Automated Identification of Thermokarst Lakes Using Machine Learning in the Ice-Rich Permafrost Landscape of Central Yakutia (Eastern Siberia)

Round 1

Reviewer 1 Report

Development of lakes in ice-rich permafrost regions exhibits complicated behaviour and is impacted by the whole combination of factors. It seems that the results strongly depends on the exact period of analyzed images. 

I suggest the authors have a look at the results of lake development analysis for the same area published in Russian Earth's Cryosphere. It actually shows that the increase of lakes area is the subject of abrupt change. Interestingly, the attached results do not show the decrease of lake area after 2012.

Minor comments:

I can not see the trend of precipitation at Figure 10. Counting the trend of precipitation from 1880 would not be correct.

Comments for author File: Comments.pdf

Author Response

Development of lakes in ice-rich permafrost regions exhibits complicated behaviour and is impacted by the whole combination of factors. It seems that the results strongly depends on the exact period of analyzed images.

It is possible that image acquisition date affects lake surface area. Ulrich et al. 2017 include images with acquisition times between August and late September. Nesterova et al. 2021 include images with acquisition times between June to September. In our study, there is only one scene from early October (October 3). I believe that our methods are in line with other, similar remote sensing studies. Similar to Nesterova et al. 2021, the scenes available to us for analysis were severely limited by cloud cover. We have highlighted the variability in image acquisition time as a possible shortcoming of this study.

“It is important to reiterate that the image acquisition months range from mid-June to early October (south study site)/September (center study site). While this is consistent with other, similar remote sensing studies of lakes in Central Yakutia [13,58]⁠, it is possible the including a relatively wide range of image months affects the observed lake surface areas. As demonstrated in Hughes-Allen et al. [6]⁠ both lake type and season can have a significant impact on carbon and GHG dynamics.”

I suggest the authors have a look at the results of lake development analysis for the same area published in Russian Earth's Cryosphere. It actually shows that the increase of lakes area is the subject of abrupt change. Interestingly, the attached results do not show the decrease of lake area after 2012.

Thank you for bringing this paper to our attention. A comparison of their results has been added to the discussion. Lines 713-728.

I can not see the trend of precipitation at Figure 10. Counting the trend of precipitation from 1880 would not be correct.

Thank you for bringing this to our attention. The Mann Kendall test has been recalculated for precipitation, temperature, and evapotranspiration. While precipitation shows no trend, temperature and evapotranspiration still show significant increasing trends since 1960.

Reviewer 2 Report

Most of the introduction is devoted to the description of climatic phenomena. This is of course good. But the subject of the journal involves a greater description of new methods of remote sensing and the results obtained using these remote monitoring methods. Therefore, it is proposed to radically remake the introduction with an increase in information about the methods and results of using these remote sensing methods. A significant part of the citations in the introduction relate to the description of climate change. Citations are needed for works with remote sensing results.

The article discusses the use of a well-known method (Convolutional Neural Networks, R-CNN) , which is not new to the readers of this journal.

I believe that the article in this form is not suitable for publication in this journal. Perhaps you should send an article to a Geoscience journal or radically change the content.

Author Response

Most of the introduction is devoted to the description of climatic phenomena. This is of course good. But the subject of the journal involves a greater description of new methods of remote sensing and the results obtained using these remote monitoring methods. Therefore, it is proposed to radically remake the introduction with an increase in information about the methods and results of using these remote sensing methods. A significant part of the citations in the introduction relate to the description of climate change. Citations are needed for works with remote sensing results.

Thank you for this comment. The introduction has been reworked to have a shared focus on climate change and permafrost dynamics, as well as remote sensing and deep learning methodologies. Please refer to lines 120-151.

The article discusses the use of a well-known method (Convolutional Neural Networks, R-CNN), which is not new to the readers of this journal.

While there are about 60 Remote Sensing articles that discuss R-CNN, there is only one other article published in Remote Sensing that discusses the implementation of this technique in permafrost landscapes (automated characterization of Arctic ice-wedge polygons Zhang, W.; Witharana, C.; Liljedahl, A.K.; Kanevskiy, M. Deep Convolutional Neural Networks for Automated Characterization of Arctic Ice-Wedge Polygons in Very High Spatial Resolution Aerial Imagery. Remote Sens. 2018, 10, 1487. https://doi.org/10.3390/rs10091487). There are no articles in Remote Sensing that discuss the implementation of R-CNN to identify waterbodies.

I believe that the article in this form is not suitable for publication in this journal. Perhaps you should send an article to a Geoscience journal or radically change the content.

Thank you for your comment. Based on the other reviewer comments and our improvement of the manuscript thanks to your and the other reviewer comments, we think that our presentation of this interdisciplinary methodology is important. We look forward to any additional comments and suggestions you have.

Reviewer 3 Report

In my opinion the manuscript will have high scientific impact and get new arguments  for researchers in the field of distant methods apply. 

The single feeble place is in the following sentence (p. 9 subchapter 2.3.) :

In the absence of field observations, radiocarbon dating of lake sediments, or geochemical 279 information of lake water, lake type classification was determined solely based on lake 280 morphology.

This way of lake classification is quite insufficient for justification of relationship between lakes types and carbon emission, and  with its genetic and time origin. For clear statement it has to do in situ research for key type of lakes.  

Author Response

In my opinion the manuscript will have high scientific impact and get new arguments for researchers in the field of distant methods apply.

Thank you for your comment.

The single feeble place is in the following sentence (p. 9 subchapter 2.3.):

In the absence of field observations, radiocarbon dating of lake sediments, or geochemical 279 information of lake water, lake type classification was determined solely based on lake 280 morphology.

This way of lake classification is quite insufficient for justification of relationship between lakes types and carbon emission, and with its genetic and time origin. For clear statement it has to do in situ research for key type of lakes.

Thank you for brining this to our attention. This paragraph has been clarified as follows: The thermokarst lakes in this region are divided into three categories based on field observations, past radiocarbon dating of lake sediments, geochemical signatures of lake waters, morphology, and a multiple-stage development model [16,36]⁠. An illustrative example from the area is presented in Figure 2. Lakes of each type have strong differences in lake physiochemistry, dissolved GHG concentration and flux. The characteristic morphology of each lake type was determined from lakes with in-situ measurements and then applied to lakes in the remote sensing images. Please refer to lines 225-233.

Reviewer 4 Report

General Comments

I read the manuscript with great interest. The formation of lakes associated with thermokarst development is an essential research subject for quantitatively assessing permafrost thawing and subsequent greenhouse gas emissions in the subarctic and boreal forests of Eastern Siberia. This manuscript can be judged as a valuable study that provides new insights into the characteristics of water bodies in permafrost regions by developing and validating a deep learning method for classifying water bodies over a wide area. It is consistent with the scope of Remote Sensing and should be published. However, a few points need to be corrected regarding the discussion of the analysis results. There are also some minor comments in the description, which should be revised appropriately and resubmitted.

1) Can the authors evaluate the detection results using elevation data to distinguish between the three types of lakes? The CA and UCA lakes should be established on alas bottom. These must have significantly lower water surfaces than the elevations of the two terrace surfaces and RT lakes. Can a free, high-resolution Arctic DEM (or similar dataset) be used to discuss the relationship with the topography, even if only in some of the identified areas?

2) When the RT lakes begin to become water bodies along the troughs of high center polygons caused by thermokarst, I believe that the shape of the lakes will be reticulate rather than circular at the initial stage. Since such early thermokarst lakes seem to be outside this study's scope, it underestimates the actual expansion of water bodies due to thermokarst. How many areas in these two areas of analysis show early thermokarst development without RT lakes?

3) I wonder if the growth/decline of the area of alas lakes in the UCA and CA types has anything to do with the expansion of artificial channels. For historical context, I think you should refer to the paper by Crate et al. (2017). Has there been any artificial lake expansion, especially in the alases along settlements and rivers (which effectively channel snowmelt water into the lakes)? The Crate et al. paper also includes a figure of long-term changes in water levels in Tyungyulu Alas, the largest in the region (Fig. 8 of Crate et al. paper), which may be helpful when discussing lake area fluctuations and weather conditions.

Crate, S., M. Ulrich, J. O. Habeck, A. R. Desyatkin, R. V. Desyatkin, A. N. Fedorov, T. Hiyama, Y. Iijima, S. Ksenofontov, C. Mészáros, H. Takakura, 2017. Permafrost livelihoods: A transdisciplinary review and analysis of thermokarst-based systems of indigenous land use. Anthropocene, 18, 89–104. DOI:10.1016/j.ancene.2017.06.001

Specific Comments

Abstract and line 199: I think the number of the alas is supposed to be about 16,000, not thousands; see Desyatkin et al. (2009)[16] and Crate et al. (2017) descriptions. 

Lines 231 and 669: Is there really a link between forest fire history and the occurrence and expansion of thermokarst lakes? Please indicate if there are any fire history sites in the study area where thermokarst (and RT lakes) has developed.

Line 679-690: The variation of Lake radius change (Fig. 5) by Ulrich et al. (2017) [13] is an important comparison for time series discussion. Is their result consistent with the expansion and contraction of lakes in CA and RT? I think there is insufficient discussion.

Citation [45]: Is this citation PANGAEA? No details are provided in the list.

Author Response

1) Can the authors evaluate the detection results using elevation data to distinguish between the three types of lakes? The CA and UCA lakes should be established on alas bottom. These must have significantly lower water surfaces than the elevations of the two terrace surfaces and RT lakes. Can a free, high-resolution Arctic DEM (or similar dataset) be used to discuss the relationship with the topography, even if only in some of the identified areas?

Thank you for this interesting suggestion. It is likely that inclusion of a high resolution DEM could improve classification of the three lake types. However, the ArcticDEM is based on photogrammetry of visible high resolution satellite images, so vegetation, particularly forest, is not removed from the DEM. The lake water level is not well represented and the bottom of the lakes are not represented at all as would be the case with radar data, for example. Unfortunately, we would also need to use a DEM whose date corresponds to the satellite image date used in this study, which is not possible for the earlier images. For reference, plesae find attached an example of some lake polygons overlaying the 2013 ArcticDEM.

2) When the RT lakes begin to become water bodies along the troughs of high center polygons caused by thermokarst, I believe that the shape of the lakes will be reticulate rather than circular at the initial stage. Since such early thermokarst lakes seem to be outside this study's scope, it underestimates the actual expansion of water bodies due to thermokarst. How many areas in these two areas of analysis show early thermokarst development without RT lakes?

Thank you for this interesting suggestion. Unfortunately, the resolution of our images (and the single band nature of the images) is not good enough to accurately identify these nascent lakes. We do see some very small thermokarst lakes appear and disappear between scenes (<5 per scene). These very small lakes might represent a stage between ponding between high centered polygons and true lake formation.

3) I wonder if the growth/decline of the area of alas lakes in the UCA and CA types has anything to do with the expansion of artificial channels. For historical context, I think you should refer to the paper by Crate et al. (2017). Has there been any artificial lake expansion, especially in the alases along settlements and rivers (which effectively channel snowmelt water into the lakes)? The Crate et al. paper also includes a figure of long-term changes in water levels in Tyungyulu Alas, the largest in the region (Fig. 8 of Crate et al. paper), which may be helpful when discussing lake area fluctuations and weather conditions.

Thank you for bringing this paper to our attention. Discussion of the results of the Crate et al. paper has been added to the discussion section. This paper has been added as a citation to the discussion of the water level of some lakes potentially being controlled by local populations for agricultural activities.

Without additional specific information about community use of alas water, I don’t know if we can say more here about human impacts on lake water level.

Abstract and line 199: I think the number of the alas is supposed to be about 16,000, not thousands; see Desyatkin et al. (2009)[16] and Crate et al. (2017) descriptions.

These numbers have been updated.

Lines 231 and 669: Is there really a link between forest fire history and the occurrence and expansion of thermokarst lakes? Please indicate if there are any fire history sites in the study area where thermokarst (and RT lakes) has developed.

An additional citation has been added at line 231.

Line 736 has been updated as follows and relevant citations added: “While forest fire events are not apparent in the scenes analyzed in this study, several recent studies indicate the link between forest fire and thermokarst lake formation.”

Line 679-690: The variation of Lake radius change (Fig. 5) by Ulrich et al. (2017) [13] is an important comparison for time series discussion. Is their result consistent with the expansion and contraction of lakes in CA and RT? I think there is insufficient discussion.

Thank you for this comment. In the Ulrich et al. 2017 paper, they track radius change for specific lakes in their study area (n= between 4-10 lakes). We have approximately 1,000 lakes per study site, making it impossible to conduct the same analysis for all of our lakes. We could, however, do this analysis for a subset of lakes in each study site. Many of the lakes (especially the UCA lakes) are half moon or banana shaped. How well will the radius change analysis as done in the Ulrich paper account for this type of morphology and/or morphology changes (ex. A half moon shaped lake infills to be more circular? Some UCA lakes are hula hoop shaped with a circular island of dry ground in the middle. Can the radius change analysis be applied to these lakes? We could consider simply not including lakes with these irregular morphologies, but will that bias our results? Including a discussion of the rate of lake radius change would be very interesting. Please let us know if you have suggestions about how to apply these methods to our study lakes.

Citation [45]: Is this citation PANGAEA? No details are provided in the list.

Thank you for bringing this to our attention. The citation has been updated.

Author Response File: Author Response.pdf

Reviewer 5 Report

This article is presenting results of lake change detection analyses that cover three different thermokarst lake types in Central Yakutia. The authors describe a semi-automated machine learning approach on Spot and Corona/Hexagon satellite data using Mask Region-Based Convolutional Neural Networks (R-CNN). Lake area changes were detected and quantified within two adjacent study areas (each about 1200 km²), which are located at the middle reaches of the Lena River closed to the Yakutian capital. In accordance to own and other previous studies, the lake types were manually classified into unconnected alas lakes, connected alas lakes, and recent thermokarst lakes. Individual and total lake area changes during seven timesteps within the last about 60 years have been discussed in relation to changes in air temperature, precipitation and evapotranspiration.

I read the article with great interest and found it generally very well written. I also think that this is an important topic, especially because the carbon source versus sink function of various permafrost-related lakes as well as their water balance has still not been fully clarified and there are still many open questions.

However, I am not clear to what extent the presented article offers new insights, since the discussions on the three influencing climatic parameters (i.e. “weather variables”) can at best only be weakly supported by the "simple" statistical analyzes and the purely qualitative physical interpretations carried out here.

However, the methodological approach offers some insights, but I miss a more in-depth discussion of the ML approach and the corresponding data basis. A methodical flowchart would also be helpful to better understand the various technical steps. A few questions remained unanswered for me, such as; Why are only panchromatic data used, even for the South side, for which there seems no Corona data were available? Especially for the problems of detecting fuzzy lake borders, the multispectral Spot data would be much better, as you know. Would a pan-sharpening also be conceivable? On the other hand, why were only SPOT data used to train the model that runs on Corona/Hexagon, too? Even if the seasonal changes of the lakes are difficult to detect with the available data, it would still have to be clarified what these differences mean for the methodology used (South side image cover from June to October) and ultimately also for the annual water balance and the seasonal differing carbon fluxes of the lakes, which was especially studied (during previous work) and mentioned here by the authors.

However, I think that this article should be published at the end, if a methodological discussion could be sharpened and further enhancements, perspectives, and consequences could be covered and discussed.

Some minor comments and suggestions are:

L134: Please check the size information of your study sides as they differ in different places from 4,500km² in total (this Line) to 1200km² each (e.g. Abstract) to 40x40 km² each (L405) to more specific area information (L419 + L428 and Table 2).

L162: Please change to: “Permafrost in Central Yakutia is generally continuous…”, because on the other hand it might be somewhat misleading with the “Eurasia” reference immediately above.

L225/226: Please explain; the lakes are larger and deeper than what? Looking to the map in Figure 2, this is clearly not the case in comparison to the Unconnected Alas Lake.

L428: There are seven scenes listed in Table 2. Please clarify.

L438: I suggest to add also the affiliation of A. Fedorov here.

L499 – 513: In my opinion, the strength of the clustering of the various lake types is not apparent from the z-scores of the NN analyzes given here. Why do the RT lakes show the strongest clustering even though the z-score, at least for the South side, is higher than for the UCA lakes? Otherwise, the CA lakes exhibit spatial clustering, similar to the UCA lakes, but with the highest z-scores. Please clarify.

L618: What other variables are meant here? Please continue or delete the end of this sentence.

Figure 1: Please mention, what are the green areas in the OSM basemap. Please change in Line186 to “…location is outlined…” and in Line 191 “The Lena River runs south to north. The Aldan River runs east to west.”

Figure 2: I would suggest adding a brief note here that the pictures are exemplary but do not correspond to the lakes on the map. Or are they the same lakes but at different times?

The axes and legend labels of many figures are too small. In the figures resampling the gray value images, details are very difficult to see.

Author Response

A methodical flowchart would also be helpful to better understand the various technical steps.

A methodical flowchart has been added to the manuscript in the methods section (Figure 6).

Especially for the problems of detecting fuzzy lake borders, the multispectral Spot data would be much better, as you know.

Agreed, multispectral data would have been much better. However, we only had access to single band SPOT images, except for 2016 and 2019, where we had both single band and multispectral images. We used the single band images for analysis for all scenes to ensure consensus in the methods.

Would a pan-sharpening also be conceivable?

The use of pan-sharpened images is certainly an interesting idea and the use of pan-sharpened images would likely improve overall lake identification. However, we are limited in our availability of cloud free images (color or black and white). Color images of any quality are not available for the earlier scenes (to our knowledge) and we would therefore not be able to apply the same methods of analysis to all scenes. Please let us know if you know of available datasets that we might have missed!

On the other hand, why were only SPOT data used to train the model that runs on Corona/Hexagon, too?

Thanks for this comment. Yes, the model was only trained on SPOT data, this was in part due to the fact that the Corona and Hexagon images had artifacts from the image digitization process that made them less than ideal to use for the training process (ex. Scratches, damage, dark marks, etc.). After training, the model predicted lakes in the Corona and Hexagon images with similar effectiveness as for the SPOT images. If lakes were predicted where artifacts existed, they were very easy to manually correct during the manual lake boundary correction process.

Even if the seasonal changes of the lakes are difficult to detect with the available data, it would still have to be clarified what these differences mean for the methodology used (South side image cover from June to October) and ultimately also for the annual water balance and the seasonal differing carbon fluxes of the lakes, which was especially studied (during previous work) and mentioned here by the authors.

It is possible that image acquisition date affects lake surface area. Ulrich et al. 2017 include images with acquisition times between August and late September. Nesterova et al. 2021 include images with acquisition times between June to September. In our study, there is only one scene from early October (October 3). I believe that our methods are in line with other, similar remote sensing studies. Similar to Nesterova et al. 2021, the scenes available to us for analysis were severely limited by cloud cover. We have highlighted that variability in image acquisition time as a possible shortcoming of this study. Please refer to lines 694-699.

“It is important to reiterate that the image acquisition months range from mid-June to early October (south study site)/September (center study site). While this is consistent with other, similar remote sensing studies of lakes in Central Yakutia [13,58]⁠, it is possible the including a relatively wide range of image months affects the observed lake surface areas. As demonstrated in Hughes-Allen et al. [6]⁠ both lake type and season can have a significant impact on carbon and GHG dynamics.”

However, I think that this article should be published at the end, if a methodological discussion could be sharpened and further enhancements, perspectives, and consequences could be covered and discussed.

Thank you for your comments and suggestions. We have updated the manuscript and look forward to your additional insights.

L134: Please check the size information of your study sides as they differ in different places from 4,500km² in total (this Line) to 1200km² each (e.g. Abstract) to 40x40 km² each (L405) to more specific area information (L419 + L428 and Table 2).

Thank you for pointing this out. The numbers in the manuscript have been rectified.

L162: Please change to: “Permafrost in Central Yakutia is generally continuous…”, because on the other hand it might be somewhat misleading with the “Eurasia” reference immediately above.

The correction has been made as suggested.

L225/226: Please explain; the lakes are larger and deeper than what? Looking to the map in Figure 2, this is clearly not the case in comparison to the Unconnected Alas Lake.

The sentence has been simplified to: “These lakes are consistently several hundreds of meters across and up to ~ 10 meters deep.” Indeed, in the old version of Figure 2, there was an exceptionally large unconnected alas lake shown. This large UCA lake is an outlier and not representative of the overall trend. Figure 2 has been updated to be more representative of lake size trends.

L428: There are seven scenes listed in Table 2. Please clarify.

Thank you for catching this typo. The correct number is seven.

L438: I suggest to add also the affiliation of A. Fedorov here.

A. Fedorov’s affiliation has been added.

L499 – 513: In my opinion, the strength of the clustering of the various lake types is not apparent from the z-scores of the NN analyzes given here. Why do the RT lakes show the strongest clustering even though the z-score, at least for the South side, is higher than for the UCA lakes? Otherwise, the CA lakes exhibit spatial clustering, similar to the UCA lakes, but with the highest z-scores. Please clarify.

Thank you for bringing this to our attention. The strength of spatial clustering is indicated by the nearest neighbor index. This value didn’t make it into the first manuscript version and has been added to the appropriate sections. The z-score indicates level of confidence. This has been clarified in the text. Please refer to lines 528-535.

L618: What other variables are meant here? Please continue or delete the end of this sentence.

This sentence has been updated: “The observed trend of decreasing UCA lake surface area compared to increasing RT lake surface area and stable CA lake surface area is likely related to differences in lake morphology and related dynamics between the three lake types and their responses to changes in temperature, precipitation, evapotranspiration, and possible other variables that are beyond the scope of this study.”

Figure 1: Please mention, what are the green areas in the OSM basemap.

Thanks for this comment. An explanation has been added to the figure caption.

Please change in Line186 to “…location is outlined…” and in Line 191 “The Lena River runs south to north. The Aldan River runs east to west.”

The figure caption has been updated as suggested.

Figure 2: I would suggest adding a brief note here that the pictures are exemplary but do not correspond to the lakes on the map. Or are they the same lakes but at different times?

Thanks for this suggestion. Figure 2 has been reworked to be more clear.

The axes and legend labels of many figures are too small. In the figures resampling the gray value images, details are very difficult to see.

Axes and legend labels have been improved for readability. I have tried to improve the readability of the gray value images and will discuss further with the journal publication department prior to publishing.

The supplementary images are so large that it’s very difficult to make readable images with scalable vector outlines of the lake polygons (I have not been successful at least). We will make the lake polygon shapefiles available in conjunction with the publishing of this article so that readers can more closely examine our results.

Round 2

Reviewer 2 Report

1. Despite the addition (very small), the introduction still does not correspond to the subject of this journal. A small addition in the introduction is clearly not enough to describe the state of the remote sensing issue. It is necessary to significantly expand the description of remote sensing methods used for the study of thermokarst lakes.

2. A new figure 6 has appeared, this figure is really needed. But it is poorly understood and little described in the text. It is necessary to make this drawing more understandable. The description should also be expanded.