the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Multi-sensor deep-learning mapping of surface and basal fractures across Antarctic ice shelves during the last four decades
Abstract. Ice shelves buttress the discharge of Antarctica’s outlet glaciers, mitigating their contribution to sea-level rise. Yet progressive damage through fracturing remains largely absent from projections, partly because long-term observations of ice-shelf fractures remain limited despite their critical influence on ice-shelf stability. Here, we introduce BASIS (BAsal and Surface Ice-shelf fracture extraction System), a deep-learning model for mapping surface crevasses and basal-fracture across Antarctic ice shelves. Unlike existing approaches based on classical edge detectors or models trained on edge-based datasets, BASIS is designed to delineate the full extent of fracture features, including diffuse basal-fracture signatures and highly damaged, chaotic regions. We train, validate, and test BASIS, based on DeepLabv3, using a unique manually labeled fracture dataset spanning 14 Antarctic ice shelves and the full Landsat observation era. Model performance approaches inter-annotator agreement, while maintaining the distinction between basal and surface fractures. We quantify biases associated with sensor transitions and observation availability and develop a framework for producing consistent multi-sensor fracture records. Annual mosaics reveal the spatiotemporal evolution of ice-shelf damage from the 1980s onward and show that basal-fracture signatures dominate the mapped damage patterns across the ice shelves examined here. At Pine Island Glacier Ice Shelf, BASIS resolves successive stages of shear-margin damage together with a progressive increase in basal-fracture. At Mertz and Borchgrevink ice shelves, it reveals extensive networks of basal-fracture signatures that are poorly represented in existing continent-wide products based on Sentinel-1 SAR imagery. Finally, BASIS demonstrates promising cross-sensor transferability to Sentinel-2 and ASTER optical imagery and, despite having been trained exclusively on Landsat data, also recovers coherent fracture patterns from SAR imagery. These results provide a foundation for large-scale, multi-sensor monitoring of Antarctic ice-shelf damage across unprecedented spatial and temporal scales.
- Preprint
(46414 KB) - Metadata XML
-
Supplement
(11181 KB) - BibTeX
- EndNote
Status: open (until 14 Dec 2026)
- CC1: 'Comment on egusphere-2026-5765', Jonathan Bamber, 06 Nov 2026 reply
-
RC1: 'Comment on egusphere-2026-5765', Anonymous Referee #1, 13 Nov 2026
reply
Shahateet et al. introduce BASIS, a new deep learning (DL) framework based on DeepLabV3 for mapping fractures on ice shelves. Unlike existing methods, their approach can distinguish between surface and basal crevasses and works with multi-sensor satellite data, which expands what is currently possible for fracture and damage mapping on ice shelves. The test data is limited to only three ice shelves, though I understand that manual labelling is time-consuming. The authors present their results for three exemplary ice shelves in informative figures and highlight changes in basal and surface fracturing, which provides new, long-term insights into crevasse evolution on ice shelves. The examples from different satellite sensors, such as Sentinel-1, show that the method transfers well across sensors, which provides valuable opportunities for future analysis on different sensors. Alongside the new method, the authors plan to publish the training and test data as well as the code, which I very much appreciate in terms of open access and reproducible research. The manuscript is well written but in parts quite long, hence any chance to shorten the manuscript and densify information should be used.
Major comments:
- The training and test data were labelled for two classes, namely basal crevasses and surface crevasses. However, the term “surface crevasse” may be misleading. Judging from the figures, the features you refer to as surface crevasses appear to be rifts filled with fast ice/mélange. I suggest either introducing a third class, “rifts”, or, if all the surface crevasses you mapped are in fact rifts, renaming them accordingly throughout the manuscript.
- You highlight that your method works for imagery back to Landsat-1 and can produce 40-year time series. How do you account for the differences in spatial resolution between the early and the more recent Landsat missions? Does a lower spatial resolution lead to fewer detected fractures, or does it not affect the results? How does the more limited availability of early Landsat data affect your time-series results (e.g., Figure 8)? Do observational gaps and longer revisit intervals affect the comparability of the time series? These points could be tested experimentally (e.g. by also reporting accuracies for Landsat 1–5, as in Figure 5), or at least addressed in the Discussion.
- The accuracies reported in Table 1 for the different delineation rounds are very close to each other (especially DR1–DR3). To demonstrate the improvement in accuracy more convincingly, I recommend reporting the mean and standard deviation across at least three model runs. This would make your point clearer and the results more robust for future comparisons. Averaging across at least three runs is now standard practice when presenting new machine learning methods.
Specific comments
Title: Consider changing the title so it better reflects the content of your manuscript. It introduces BASIS as a new framework to detect fractures. With the current title I would expect pan-Antarctic fracture time series over the last four decades.
L20: Where do the 61% come from? Please check the reference.
L126: “From Landsat 4 onward, satellites operate” is not correct. Landsat 6 never operated.
L148: For the reader it would be very helpful to have a figure showing different labels of basal and surface crevasses to understand how the training and testing data looks like. Maybe just shifting Figure 3 here would help.
L158: width > 30 m by mapped features that are at least three pixels wide? Shouldn’t it be 90 m?
L177: You state the initial training was based on 318 patches. How many images did this include? Furthermore, what are your input channels? RGB? Please make sure to describe the training parameters precisely for reproducibility.
L180: Please state what hardware (GPU) you used. A batch size of 56 seems quite large for a tile size of 512, how did you define it? Did you test different batch sizes?
L183: You trained for 1200 epochs but which one was the one with the highest validation IoU?
Figure 2: I think this figure is quite busy and hard to understand as the image caption is not so informative. Maybe (as mentioned earlier), create one figure on visual examples of training and testing data and labels, then a second one with a) and b) showing the maps on fracture class distribution and sensor distribution. Shift the legend to the according map to make it easier to read.
L262: Consider putting all these numbers into a structured table. This would make it easier to read.
L653: Make sure to state the use of AI for your research e.g. for coding and writing. I am quite confident that at least parts of this manuscript where LLM written/re-phrased/corrected, which is fine, just make sure to state it correctly.
Citation: https://doi.org/10.5194/egusphere-2026-5765-RC1
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 422 | 269 | 39 | 730 | 92 | 41 | 47 |
- HTML: 422
- PDF: 269
- XML: 39
- Total: 730
- Supplement: 92
- BibTeX: 41
- EndNote: 47
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Thanks for submitting this paper to EO, which I read with interest. We have also been working on detecting damage on ice shelves and looked at trying to differentiate different classes of damage but concluded that this was ambiguous from optical imagery alone. I am curious, therefore, how robust your distinction between basal and surface crevassing is? Looking at Fig 4, although it's difficult to be certain, and comparing with our own results, it seems like some of the damage identified as basal in Dotson is snow-filled surface crevasses and some of the damage defined as "surface crevasses" are actually full depth rifts. It would be useful to see what high resolution surface topography (say REMA or Icebridge data) showed over the features identified as basal crevasses and to have some sort of validation of the partitioning of damage into surface and basal. Additionally, a rift is essentially both surface and basal in nature so should, perhaps, have a separate category or should be defined as both. To have full confidence in the classification used here it would be useful to see some validation of the classes.