| 26 Mar 2026
An Automated Method for Polynya Detection Using a Geomorphon Algorithm
Abstract. Polynyas, persistent areas of open water within sea ice, are critical features of polar marine systems, facilitating ocean-atmosphere heat exchange, deep water formation, nutrient cycling, and biological productivity. However, current remote detection methods, typically based on sea ice concentration thresholds, often struggle to capture the complex morphology of polynyas, especially fine-scale coastal features, and can be time-consuming to use and inconsistent across spatial scales. This study presents a novel application of a geomorphon pattern recognition algorithm, originally developed for terrestrial landform classification, to automate polynya detection using sea ice concentration data. Focusing on two key Southern Ocean regions, the Weddell and Amundsen Seas, we assess the algorithm’s performance through a comprehensive sensitivity analysis, involving 96 and 144 parameter combinations respectively, and compare the results to polynyas identified using a traditional sea ice concentration threshold-based method. By identifying morphological analogues, such as depressions and valleys in sea ice concentration data, the geomorphon method effectively captures spatial patterns and areal extents of polynyas, closely aligning with results from traditional threshold-based approaches and literature reports. The method's scalability and self-adaptive lookup distance allows detection of both large-scale open water and small coastal polynyas. Application of analytically rescaled parameters to an independent passive microwave sea ice concentration dataset further demonstrated transferability across datasets and spatial resolutions without additional optimisation. Critically, its automated nature enables rapid processing of time series data, up to two orders of magnitude faster than traditional methods, making it well-suited for investigating long-term polynya dynamics. By enabling consistent detection across large datasets, the method provides a framework to support investigations into climate-sensitive ocean processes, including air-sea fluxes, water mass formation, carbon cycling, and ecosystem dynamics in polar regions.
Review of
An automated method for polynya detection using a geomorphon algorithm
by
Hurst, M., and L. Boehme
Summary:
This paper demonstrates how a method that is used by the geological land surface community to delineate topographical features in DEMs can be applied to sea-ice concentration data to delineate polynyas. Polynyas are openings in the sea ice cover that can be considered as depressions in a land surface topography the authors suggest. This is quite cool. The paper describes the method very well and evaluates its technical performance. The paper provides examples of time series of polynya area for two regions in the Southern Ocean derived from sea-ice concentration data of an ocean reanalysis. The paper utilizes mainly monthly data but also shows daily data. While the demonstration itself is quite convincing, the paper is very light when it comes to the evaluation of the obtained results. In addition, little to no information is provided about potential uncertainties of the approach itself - taking into account uncertainties in the assumptions and input data.
General Comments:
GC1: Evaluation:
I can understand that this manuscript is kind of a proof-of-concept study and that because of that you focused a bit more on demonstrating the technical skills of your approach. However, your attempts to evaluate the results obtained, i.e. the temporal development of the actual polynya area, are not sufficiently mature and should be improved in two ways. One of these ways would be find more actual estimates of the size of the polynyas you investigated in the published literature and compare your results with those. Another one would be to find more examples where you map the polynya area maps obtained with your approach onto alternative sources of sea-ice cover information - the easiest of these would be high-resolution sea-ice concentration maps (e.g. of the ASI algorithm based on AMSR2 data at 3.125 km grid resolution). Of course you could also try to find more maps of the kind shown in Fig. 8 and perform an inter-comparison but I guess that this is too much work at this stage because i) it would not be sufficient to dig out 2-3 more of such images and ii) you would need to find a quantitative means to do this intercomparison.
GC2: Seasonal development of signatures / input data uncertainties:
Even though the product onto which you apply the geomorphon algorithm primarily is based on numerical modeling I recommend to add information about the uncertainty and credibility of the sea-ice concentration data you used as input. This drives the uncertainty and plausibility of your results to a considerable amount. The same applies to an even larger degree to the results that are based on sea-ice concentration data sets based on satellite observations. As detailed in the specific comments it is very likely that the seasonal cycle of how sea-ice concentrations can be retrieved inside and in vicinity of a polynya leads to a different uncertainty / credibility of the sea-ice concentrations input into the geopmorhon algorithm. It appears therefore to me mandatory to come up with a sensitivity analysis that takes into account the influence of the seasonally varying sea-ice concentration bias in and around a polynya.
GC3: Overselling:
I might be over-critical but neither the discussion of the uncertainties inherent in the retrieval thanks to the input data nor your attempts to evaluate the results are convincing enough to make statements about the usefulness and application potential of the geomorphon approach. The manuscript is clearly overselling the approach, neglecting that the state-of-the-art of polynya detection and mapping is potentially at a different level than what is written in the manuscript. To transfer the geomorphon approach from a land to a sea ice application is already cool enough and you demonstrated technically that it seems to function well. But clearly there is much more work that needs to be done before the approach can be applied in a credible and reliable manner to the suite of sea-ice concentration data sets that is around.
GC4: Other products:
The manuscript would benefit from a more comprehensive layout of existing other products and approaches to derive the location and extent of polynyas. This would have two advantages. One is that a reader gets a better feeling about how unique your approach is - hence a worthwhile addition to your manuscript. The second advantage is that you could use results of these other approaches for the evaluation - a mandatory, yet not convincing element of your manuscript (see GC1).
GC5: I suggest to be more concise about potential application areas. Perhaps you choose an ocean reanalyses for a particular reason but your manuscript does not state this in a sufficiently clear manner. I think you could increase the value of your work when you indiecate a bit more clearly, that your intention was to develop something for numerical model outputs - simply because you did not discuss the uncertainty sources that are linked to the choice of the input data.
Specific Comments:
L82: Kern et al. (2007) used the so-called polynya signature simulation method (PSSM) (as the title of that paper actually states) to derive the polynya area. It is one of the few studies NOT using sea ice concentration thresholds to define a polynya.
L111-116: Passive microwave based algorithms usually come up with a grid resolution of 12.5 km or finer. I am wondering which re-analysis products you had in mind when you wrote about "high-resolution" gridded fields in the context of re-analyses. It would be good to have a few references here. Also, I was wondering which input data sets these ocean reanalyses use. Any fine-resolved information about the sea-ice concentration must come from somewhere and there are not too many products around that go, for instance, below 12.5km.
L122: Kern et al. (2007) used the PSSM which was developed in the 1990s. I invite you to carefully study the references in that paper as well as in Kern, 2009: GEOPHYSICAL RESEARCH LETTERS, VOL. 36, L14501, doi:10.1029/2009GL038062 to understand the advantages (and disadvantages) of that method and to complete the list of approaches that exist. There are certainly more.
L127-129: This sentence is a bit too dense. Thresholding approaches (using sea-ice concentration or sea-ice thickness thresholds) struggle in general - whether the polynya is irrregularly shaped or not and whether the coastal geometry is complex or not. Sea ice formation in polynyas often involves streaks of new ice that align along the principal wind direction and by that create a quite heterogeneous sea ice distribution within the polynya. Hence, what you write about here are two different aspects.
L130: Please provide an example (reference) where such a manual correction has been applied.
L131/132: I doubt that the methods using thresholds are overly time consuming. On the contrary. Some of these even have the advantage that one can retrieve polynya area and landfast occurrence in one go - e.g. Nihashi and Ohshima (2015: DOI: 10.1175/JCLI-D-14-00369.1)
L133-136: "Consequently ... accessible way." --> I can understand that you want to come up with a credible motivation of your study. But ... any properly tuned / developed polynya detection method based on passive microwave data enables automated polynya area retrieval. I strongly recommend that you tone down these statements.Â
Perhaps your main motivation could come from the fact that - as you described already - the coarse resolution of passive microwave data causes artefacts (key word: land-spillover effect and smearing), causes too small polynyas to not be detected, and causes difficult-to-quantify retrieval uncertainties. Using sea-ice concentration thresholds is affected by the low bias in basically all sea-ice concentration algorithms for thin ice. Using sea-ice thickness thresholds is affected by the additional data that are typically required to derive the thickness of the predominantly thin sea ice (e.g. atmospheric reanalyses which are not necessarily overly realistic in polynya areas). And you have mixed-pixel issues - in the Antarctic but also in the Arctic, e.g. the Kara Sea - also due to the presence of landfast ice.
L146: "spaceborne infrared imagery" does not have "coarse spatial resolution" - at least not compared to passive microwave data and also not to the data you actually used in your study.
L149-151: "Collectively ... type" --> I would say that also all approaches that are based on passive microwave data - be it using sea-ice concentration thresholds or be it using sea-ice thickness thresholds - are automated, data-driven techniques. So I don't understand for what we need this sentence here.
In addition, I suggest that you check the existing literature a bit more to complete the list of approaches. I mentioned the work of Kern et al. (2007) already. You might want to take a look into the paper by Aparicio et al. https://egusphere.copernicus.org/preprints/2025/egusphere-2025-4480/ and there are other studies that should perhaps be mentioned, e.g. by Ciappa, A., et al. who studied Antarctic polynyas using MODIS and/or COSMO SkyMed satellite data. That discussion should perhaps also check other optical imagery from  Sentinel-1 / -2 / -3 approaches that might exist, or Sentinel-1 SAR approaches and also take into account that both AMSR2 and AMSR3 allow for about 3 km resolution sea-ice concentration data.
L165/166: Please bear in mind that using sea-ice concentration as the base layer for the geomorphon algorithm introduces - potentially - the same biases that are common when using a threshold based algorithm - namely the low bias over thin ice and mixed pixel issues.
L184-191: Why did you chose a reanalysis data set for the sea-ice concentration? What is your motivation to prefer a numerical model result over a data set based on observations? What is the source of the sea-ice concentration information in that model, how has that been evaluated and how realistic is it? How well, do you think, is the model able to represent polynya conditions in general?
You state "daily and monthly" output as if this was an advantage for investigating polynya dynamics. Since polynyas can change on sub-daily time-scales (e.g. van Woert, 1999; Kern et al., 2007) it would have been of advantage to use (observational) data that allow to take sub-daily time-scales into account and by this have a more adequate representation of sub-daily dynamics.
How up-to-date is the land mask used in the model. What is its source?
L200: Selecting data from December means that you are selecting summer conditions - a time of the year when the delineation of polynyas with basically all kinds of remote sensing data is particularly easy - for many reasons.Â
L202-204: You selected monthly mean data to account for short-term ... fluctuations ... influence polynas on DAILY time scale ... I don't understand why you select monthly data.
L207-210: I don't understand what you state in this paragraph. In the previous paragraph you wrote that you are studying two polynyas for just one month each. So, a long-term consistent data set is not important?
Also, you did not yet specify the time period for which the modeled sea ice fraction data set would be available (in principle).
L220: All coastal polynyas in the Southern Ocean might be not just bounded by the coastline but by land-fast sea ice and/or by an ice shelf or glacier front. How is this taken into account by your approach.
Fig. 3 / Fig. 1: I recommend to show the same region in Fig. 1 as you show here in Fig. 3. This would make it a bit easier to assign the 3-D structure shown in Fig. 3 to the map and the sea-ice concentration distribution shown in Fig. 1.
L274-280: What you write in this paragraph sounds a bit complicated. If you would have used a sea-ice concentration product based on observations you might have benefitted from a product on the so-called EASE2 grid. Most of these products come with cartesian coordinates (i.e. x and y in meters) anyways and it would be much easier to prepare these for the purposes of the geopmorphon approach.
L298: "thresholds" --> "differences" - see my comment to Table 2.
Table 2, caption: I don't think that "sea ice concentration threshold" is the right term to use here. If I understood your explanation of how you derive the flat parameters then the values you are listing here denote the sea-ice concentration difference between two neighboring grid cells; they have nothing to do with any threshold to define a polynya.
L323-326: After all the disadvantages you have been listing up earlier in your paper, I am surprized to see that for the evaluation of your approach you use products of exactly those methods that you were criticising and/or aiming to improve with your approach. I am therefore wondering whether you haven't considered to use independent, higher-resolution estimates of the polynya location and area. The most obvious thing to do would be to take a look at MODIS and/or VIIRS and/or AVHRR imagery. Another alternative would be using Sentinel-1 SAR imagery. Readers might appreciate clarification in this regard because as written you seem to contradict yourself.
L332-334: You chose an opposite set of values than before (there you set open water to 0 and land to 1; here you set open water to 1). Why?
Why are you utilizing a "new polynya detection threshold variable"? Why new?
The work of Mohrmann et al. utilized CMIP6 climate model data with a substantially coarser grid resolution. Their paper clearly reveals that the definition of the sea-ice concentration threshold needed to define polynyas is strongly dependent on the grid resolution. You yourself stated above that sea-ice concentration thresholds that have been used to define polynyas range between 0.2 and 0.8. You are at the lower end. Why?
L339: "were manually drawn" --> but why? What the advantage of chosing a manual delineation over a statistical one where you simply assign all grid cells below the threshold to open water / polynyas and all grid cells above to sea ice?
L390: Why only the Amundsen Sea region? And: For which time period did you perform this comparison? Also: so far you used monthly mean sea-ice concentration maps. Here the temporal resolution is daily. Why?
L415-448: For somebody who is not overly well familiar with the geopmorphon algorithm it would be cool to learn a bit more about how the different parameter values chosen actually map to sea-ice concentration values and gradients. Why? Because especially for the December 2020 case of the Amundsen Sea the border between the open water and the sea ice is possibly very sharp and distinct because this is an example of summer melt; there is no new ice formation within the polynya. Hence, your december 2020 example will possibly work well for a set of parameters that represent high sea-ice concentration gradients from one grid cell to the next. During winter, however, with new ice formation, sea-ice concentration gradients are less sharp and the geomorphon algorithm should work better with parameter settings that take into account better weaker gradients in the sea-ice concentration. Is this the case? Did you investigate this? I would hypothesize that you cannot use the same set of parameters throughout the entire year. While what you show for the Weddell Sea open ocean polynya might work the way shown, I doubt that for the Amundsen Sea polynya the results shown for the freezing season are as accurate as during the melting season - simply because of a change in the sea-ice concentration gradient.
Figure 6: I think, since you superposed a lat/lon grid onto the maps you do not need to show the arrows that are supposed to point to the North but are not universally correct across the maps.
I suggest to superpose an isoline that marks the ice edge to highlight better where sea ice transitions into open water.
In the caption you should state the temporal resolution and the date of the maps shown.
L484-492: You have not worked out very well what the advantage of your algorithm is over using a simple sea-ice concentration threshold-based algorithm where polynyas are identified by assigning all grid cells below a certain sea-ice concentration threshold (which might be grid-resolution dependent) to a polynya.Â
Apart from that I am not entirely sure I "buy" your argument that you method is particularly fast. You might want to run both methods in parallel - the geomorphon one and an automatic sea-ice concentration threshold one (using R or IDL or python).
L499-502: It would be cool if you could show an example where these geomorphon features map onto real-world sea-ice concentration distributions so that a reader is able to better understand what you are writing here.
L526-529: Here you begin a discussion about the actual sea-ice concentration values within a polynya I voiced this as a missing aspect of your paper already earlier. Please expand on this along the lines I mentioned in my respective comment.
L533: 130 to 20 000 km² ... this is three orders of magnitude. I suggest to be much more specific and find literature where the area of the Amundsea Sea polynya(s) are reported in more detail and not with such a range. This is quite unspecific and does not "help" with the interpretation of your results I find. Maybe the various papers by Arrigo et al., by Kern, by Tamura et al. or by Nihashi and Ohshima might help in this regard.
L534/535: The values you specify here fall outside the range you mentioned just a line before ...
In case of the Amundsen Sea polynya, I recommend that you check the daily development of the polynya location and appearance - just to exclude that the polynya has connected to the open water north of the ice edge - which happens regularly during summer.
L539-540: "Unlike ... austral winter." I doubt this statement is correct because the Amundsen Sea polynya can be identified very well in sea-ice concentration maps based on microwave radiometry and is certainly among the polynyas that can be detected with a threshold based method - simply because it is large enough. If you would speak about the Terra Nova Bay polynya then your statement might be correct.
L546-552: Exactly! I was very surprized to see that your aim was to identify polynyas in monthly mean sea-ice concentration fields. I was close to recommend that you move away completely from using and demonstrating results that are based on monthly sea-ice concentration data - especially if the focus is on coastal polynyas.
L558/559: Highlighting the spectrum of polynya sizes that can be retrieved is certainly a good thing but you should come up with an uncertainty estimate and with an idea how reliably polynyas can be obtained over this large range. The uncertainty is not only driven by the geomorphon approach itself but also by the uncertainty of the input sea-ice concentration data which is unknown - at least as written.
For a sea-ice concentration threshold based algorithm it is relatively easy to come up with a sensitivity - by simply changing the thresholds used. Â In case of the geomorphon approach I don't see a straightforward way to carry out such an uncertainty assessment.
L563: I don't think that products that are used to derive long-term time series of polynya dynamics - such as those derived by Tamura et al., or Nihashi and Ohshima, use manual delineation. I am not aware of a method that is used seriously in the sea ice community that uses manual delineation. Whether "static" is the correct setting is something you need to check for the geopmorhon algorithm as well for the reasons I stated earlier.
L562-583: This entire paragraph should be condensed substantially - because i) manual delineation of polynyas is not the state of the art; ii) using monthly data to study polynya dynamics is not a common thing to do; iii) there exist retrieval methods that automatically derive polynya area using sea-ice concentration of sea-ice thickness threshold based methods.
L594/595: "which often rely on fixed ... variability." --> I don't think this statement is correct as it stands. Again this entire paragraph aims for highlighting the apparent advantages of your approach beyond that it is by itself cool to transfer that method from land applications to sea ice applications. I believe you would do good to be a bit more modest and simply consider that for many applications it is sufficient to look into ~3 km sea-ice concentration maps to check whether there is a polynya or not. Other applications might simply use optical imagery or SAR imagery which are increasingly available and offer considerably finer spatial resolutions.
I also would like to stress - again - that possibly researchers interested in monitoring the Weddell Sea polynya will use a different approach than researchers that are after monitoring Antarctic coastal polynyas around the continent or with a focus on a particular coastal polynya. And while your study is clearly a right step into the direction to spread the word and to invite people to use the geomorphon approach, substantially more work is needed to assess the uncertainty and the maturity of the approach.
Fig. 8 and the paragraph describing and assessing it: These are melting season cases - e.g. easy cases because there is no new ice formation in the polynyas. In both cases the visible images clearly show a different shape and morphology of the polynyas as identified by the geomorphon approach. This figure (and the way you interprete it) clearly confirms what I stated in my previous comment: There is substantilly more work required to demonstrate the quality of the algorithm. The figure also suggests - as said before - that you should perhaps tone done parts of your statements about what the approach can do and what it cannot do to not mislead readers. I suggest that you modify Fig. 8 in a way that you overplot both maps of the polynya in one map - by using transparent colors for instance. That way you could visualize much better how the two products "match". And I suggest you come up with a follow-up study where you present results of an adequate evaluation study - using independent estimates of polynya location and size - or simply high-resolution sea-ice concentration maps or classified Sentinel-1 SAR imagery.
L610-619: Again you "praise" the advantages of your method - without having really proven that your application of the method to ASI-algorithm sea-ice concentration data provides accurate data - simply because you did not perform an adequate evaluation. Please consider toning down statements that oversell your approach.
L635-642: Also this paragraph you need to tone down in light of the discussion provided. The main point to keep it that it is cool to port the approach from a land application to a sea ice application. This step is successfully done. The rest needs work - evaluation and checking the seasonally changing ice conditions in a polynya and their impact on the choice of the geomorphon algorithm parameters.
Â
Editoral Comments / Typos:
L125: Please be more specific: What is a "sea ice threshold"?
L144: What do you mean by "manual validation"? Shouldn't all these products be validated?
L237/238: "Notably .. manner" --> How?
L335/337: Typo: Morhmann --> Mohrmann
L473: "corresponding ... threshold" --> better: corresponding to a grid cell-to-grid cell change in sea-ice concentration of 0.4.
L523: As I mentioned already, I think this term is not well chosen and should rather be termed "grid cell-to-grid cell sea-ice concentration difference"
L524: I note that the Barber et al. (2001) paper is quite specific, dealing with the North Open water polynya which is in the Arctic. How well does this paper fit to the polynyas you have been investigating in your paper?
L538/539: I suggest to also mention the data and the method that is used by Macdonald et al. (2023) - Envisat-1 SAR and 3.125 km AMSR2 sea-ice concentration data - hence they are using a quite advanced approach with "true" high-resolution data.
L541 ... suggests." I would start a new paragraph here to separate the results of the Weddell Sea polynya better since this is a different polynya type.
L544/545: What do you mean by "winter months" here? September? Not clear.
L569: "relative to the conventional .. method" --> "relative to our conventional ... method that involves manual delineation ..."
L644-654: Certainly your approach has potential to be applied to various sorts of data - also other data sets based on numerical modeling such as the CMIP6 model suite - but all this in parallel to a proper evaluation of the results. I suggest that you re-formulate this paragraph accordingly.