the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Jun 2025
A machine learning, multi-band spectral reflectance clustering approach for examining physical transformations in landfast sea ice environments affected by spring freshets
Abstract. Spring freshets account for more than 50 % of the annual terrestrial freshwater discharge into coastal margins in the Alaskan Arctic. Given the usual timing of Arctic freshets, much of this freshwater is discharged into coastal waters that are still covered by landfast sea ice formed the prior winter. This riverine freshwater both floods the sea ice surface and creates freshwater plumes immediately underneath landfast ice. We employed machine learning clustering algorithms to identify and characterize spatial and temporal variability in spring freshet overflows in the Alaskan Arctic, using the Sagavanirktok River as a model system. Multiband imagery from Landsat 8/9 OLI at the mouth of this river during the 2016 spring freshet were examined using the Caliniski-Harabasz method, which identified five unique clusters putatively representing areas of dry ice and snow, wet ice and/or snow, snow-free ice, ice-free open water, and areas of spring freshet overflow. A Gaussian Mixture Model algorithm, used to estimate cluster purity, indicated that the cluster representing freshet overflow is the most distinct from other clusters. Applying these approaches to an unusually comprehensive time-series of ten OLI images from 2022 revealed interesting spatial and temporal dynamics of these clusters as the freshet evolved, including the maximum spatial extent of freshet-flooded ice (271 km2) occurring 2 weeks after peak estimated volumetric discharge, and the persistence of organic material-laden freshwater on top of landfast ice up to 10 km offshore until complete ice loss in early August.
- Preprint
(12057 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2025-1450', Anonymous Referee #1, 03 Aug 2025
This manuscript utilizes Landsat imagery and unsupervised classification to divide the landfast sea ice region into five classes. While it successfully distinguishes the Arctic river freshet class—a feature that has not previously been separated in this context—there remain significant challenges in validating these classifications. Although references for the classes are provided, the manuscript lacks sufficient spectral interpretation of each class. If the authors can strengthen the validation of the classifications and provide a more thorough spectral interpretation, I believe the manuscript could be suitable for publication.
General comments:
Figure 2: While it is clear from Figure 2 that five classes are optimal, it would be helpful to include a spectral interpretation explaining why the Calinski-Harabasz (CH) score increases as the number of clusters increases.
P10, L184-187: There is insufficient explanation regarding the references for the spectral patterns used to assign class similarity. Since the classification is based on the similarity to previously reported spectral patterns, I recommend including explicit figures showing these reference spectra and providing a detailed description of how they are spectrally similar to Figure 3.
3.2: While the unsupervised clustering approach and subsequent environmental interpretation are interesting and novel, the study would greatly benefit from additional validation. I recommend that the authors perform at least some visual inspection by comparing classified cluster maps with true color RGB imagery at selected locations. Where possible, manual labeling or comparison with reference data (in situ measurements, prior literature, or high-resolution imagery) should be used to quantify the accuracy of the unsupervised classification. This would significantly strengthen the conclusions.
4.0 discussion: While the discussion covers several broad topics and acknowledges major limitations, I found it somewhat superficial in its engagement with the most critical scientific and methodological issues. The authors could strengthen this section by providing deeper analysis of (a) the physical and biogeochemical implications of the identified clusters, (b) the robustness and potential biases in their unsupervised classification, and, (c) how their findings relate to prior studies on Arctic melt pond/freshet dynamics, could concretely address the limitations identified. Such deeper discussion would significantly improve the impact and credibility of the manuscript.
I recommend moving Section 4.4 (“Blue-band ratios as an indicator of CDOM absorption in cluster spectra”) to the Results section, as it presents concrete findings based on quantitative analysis rather than discussion. In addition, the “Future directions” subsection should be shortened, and I suggest that the authors prepare a new, concise Conclusion section to summarize the key findings and contributions of the manuscript.
Minor comments:
Please include the specific formulas and a more detailed explanation of the CH index, GMM, and k-means clustering methods in the Methods section.
Figure 4: To avoid confusion, please apply a more precise land mask to the figure.
P14, L235: Please explain the rationale for focusing on the year 2022.
Figure 7: The water temperature shown in Figure 7 is not discussed in the main text. Please include an explanation in the manuscript.
Citation: https://doi.org/10.5194/egusphere-2025-1450-RC1 -
AC1: 'Reply on RC1', Luka Catipovic, 20 Aug 2025
Hello and thank you for your comments, I appreciate the feedback. I had a few questions on some of your comments I'd like clarified before I dig into the edits.
First, in section P10, L187-187 you mention you would like to see an explicit figure showing the reference spectra from literature and how they are spectrally similar to the classification spectra. However, these in situ spectra are already included in figure 3 as dotted lines plotted on the same axis as the cluster centroids for all ice/water types except the freshet flood water, whose absence from in situ literature is discussed in the text. If you saw these and had something else in mind please let us know.
Second, in comment/section 3.2 you mention the need for a comparative analysis between true color images and classified cluster maps by visual inspection. Figure 7 shows the entire 10 image time series in both true color RGB imagery (top row) and classified cluster maps (second row). These true color RGB images are derived from the same high resolution remote sensing imagery (30 m pixel size) used for the classified maps and show the development of each cluster throughout the year supported by the meteorological analysis in the text (I will add analysis regarding the water temperature). Again, please let us know if you had a more specific analysis in mind.
Thank you again for your thoughtful and thorough comments!Citation: https://doi.org/10.5194/egusphere-2025-1450-AC1
-
AC1: 'Reply on RC1', Luka Catipovic, 20 Aug 2025
- RC2: 'Comment on egusphere-2025-1450', Alice Bradley, 21 Aug 2025
-
RC3: 'Comment on egusphere-2025-1450', Anonymous Referee #3, 24 Aug 2025
This study proposes that river freshet on landfast ice has a unique spectral signature that allows it to be differentiated from other types of wet ice and melt ponds. This spectral signature arises from the terrestrially derived sediment and organic matter carried by the river. The authors use unsupervised clustering algorithms to preliminarily identify different surface types in satellite imagery, and find one surface type that they categorize as freshet on the basis of its proximity to a river mouth and an uptick in river discharge. However, there are not any in situ observations available to validate this interpretation of the imagery.
While I find the idea of studying freshet using its unique optical properties compelling, I do not think this paper is appropriate for publication at this time. My primary concerns are:
- An extremely limited training dataset. The authors define the freshet spectral signature based on a single satellite image of a single river mouth
- No testing/validation procedure. I understand that unsupervised algorithms are exploratory and not meant to have training/testing data like supervised algorithms. However, I think the unsupervised approach taken here is best used a first step to tentatively identify a spectral signature of freshet that can then be used to test against in situ measurements of freshet observations (which the authors note are not available) or independent observer-annotated satellite images in which areas of freshet have been delineated.
- No discussion of how applicable this identified spectral signature is to other watersheds. The characteristic organic matter and sediment producing the spectral signature likely varies among rivers based on their upstream properties, so defining the freshet signature using one river may be quite limiting.
I find the concept of a distinguishable freshet signature compelling, and I agree with the authors that these phenomena are understudied and challenging to observe. I strongly encourage the authors to increase the ambition of their analysis. While the Arctic coasts are often cloudy in summer, and high spatial resolution satellite imagery presents some computational challenges to work with, I think it would be entirely reasonable to expect this study to put much more effort into identifying additional satellite images of freshet to use for initial unsupervised clustering and later testing/validation. The authors make this suggestion themselves when they write “We recognize these limitations and encourage future users to not apply such unsupervised ML models with the expectation of identifying a set of features, but instead to use the features found by the algorithm as a starting point for examining the system under study.” I do not agree that this task is outside the scope of the current study. I would expect this study could produce something closer to the scale of a database of ten years of pan-Arctic freshet events, rather than a speculative spectral signature of a single event.
Minor comments:
Introduction: Please include some of the literature on identifying melt ponds in remote sensing imagery. For example, Rösel et al. (2012) use a machine learning algorithm to create a pan-Arctic, multi-annual melt pond dataset from satellite imagery.
Rösel, A., Kaleschke, L., and Birnbaum, G.: Melt ponds on Arctic sea ice determined from MODIS satellite data using an artificial neural network, The Cryosphere, 6, 431–446, https://doi.org/10.5194/tc-6-431-2012, 2012.
Figure 1: Please make the satellite image much larger. The entire analysis is premised on this image providing “a clear example of the freshet flooding on ice” (Line 110), however in this figure it is barely visible
Line 148: “An unusual number of cloud-free scenes” - I understand that this is a cloudy time of year, but if days cloud-free enough to apply your method are so rare, it really limits the potential utility of your method
Figure 3: Please add inferred / interpreted surface type to cluster titles
Figure 7: Here the fallibility of relying on an "expert observer" with no in situ observations to validate the method is particularly stark - this "freshet" sequence doesn’t make much sense to me. From the classified images it seems like first the landfast ice completely melts (the “deep water” cluster appears at the river mouth on 5/31) but then somehow the freshet ends up on top of the ice offshore of the open water? Maybe I am misunderstanding where the landfast ice edge is located - a land mask on the classified images would help
Line 246-247: “spatial changes in air temperature” – please include net surface heat fluxes (i.e., from reanalysis) which may be more indicative of surface melt on Figure 7. While not my area of expertise, there must be so much literature about onset of melt pond formation and the heat flux/temperature triggers that precede the appearance of wet ice – please refer to some of that.
Line 274-275: Update to “Four of these groups show [delete "significant"] similarity to in situ surface reflectance…” I don’t think you did any significance testing, and as you acknowledge, there is no observed freshet spectrum for comparison against the fifth group
Lines 323-235: “k-means and GMM models were able to identify a cluster represented by a spectrum that bears resemblance to the expected spectra of freshet water overlaying sea ice” – expected on what basis?
Lines 328-329: “Such information would be highly valuable, for example, in studies that examine how on-ice flooding melts overlying snow cover” - I agree that this would be an interesting application, but how would you account for the dilution of the spectral signature of the freshet by the melted snow?
Lines 339-357: It’s confusing to me that this text is here – it seems like it belongs with the previous discussion of this image sequence in Section 3.3
Lines 502-516: These studies feel quite tangential to the results presented here – perhaps this material is more appropriate for the introduction
Citation: https://doi.org/10.5194/egusphere-2025-1450-RC3
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 358 | 44 | 20 | 422 | 6 | 26 |
- HTML: 358
- PDF: 44
- XML: 20
- Total: 422
- BibTeX: 6
- EndNote: 26
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1