obviously”

Keywords: the authors can remove “algorithm” and “the”

Line 83: “so on”?

Lines 85-93: lack of the main verb

Line 169: which is the

Line 170: full stop and “However, the traditional …”

Line 176: never used “So” at the beginning of the sentence.

Figure 4: the distance scale should be in “Km”.

Line 216: data were

Author Response

请参阅附件。

Author Response File: Author Response.docx

Reviewer 3 Report

The authors describe the proposed method of glaciers borders delimitation. This method was applied to the glaciers on the Tibetan Plateau and showed good results. All steps of the work are described very detailed and well-illustrated.

I have only 2 questions:

1. Paragraph 4.1.2, line 318. It is not clear the criteria, which show that the classification accuracy did not improve significantly. Especially for the second classification as 25 is a maximum number of the features. Figure 8b shows that after 8 or 9 features the increase of accuracy is not high.

2. Paragraph 4.1.3, line 325. The authors mentioned the number of sample points. But points cannot have such characteristics as texture. It must be described in more details which objects (points or polygons) were used for RF classification.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

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

Comments on Hu et al. – Remote Sensing
The manuscript by Hu et al. presents a new double random forest method for mapping glaciers on the
Tibetan Plateau. The authors first present their methodology, and then apply it over a large region.
While the double-RF method is interesting, I am not convinced that it represents a very novel result. In
essence, the authors (1) use Google Earth Engine to create composite images, (2) train two random
forest models based on some hand-clicked points, and (3) combine the results of the two RF outputs
into a single glacier map. While I think that this approach is in principle a good one – mapping both
glaciers and debris-tongues at the same time is difficult – I am not sure that a whole paper needs to be
written about it. Efforts to map the clean parts of glaciers are well-established, and can be done
accurately with much simpler methods (for example, band ratios). Using a long-term composite band
ratio with a simple threshold (or Otsu thresholding) is a quick and efficient way to accomplish the first
half of their work (e.g., the first RF model). A comparison of such an approach to the RF approach would
be necessary to show that their more complex method is really adding something to the discussion.
The case for using an RF classifier for debris-covered glaciers is stronger – indeed this has been done
before in the literature (for example, https://www.mdpi.com/2072-4292/4/10/3078). To show that their
new method is ‘novel’, it would be important to much more fully assess their training data, and to
compare their results to previous work. I also find the training dataset to be much too small to map
glaciers over such a large and diverse area.
Finally, there is very little comparison to the overall glacier area mapped – we see statistics about
comparisons to the training dataset, but not to how much of the glacier inventory is well-mapped by this
method. In the few figures where such overlaps are shown, the results are not convincing.
I have included other specific comments below.
Specific Comments
Introduction – Missing some references on Debris-Covered glacier mapping (e.g.,
https://www.researchgate.net/profile/Manfred-
Buchroithner/publication/228753609_Automated_delineation_of_debriscovered_
glaciers_based_on_ASTER_data/links/0deec51ab92fe6894a000000/Automated-delineation-ofdebris-
covered-glaciers-based-on-ASTER-data.pdf,
https://ieeexplore.ieee.org/abstract/document/1025770,
https://tc.copernicus.org/articles/9/1747/2015/)
Line 157 – What was the chosen cloud threshold? How was this balanced? What statistics were used?
Line 199 – How were these samples chosen? How many samples for each type? Are these end-members
representative across all glaciers in the Tibetan Plateau? I am sure that the debris spectra is different
between the Tien Shan and Karakoram, for example. As the whole premise of the study is based upon
these training samples, much more detail about how, on which images, who, etc collected this training
data is required so that their validity can be assessed.
Line 233 – How big of a temperature difference is there between debris-covered glaciers and nearby glaciers? During much of the year, they will be very similar temperatures.
Figure 5 – I don’t see how a lower cloud threshold (40) could result in less pixels than a higher cloud score (60). Wouldn’t 40 mean that there are more available images? Or did this somehow increase misclassified pixels? Further explanation is needed here.
Line 355 – Were these sample points polygons or single pixels? How were they chosen? Their distribution on a map would be good to see. Furthermore, what are the ‘Others’ that are such a large part of the sample? Why is this class so much larger than the target (glacier) classes?
Figure 9 – Is this really a 99.7% accurate classification? I see MANY missing pixels here – both on debris tongues, as well as extra pixels outside of the red glacier outlines. Are the accuracy statistics only related to your hand-chosen training samples, rather than as compared to the large-scale glacier inventory data?
Section 4.3 – What is added here that is not already available via the large-scale glacier inventory or the RGI? The same spatial/elevation/area distributions can be gotten from that dataset as well.
Figure 13 – Why are there no glaciers pointing due north (0 degrees)? It looks like there is an issue with how the data is binned if there are no glaciers in one direction.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The presented paper is devoted to an important part of remote sensing methodology - delineating the glaciers, especially in the problem areas, covered by debris, shadow, snow or clouds. The novelty of this approach is the use of open access resources of Google Earth Engine and double-RF output. Actually, there are glacier outlines for Tibetan Plateau, which were used to verify the results, so, it is more about methodology, than mapping new territories. The presented results show high agreement with existing datasets (if calculated correctly) , which is good. The regional issues are not described in detail, which makes it a bit difficult to understand the real advantage of the proposed method to using on unstudied regions. Would be useful to add more detail about combining automated patterns to final contours and accuracy calculation. 

Some small corrections are:  Line 39 - better use ice-accumulating instead of ice-forming

Figure 13 - only two types are shown, please correct the title.

Being not state of the art, this work is an example of possibilities of open access resources in remote sensing applications. One could suggest shortening of the text and reducing the number of figures, focusing more on the main goal - automatic delineation of the problematic glacier parts. 

 

Author Response

请参阅附件。

Author Response File: Author Response.docx

Reviewer 3 Report

The manuscript entitled “A new automatic extraction method for glaciers on the Tibetan Plateau under clouds, shadows and snow cover” submitted to Remote Sensing. The authors aim to (1) propose remote sensing indexes for identifying glaciers in complex environment, (2) proposal an automatic extraction method of glaciers based on RF and GEE, and (3) compare the proposed method with other common methods. In general, the technical details were well explained, especially the section about feature selections, and I have no concern. The study may be considered for publication, but the authors have to improve the quality of writing:

1. Please revise the manuscript more carefully for grammar mistakes, especially articles. There are long sentences and should be separated into two-three sentences. The meaning of some sentences is not clear (maybe a professional proofreading service will be helpful?).
2. The introduction should be revised, especially the second and third paragraphs; the authors do not clearly show the gaps in literature. The 2^nd paragraph is too long and does not provide a clear message. The third paragraph (Google Earth Engine) is too general. If the authors would like to introduce GEE, it should be in the methodology. In the introduction, the authors may present who has used GEE, how do they use GEE for glacier classification, and what are the limitations?
3. The authors rather introduced a methodology than the case study of the Tibetan Plateau, so they should modify the text accordingly.
4. The authors can add a table showing data used in this study and their references and can reduce the number of words for Data sources in Section 2.2.
5. It is not very clear to me the relationship between “Dataset screening” and “obtain more available images”.
6. Where do data used to create Figure 4 come from? If it is from this study, how many sampling points were taken?
7. “Table 1 presents the 25 features constructed in this study” should be in a separate line.
8. Please remove the border line in figures (e.g., Figure 5).
9. Table 2 is not necessary. Also, figure 12a is not necessary. Data in Figure 12a are slightly different from Figure 12b.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments on Hu et al. – Remote Sensing Firstly, I want to thank the authors for taking the time to revise the paper. I appreciate the extra figures, as they have helped make the work clearer. However, some of my original concerns remain. Mainly: Finally, there is very little comparison to the overall glacier area mapped – we see statistics about comparisons to the training dataset, but not to how much of the glacier inventory is well-mapped by this method. The presented accuracy statistics refer only to a small manual sample – not to how well the algorithm performs over the whole study area. No statistics about the total area well-mapped/mis-mapped is given (e.g., if the red polygon outlines of the glacier inventory sum up to 100,000 sq km, how many sq km is mapped by your algorithm? How much extra that is not in the glacier inventory is mapped?). This is crucial information – we can’t interpret the accuracy of a classification at 30m over the entire Tibetan Plateau based on a few hundred training points. There are millions of pixels potentially mapped or mismapped – that would be a much better way to know how well the algorithm performs. In principle the double RF method is good – but without knowing how well it actually performed over the entire area, it is difficult to fully evaluate this study. Specific Issues Abstract – the given percentages for overall accuracy seem incredibly high. This does not agree with the images (Figure 10) which still show significant white area which has not been captured by the algorithm. For a nearly 100% accurate algorithm, I would expect that the full glacier outlines are filled in by the algorithm! If this accuracy just refers to the training data, then this is misleading – well-chosen training data should of course always be captured by the algorithm. What would be more useful for a statistic here is how much of the glacier area is captured by the algorithm, and how much extra non-glacier area is also captured (so, with regards to the red polygons in Figure 10, not the training data). Line 39 – Extra ‘The’ can be removed as first word. Line 121 – Could you not use all images since you construct a minimum NDSI? The annual minimum NDSI could also help with the varying seasons around the Tibetan Plateau. This also might further help with missing data due to cloud cover. Line 307 – What is the total area of those sample points? They aren’t just single points, but rather polygons. It would also be important to know how much area is covered, so what the relative number of pixels used to train each class was, not just the number of locations used. I assume not all training areas are the same size and shape, from the figures in the paper. Line 316 – You should defend this statement. How much more accurate? In what way was it more accurate? Some explanation is needed here. Tables 4/5 – Why are some classes missing from the verification? I would expect an even distribution of all classes to remain in the validation data – otherwise you are only checking how well you did in a subset of the classes, not across all land cover types. Figure 11 – I’m a bit confused how you got the accuracy statistics for the other methods. What data are you comparing the band ratio approach to here? I would be very surprised if the simple band threshold approach got you close to the same accuracy over all areas (including debris cover) as your double RF method did. The total glacier area mapped would also be very different (Figure 15). What is the accuracy/kappa referring to here? Did you use your same training points and a supervised classifier? Figure 13 – Why are the bars offset on panel (b)? The axis is also misleading – those aren’t negative elevations, but rather a different way to visualize panel (a). Line 420 – Using an annual (instead of summer-only) minimum NDSI composite might help with shadows as well – the illumination angle of glaciers will change throughout the year, so might pull some glaciers out of the shadow in some Landsat images. Line 469 – This discussion of errors with the training dataset should be in the Methods, and more details should be given about how the samples were chosen. For example, were all samples chosen by one person? Did multiple people check the polygon outlines to see if they agree on the classification? The whole study is based on these training data, so this should really be made clear. 

Reviewer 3 Report

The manuscript entitled “An new automatic extraction method for glaciers on the Tibetan Plateau under clouds, shadows and snow cover” submitted to Remote Sensing journal. The authors handled well my comments from the previous round.

In the revised version, please cross out/remove redundant texts; in the current form, the authors only highlight new adding texts, which made readers difficult to track changes.

Please check again for minor grammar mistakes

Line 134: include

Figure 4: do not fill the polygon

Line 590: dimension disaster?