the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 29 Feb 2024
Automated snow cover detection on mountain glaciers using space-borne imagery
Abstract. Tracking the extent of seasonal snow on glaciers over time is critical for assessing glacier vulnerability and the response of glacierized watersheds to climate change. Existing snow cover products do not reliably distinguish seasonal snow from glacier ice and firn, preventing their use for glacier snow cover detection. Despite previous efforts to classify glacier surface facies on local scales, a unified approach for monitoring glacier snow cover on larger spatial scales remains elusive. We present an automated snow detection workflow for mountain glaciers using supervised machine learning-based image classifiers and Landsat 8/9, Sentinel-2, and PlanetScope satellite imagery. We develop the image classifiers by testing numerous machine learning algorithms with training and validation data from the U.S. Geological Survey Benchmark Glaciers. The workflow produces daily to biweekly time series of several glacier mass balance and snowmelt indicators (snow-covered area, accumulation area ratio, and seasonal snowline) from 2013 to present. Workflow performance is assessed by comparing automatically classified images and snowlines to manual interpretations at each glacier site. The image classifiers exhibit overall accuracies of 92–98 %, Kappa scores of 84–96 %, and F-scores of 93–98 % for all image products. The median difference between automatically and manually delineated median snowline altitudes, along with the interquartile range, averages 27 +/- 79 m across all image products. The Sentinel-2 classifier (Support Vector Machine) produces the most accurate glacier mass balance and snowmelt indicators and distinguishes snow from ice and firn the most reliably. Yet, the Landsat- and PlanetScope-derived estimates greatly enhance the temporal coverage and frequency of observations. Additionally, the transient accumulation area ratio produces the least noisy time series, providing the most reliable indicator for characterizing seasonal snow trends. The temporally detailed accumulation area ratio time series reveal that the timing of minimum snow cover conditions varies by up to a month between Arctic (63° N) and mid-latitude (48° N) sites, underscoring the potential for bias when estimating glacier minimum snow cover conditions from a single late-summer image. Widespread application of our automated snow detection workflow has the potential to improve regional assessments of glacier mass balance, water resources, and the impacts of climate change on snow cover across broad spatial scales.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(7191 KB)
-
Supplement
(2693 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(7191 KB) - Metadata XML
-
Supplement
(2693 KB) - BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-548', Anonymous Referee #1, 22 Apr 2024
Please see attached pdf file for review comments.
- AC1: 'Final response', Rainey Aberle, 09 Jul 2024
-
RC2: 'Comment on egusphere-2024-548', Anonymous Referee #2, 02 Jul 2024
Summary:
The study by Aberle et al. presents an analysis of using machine learning approaches to classify glacier surfaces facies and to track the spatial and temporal patterns of snow cover on mountain glaciers from satellite imagery. The authors test nine machine learning algorithms on a robust training and validation dataset from five USGS Benchmark Glaciers, using Landsat, Sentinel-2, and PlanetScope imagery. They find that the nearest neighbors and support vector machine models provide the best classification results, and use these models to classify 3000+ individual images of the study glaciers. They then derive daily to bi-weekly timeseries of snow-covered area, accumulation area ratio, and snow line altitude/locations. The study would be a welcome and novel addition to the growing research on remotely-sensed snow cover and glacier dynamics, but I recommend that revisions are made to improve the manuscript quality and to better represent the derived datasets.
Major comments:
- The introduction does a good job of presenting the benefits of and need for improved snow detection methods. However, it should be expanded to include greater discussion and acknowledgement of the existing body of research that has focused on snow identification on glacier surfaces.
- Individual portions of the methods section are well written, but I suggest reorganizing it to make the entire story easier to follow. Specifically: 1) the study area should be moved to its own section, rather than included in the methods, and 2) Sections 2.2-2.4 could be reorganized to follow the structure of Figure 4 to make it easier for the reader to follow. For example, large parts of imagery selection and pre-processing are currently included in Section 2.4 but may fit better in section 2.2, and the seasonal snowline identification could be broken out into its own section.
- The development of reproduceable and extendable code/methods is an important aspect of this project, however I find that the areas of the manuscript where the authors discuss details of the code to be confusing and distracting (e.g. lines 226-227, 234-235). I would encourage the authors to consider whether these details are better suited to be included in the supplemental information or as details on the github page.
- The authors have put considerable effort into developing the methods and creating a thorough dataset. However, I think that more space should be used to present details of the derived products. For example, Figure S3 contains many useful insights that would be better suited for the main manuscript. Specifically, it highlights the differences in temporal resolution between the different imagery products, and well as how consistent (or inconsistent in the case of PlanetScope imagery) the derived products are. Highlighting these results in a main-text figure would improve the presentation of the findings (or perhaps a subset of this figure, such as only a single glacier, or only a subset of years, such that the details of the plot are more easily seen). Other questions which are raised in this figure and throughout the manuscript which could be elaborated on include: are you able to identify significant interannual-variability in the glacier snowline elevation and AAR from these products? How would the results compare when using only a single imagery source, rather than a blend of all imagery as you have done here?
Minor comments:
Line 68: has -> have
Line 69: I would suggest rephrasing “images with spatial resolutions of 1 km or more” to remove the specific number, as most commonly-used satellite imagery is finer spatial resolution.
Line 92: I found that these two points (particularly point 1) were difficult to read. You might consider simplifying or restructuring the sentence here.
Line 109: It should be clarified that the manually generated snow cover observation were made from satellite imagery, rather than from in situ observations.
Line 132: It was a bit confusing to see Emmons Glacier included in this figure immediately after the study area section, where it was not mentioned. Perhaps the details on how it is used should be included earlier in the manuscript to avoid this confusion.
Line 147: The reference to Figure 3 should be to Figure S1, I believe.
Line 204: The inclusion of nine separate ML models is impressive and thorough. Additional information should be included for each (likely in the supplement, I would think) on the specific hyperparameters used for each.
Line 252: How are the masked areas treated in the process of making these histogram? Are the masked pixels included in the glacier elevation bin histogram?
Line 257: What is included in the no-data mask here? Is it only cloudy pixels? Cloudy pixels and off-glacier areas?
Line 254-255: I worry that this may cause a consistent negative bias in the snowline altitudes which are derived. Was a similar approach used to remove sparse snow patches at low elevations to ensure that these snow-ice boundaries were not included?
Line 340: typo for “instils”
Line 343: How are the differences in timing of the manual vs automated snowlines treated in this comparison? What is the range of differences?
Line 343-352: I don’t think the +/- symbol should be used for the IQR numbers here. Including the actual min/max of the IQR would be a more useful metric. eg “… differ from manually delineated snowlines by a median of 116 m (IQR 20–259 m) in ground distance …”
Line 344: including a figure (scatterplot) showing the relationship between automated vs manually-delineated snowline altitude would be a useful addition to highlight the accuracy of the automated methods.
Line 370: Is “the ranges in transient AARs are much larger (0.1–0.3)” referring to interannual ranges in AAR, intra-annual, or range amongst the glaciers?
Line 425: I believe this is the first time that cloud shadows are discussed. The methods should be more explicit that cloud shadows are identified and removed from the imagery.
Lines 440 & 455: I was initially a little confused by these statements on more heavily weighting Sentinel-2 SR observations. I would suggest rephrasing these to make it more clear that the suggestion is for when observations from multiple sources are being synthesized.
Supplement Line 72: typo (repeated words)
Figure 1: A note should be made in the caption (and/or an asterisk added) to acknowledge that the NDSI bands indicated for PlanetScope imagery is not the typically-used SWIR band.
Figure 2: I find that having only the elevation makes the setting of these glaciers difficult to interpret. I would suggest including a background hillshade on each glacier to accentuate the local topography, and perhaps use a colormap with more breaks to better highlight changes in elevation (such as the matplotlib “terrain” or “gist_earth” colormaps). The authors may also consider removing the Easting/Northing grid labels from each panel and instead include an inset scale bar for each, to allow more space for each panel to be larger.
Figure 5: I don’t feel that it is necessary to include both SCA and AAR here, as the patterns of each are nearly identical. I personally find the AAR to be a more useful metric in this visualization, as it allows direct comparison between the glaciers.
Figure S3: I find it difficult to tell the difference between the Sentinel-2 SR and TOA markers. Could a different color or shape be used to better highlight the difference between them?
Citation: https://doi.org/10.5194/egusphere-2024-548-RC2 - AC1: 'Final response', Rainey Aberle, 09 Jul 2024
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-548', Anonymous Referee #1, 22 Apr 2024
Please see attached pdf file for review comments.
- AC1: 'Final response', Rainey Aberle, 09 Jul 2024
-
RC2: 'Comment on egusphere-2024-548', Anonymous Referee #2, 02 Jul 2024
Summary:
The study by Aberle et al. presents an analysis of using machine learning approaches to classify glacier surfaces facies and to track the spatial and temporal patterns of snow cover on mountain glaciers from satellite imagery. The authors test nine machine learning algorithms on a robust training and validation dataset from five USGS Benchmark Glaciers, using Landsat, Sentinel-2, and PlanetScope imagery. They find that the nearest neighbors and support vector machine models provide the best classification results, and use these models to classify 3000+ individual images of the study glaciers. They then derive daily to bi-weekly timeseries of snow-covered area, accumulation area ratio, and snow line altitude/locations. The study would be a welcome and novel addition to the growing research on remotely-sensed snow cover and glacier dynamics, but I recommend that revisions are made to improve the manuscript quality and to better represent the derived datasets.
Major comments:
- The introduction does a good job of presenting the benefits of and need for improved snow detection methods. However, it should be expanded to include greater discussion and acknowledgement of the existing body of research that has focused on snow identification on glacier surfaces.
- Individual portions of the methods section are well written, but I suggest reorganizing it to make the entire story easier to follow. Specifically: 1) the study area should be moved to its own section, rather than included in the methods, and 2) Sections 2.2-2.4 could be reorganized to follow the structure of Figure 4 to make it easier for the reader to follow. For example, large parts of imagery selection and pre-processing are currently included in Section 2.4 but may fit better in section 2.2, and the seasonal snowline identification could be broken out into its own section.
- The development of reproduceable and extendable code/methods is an important aspect of this project, however I find that the areas of the manuscript where the authors discuss details of the code to be confusing and distracting (e.g. lines 226-227, 234-235). I would encourage the authors to consider whether these details are better suited to be included in the supplemental information or as details on the github page.
- The authors have put considerable effort into developing the methods and creating a thorough dataset. However, I think that more space should be used to present details of the derived products. For example, Figure S3 contains many useful insights that would be better suited for the main manuscript. Specifically, it highlights the differences in temporal resolution between the different imagery products, and well as how consistent (or inconsistent in the case of PlanetScope imagery) the derived products are. Highlighting these results in a main-text figure would improve the presentation of the findings (or perhaps a subset of this figure, such as only a single glacier, or only a subset of years, such that the details of the plot are more easily seen). Other questions which are raised in this figure and throughout the manuscript which could be elaborated on include: are you able to identify significant interannual-variability in the glacier snowline elevation and AAR from these products? How would the results compare when using only a single imagery source, rather than a blend of all imagery as you have done here?
Minor comments:
Line 68: has -> have
Line 69: I would suggest rephrasing “images with spatial resolutions of 1 km or more” to remove the specific number, as most commonly-used satellite imagery is finer spatial resolution.
Line 92: I found that these two points (particularly point 1) were difficult to read. You might consider simplifying or restructuring the sentence here.
Line 109: It should be clarified that the manually generated snow cover observation were made from satellite imagery, rather than from in situ observations.
Line 132: It was a bit confusing to see Emmons Glacier included in this figure immediately after the study area section, where it was not mentioned. Perhaps the details on how it is used should be included earlier in the manuscript to avoid this confusion.
Line 147: The reference to Figure 3 should be to Figure S1, I believe.
Line 204: The inclusion of nine separate ML models is impressive and thorough. Additional information should be included for each (likely in the supplement, I would think) on the specific hyperparameters used for each.
Line 252: How are the masked areas treated in the process of making these histogram? Are the masked pixels included in the glacier elevation bin histogram?
Line 257: What is included in the no-data mask here? Is it only cloudy pixels? Cloudy pixels and off-glacier areas?
Line 254-255: I worry that this may cause a consistent negative bias in the snowline altitudes which are derived. Was a similar approach used to remove sparse snow patches at low elevations to ensure that these snow-ice boundaries were not included?
Line 340: typo for “instils”
Line 343: How are the differences in timing of the manual vs automated snowlines treated in this comparison? What is the range of differences?
Line 343-352: I don’t think the +/- symbol should be used for the IQR numbers here. Including the actual min/max of the IQR would be a more useful metric. eg “… differ from manually delineated snowlines by a median of 116 m (IQR 20–259 m) in ground distance …”
Line 344: including a figure (scatterplot) showing the relationship between automated vs manually-delineated snowline altitude would be a useful addition to highlight the accuracy of the automated methods.
Line 370: Is “the ranges in transient AARs are much larger (0.1–0.3)” referring to interannual ranges in AAR, intra-annual, or range amongst the glaciers?
Line 425: I believe this is the first time that cloud shadows are discussed. The methods should be more explicit that cloud shadows are identified and removed from the imagery.
Lines 440 & 455: I was initially a little confused by these statements on more heavily weighting Sentinel-2 SR observations. I would suggest rephrasing these to make it more clear that the suggestion is for when observations from multiple sources are being synthesized.
Supplement Line 72: typo (repeated words)
Figure 1: A note should be made in the caption (and/or an asterisk added) to acknowledge that the NDSI bands indicated for PlanetScope imagery is not the typically-used SWIR band.
Figure 2: I find that having only the elevation makes the setting of these glaciers difficult to interpret. I would suggest including a background hillshade on each glacier to accentuate the local topography, and perhaps use a colormap with more breaks to better highlight changes in elevation (such as the matplotlib “terrain” or “gist_earth” colormaps). The authors may also consider removing the Easting/Northing grid labels from each panel and instead include an inset scale bar for each, to allow more space for each panel to be larger.
Figure 5: I don’t feel that it is necessary to include both SCA and AAR here, as the patterns of each are nearly identical. I personally find the AAR to be a more useful metric in this visualization, as it allows direct comparison between the glaciers.
Figure S3: I find it difficult to tell the difference between the Sentinel-2 SR and TOA markers. Could a different color or shape be used to better highlight the difference between them?
Citation: https://doi.org/10.5194/egusphere-2024-548-RC2 - AC1: 'Final response', Rainey Aberle, 09 Jul 2024
Peer review completion
Journal article(s) based on this preprint
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 709 | 270 | 61 | 1,040 | 80 | 38 | 39 |
- HTML: 709
- PDF: 270
- XML: 61
- Total: 1,040
- Supplement: 80
- BibTeX: 38
- EndNote: 39
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Rainey Aberle
Ellyn Enderlin
Shad O'Neel
Caitlyn Florentine
Louis Sass
Adam Dickson
Hans-Peter Marshall
Alejandro Flores
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(7191 KB) - Metadata XML
-
Supplement
(2693 KB) - BibTeX
- EndNote
- Final revised paper