Vegetation-Ice-Bare Land Cover Conversion in the Oceanic Glacial Region of Tibet Based on Multiple Machine Learning Classifications
Round 1
Reviewer 1 Report
This study uses machine learning algorithms for the identification of land cover classes in a glacial landscape of Southeast Tibet. High accuracy was reported for all machine learning methods, with back-propagation neural network showing the highest accuracy over CNN and Random Forest. The authors were also able to demonstrate the change in landscape indices between 1990 and 2016 and showed a brief connection to changes in temperature and precipitation data. Some improvements and clarifications are needed within the article. The methodology for the selection of training samples needs to be clarified, it is unclear how spectral indices were used to select different classes and how the training samples were split between four images. Additionally, further comparison to additional literature is needed to place the results in context. Also there are several contradictory statements that need to be corrected throughout the text.
Introduction: The introduction provides a good overview of how glaciers are changing in China, however, there is a disconnect between this and the rest of the article. The article does not focus on glaciers specifically but instead on landscape change (vegetation and bare soil). Some further exploration of the literature on changes in mountain land cover would also be helpful to provide additional context.
Line 97-103: Make sure all of your research questions are numbered for consistency. I would also suggest having the questions in the order that they appear in text, i.e. machine learning first and then landscape patterns.
Figure 1: I would recommend moving this figure to after it is first mentioned on Line 112, below the Study Site section
Lines 122-132: The information given in these lines could also be presented as a Table, I think this would make it more accessible to the reader.
Additionally, why were these four years selected? There needs to be some additional support for the selection of a small sample size. Do these images capture the variation in the region or would it be useful to have a larger time series? Why are having images in June-September ideal? Is this when glacial minimums are observed?
Equation 1. The equation given for NDVI is incorrect, it should be NIR – Red/NIR + Red.
Lines 162-166: It is unclear how the points were divided into the six land cover classes based on the values of the index. This should be addressed, what thresholds/bounds were used for each class? Which indices were used to define each class? For example, how was NDVI used to identify vegetation, what was the minimum NDVI needed for a sample point to be labelled vegetation? What was the accuracy of this method?
Were the random points spread out across all images or were they collected for only one image? The methodology used in this section is unclear. If 20,000 points were collected for each image, it would be useful to include the breakdown per class for each image. If the 20,000 points were collected across the entire image dataset it would be good to see how the points are distributed across the four images (i.e. 5000 points per image).
What was the spectral criteria for deleting points? Was the selection of these pixels done according to observed bias?
Lines 234-245: It would be helpful if the references for these indices were provided. This would help to support the statements made it the text. Alternatively, references should be provided for the equations given in Table 1
Section 4.1: Direct comparisons of accuracy from other research would help to strengthen this section. For the BPNN, are there other examples of the application of similar machine learning methods in the peer-reviewed literature? Same for CNN and RF. Do similar applications of these machine learning methods return the same or similar results? Alternatively, how do these results improve on the classification of these land cover types by other machine learning or non-machine learning classification approaches?
Lines 435-436: Is there other research to support these benefits of BPNN?
Section 4.2: The comparison with meteorology data is good, however, I am concerned about the use of data only 1 month before image acquisition. Is there any influence of different seasons on this area? For example, how do the patterns reported compare if an average of 3 months before the acquisition month were considered? When the text states that averages were calculated one month prior, does the mean 30 days? Or does it mean that if the image was acquired in May that the April average was used? This could dramatically affect the averages.
Additionally, the figure should be presented immediately following Section 4.2. I would also suggest switching colours on the plot so that precipitation is blue and temperature is red.
Connecting back to my previous comment concerning the sequence of images used, a more complete time series would help to show further variation in these meteorology data for this region.
Lines 461-462: This statement is contradictory. “Glaciers respond positively to changes in air temperature but respond negatively to changes in precipitation”, this makes it seem as though when air temperatures increase glaciers also increase but shrink when there is increased precipitation. The next sentence, “The increase in glaciers caused by the increase in annual precipitation is not enough to offset the increased ablation due to rising temperatures.” is more effective.
Line 471-472: This statement is opposite of what these data show, temperature values decrease and precipitation increases. Make sure that these data are correct.
Lines 478-479: The wording, “Vegetation Age”, needs to be clarified. Are the authors referring to the areal coverage by vegetation?
General Discussion Comment: It would be good to see some discussion on the overall performance of the algorithms in the short coming section. What classes were easily confused? For example, were ice and cloud cover well discriminated? What affect does the overall accuracy of the methods have on the results of land cover change?
Wording/Writing Style Comments
Overall some work needs to be done with the writing,
Line 36-40: The word ‘shrinkage’ is used 3 times in two sentences.
Line 52: “temperatures rise lightly…”, slightly?
Line 57-58: “spatiotemporally detected”, the wording is awkward. Consider changing to, “need to be monitored” for example.
Line 94-97: The statements are contradictory, one says that there is insufficient data to perform remote sensing analysis and the next sentence says that remote sensing analysis was performed.
These are some example, ensure the paper is closely checked for sentence structure and contradictory statements.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
General comment:
The paper discussed an environmental issue of glacier decrease and the suitable methodology for the analysis. It is an important and interesting topic for the readers of this journal.
The methodology is straightforward and the results are easily understood. Besides the cross comparison of 3 machine learning techniques, I would be more interested in the science.
In the discussion section, authors could reference some paragraph from the paper mentioned in the introduction section to emphasize the points. It will be more persuasive than just describing how lines go up and down in the plots.
Besides the above, there are few other comments as follows.
Other comment:
Title : From your article, it seems more suitable using "Land Cover Conversion" rather than "Land Interactions".
Figure 2: Label at top right, replace "BP" by "BPNN"
Figure 3: Adding "Cloud Free" to the caption will be more clear
Figure 4: Using color with small marker are distracting when you have two different scale in one plot.
I suggest using different line styles. For lines to left y-axis, use the solid line and dash line for those to right y-axis.
Figure 4: Instead of doing the above, Use 6 plots in two rows to plot 6
different index could be even more clear for the readability.
Figure 6: lack of definition. cannot read some numbers and words.
Figure 7: The horizontal labels are not clear. You can use the abbrevation and "to" make it more clear. e.g. replace "Vegetation-Ice" to "Veg to Ice" and replace "Vegetation-Bare land" to "Veg to Bare", same for others. Explain those abbrevation in the caption is fine.
Figure 8: Should "Monthyly precipitation" be "Monthyly average precipitation" as Ln477?
Table 1 : The self-consistency of labels within the same table. e.g.
1)A in PD and M in AREA_MN represent the same quantity. Why not using the same label.
2)A in PD and A in ED use the same label with different definition. If they represent the same quantity, it is more clear to use the same definition.
3)N in PD, there is no definition of it.
Ln164 : Replace "glaciers" with "ice" to be more consistent with the first line in the same paragraph.
Ln159-Ln168: Please explain more clearly how do you know the class
for those training points, 20000 random points ?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The authors have addressed most of the comments and made sufficient changes to the manuscript. I have some additional comments pertaining to general writing that should also be made.
Line 37: “and the annual percentage”, the ‘and’ can be removed.
Line 63: “particularly” should be changed to particular.
Line 82: “remote sensing classification tasks”, the ‘tasks’ can be removed.
Lines 103-108: The numbering of the questions can be removed, having the questions stated one after another should be okay. The word especially can be removed.
Line 173: This comment was not addressed from my previous review; a table is needed to show how the training/validation points are divided between the different classes. The percentage would be the most useful as it would also help to provide some additional context for the discussion.
Section 3.2: Was an accuracy assessment performed on the interpolated maps? This could be added quickly to showcase how the new results improve over the previous classification.
Section 3.4: The heading on this section should be changed, possibly “Relation between Landcover change and Vegetation” or be combined with section 3.3.
Figure 7: I have some concerns about how results are displayed in this figure. Why are the bin labels different between all the results? 1990-2000 shows vegetation to ice but all other set show ice to vegetation, is this because it is the transition between the different year? Also, the order of the bins should be adjusted. The first one goes vegetation to ice, vegetation to bare, and bare to ice, however, in the other panels bare to veg and ice to bare are switched.
Author Response
Please see the attachment.
Author Response File: Author Response.doc
