A multi-decadal dataset of surface damage on Antarctic ice shelves (1999&ndash;2024)

Tang, Leyue; Bamber, Jonathan L.; Li, Tian; Qiao, Gang

doi:10.5194/essd-2026-414

Preprints

https://doi.org/10.5194/essd-2026-414

Preprints

29 Jun 2026

| 29 Jun 2026

Status: this preprint is currently under review for the journal ESSD.

A multi-decadal dataset of surface damage on Antarctic ice shelves (1999–2024)

Leyue Tang, Jonathan L. Bamber, Tian Li, and Gang Qiao

Abstract. Many Antarctic ice shelves have undergone accelerated thinning and retreat in recent decades, weakening their buttressing effect on grounded ice and increasing the risk of further sea-level rise. Surface damage, including crevasses, rifts and heavily fractured areas, is an important indicator of ice shelf structural integrity, but there is limited understanding of its long-term evolution across Antarctic ice shelves. Here we present a new surface damage dataset for nine representative Antarctic ice shelves, derived from Landsat optical imagery covering the period 1999–2024. These ice shelves include Amery, Brunt, Crosson, Dotson, Holmes, Larsen B, Pine Island, Thwaites and Totten, and encompass a range of change behaviours from relatively stable to rapidly changing systems. A deep-learning image segmentation model was trained on a manually annotated dataset from diverse Antarctic ice shelves to automatically map surface damage. To extend the usable record, Landsat 7 scan-line-corrector-off imagery was restored using a diffusion-model-based framework fine-tuned for Antarctic imagery. The final dataset contains 170 surface damage maps at 30 m resolution, each representing a single ice shelf for a specific year. Temporal coverage varies among ice shelves owing to differences in the availability of usable imagery. The model achieved a mean intersection over union of 0.845 on the test set and 0.822 on an independent validation ice shelf not included in model training. The dataset demonstrates good multi-temporal consistency, supporting its use for time-series analysis. Among the nine ice shelves, Pine Island, Thwaites and Larsen B show the most pronounced surface damage changes during the study period. Compared with existing studies, this dataset provides improved temporal continuity over multi-decadal timescales at substantially finer spatial resolution, offering new insights into the long-term evolution of Antarctic ice shelves and contributing to a better understanding of ice shelf instability. The dataset is publicly available at https://doi.org/10.5281/zenodo.20425951 (Tang et al., 2026a).

Received: 30 May 2026 – Discussion started: 29 Jun 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Preprint (PDF, 2732 KB)

Supplement (1312 KB)

Download & links

Leyue Tang, Jonathan L. Bamber, Tian Li, and Gang Qiao

Status: final response (author comments only)

RC1:
'Comment on essd-2026-414', Sepideh Jalayer, 06 Aug 2026
This manuscript presents a valuable Antarctic ice shelf surface damage dataset. However, several parts of the methodology, model evaluation, validation and dataset description need to be explained more clearly. The following comments need to be addressed before the manuscript can be considered for publication.
Section 2.3.1 (Lines 188–197): "close in time" is not clear. Please specify the time gap range between the Landsat 8 images and the corresponding Sentinel-1 and ICESat-2 data used for annotation.

Section 2.3.1, page 8: Only Landsat 8 images from the 2020 austral summer are used for the training and test datasets? and the final dataset covers 1999–2024 and includes Landsat 7, Landsat 8 and Landsat 9 images. Please explain why this is sufficient to represent the full dataset or provide additional validation for the other sensors and time periods. This part needs to be explained more clearly in the manuscript.

Section 2.3.2: It would be helpful to add a figure to show the segmentation model architecture, including example input images, the SegFormer model and the corresponding output surface damage map.

Section 2.3.2: Please explain why you used SegFormer model for this study. Did you compare it with other segmentation models such as U-Net, DeepLabV3 or other transformer based models? Need to briefly clarify the performance comparison and explain why you selected SegFormer model over the other models. This part should be explained more clearly in this manuscript.

Section 2.3.3: Please clarify how the final model was selected. Was a validation dataset used to choose the best model, or was the model from the last training epoch used?

Could you please briefly explain how the hyperparameters (learning rate, batch size, number of epochs, etc) were selected for training?

Section 2.3.3: Please clarify which data augmentation methods were used during training and whether they were applied to all training images?

Please explain more clearly how the training and test datasets were split. To better demonstrate the model's generalizability, please consider using spatial cross-validation, for example, by training on ice shelves in some regions of Antarctica (e.g., western Antarctica) and testing on ice shelves in other regions (e.g., eastern Antarctica). This can provide a better evaluation of the model's performance in unseen regions.

Section 2.4.2: Since the final products were manually refined, please explain how consistency was maintained throughout the dataset. For example, were the same criteria applied to all ice shelves and years?

Section 2.4.2: Please describe the manual refinement process in more detail. What types of edits were made, and were only obvious errors corrected, or were the predicted damage boundaries also manually modified?

Section 3.2.3 (Lines 373–385): The DiffGF restoration is evaluated using only one simulated example. Since restored Landsat 7 images are used in many parts of the dataset, please consider testing the method on more regions and different surface conditions to better show that it works well. Also, Table 4 reports the results for the whole image. It would be more useful to also report the results only for the reconstructed SLC-off gap areas, where the restoration is actually applied.

Please define surface damage more clearly. It would be helpful to explain exactly which features are considered damage and which are not. Based on the example images, the mapped features appear to include crevasses, rifts, and fracture networks, which are also related to surface roughness and deformation. Please provide a clearer definition of what is considered surface damage and explain how it is distinguished from normal surface roughness or naturally crevassed ice.

Section 3: The uncertainty analysis is useful; however, it would be much more informative to include spatial uncertainty maps (or confidence maps) along with the damage maps. For example prediction entropy or the standard deviation/variance of predictions from multiple runs with different random seeds or hyperparameters.
In this study, quantifying data uncertainty is particularly important because it can help identify label noise, inherent image noise and uncertainties in the manual annotations.

Section 3: Please provide validation results for additional ice shelves, other sensors and different years. If possible, please include quantitative evaluation. This is particularly important because there are differences among these sensors, and it would better show that the model performs consistently across different years, not only on the current test data.

Please provide some software and computing details used to train and run the model, such as the programming language, main libraries and GPU.

It is required to add a map to show the spatial distribution of the training and test images using different colors. This can help readers understand their spatial distribution and check whether the test images are well separated from the training images. Please also show any validation images separately, if a validation set was used.
Citation: https://doi.org/10.5194/essd-2026-414-RC1
RC2: 'Comment on essd-2026-414', Ted Scambos, 13 Aug 2026

Review of Tang et al. ESSO --
A multi-decadal dataset of surface damage on Antarctic ice shelves (1999–2024)
The paper presents a potentially useful data set of ice shelf surface damage based on visible-NIR imagery and a machine-learning approach to detecting the surface features indicating damage. The paper is well-written, virtually zero issues with English or grammar, and could be considered publishable as it is, with proper qualification.
But. Ice shelf damage comes in three types: surface damage, (a) rifts and (b) crevasses, and (c) basal fractures. All three weaken shelves, but in differing ways that would be important to modelers seeking to use your data set to explore various ice shelf vulnerabilities. So, while the data set might be valid within the scope of what has been done, and the title, it would be hard to implement for the most logical application(s).
The authors detect surface damage on the basis of sharp linear breaks in pixel brightness – what are the sources of false positives here? Frost patches can produce linear breaks in brightness, were these detected? Please include more of discussion of the false positives, negatives, etc. – what were the causes of these?
In several of the maps of results you _are_ showing surface damage above bottom crevasses, and this stems from surface ruptures initiated by deformation around bottom crevassing (there are many papers to cite here – one is by McGrath for the Larsen C). This is very prevalent on Scar Inlet (your ‘Larsen B’ test area) and the northeastern Larsen C, as well as in the floating parts of Thwaites Western Ice Tongue. This kind of damage is especially common where the ice is thick (in other words, where buoyancy and lateral deformation forces are large).
It would be interesting to consider how to modify the algorithm or segmentation pre-processing to facilitate bottom crevasse detection. This would likely involve a larger cluster of pixels in the segmentation regions and inclusion ice thickness mappings for scaling the detection of the surface undulations that mark bottom crevasses. There are also notable differences in bottom crevasse surface manifestations between warm-water cavities (Bellingshausen-Amundsen Sea coasts) having to do with basal erosion due to melting.
A little more detail on what was manually identified as surface damage would be good to describe. This might only require a few sentences in Section 2.3.1.
In your temporal consistency checks (Section 3.2.5 and 3.3), did you look for seasonal variations in damage extent or detectability? Seasonal variations in snow bridging of crevasses should lead to some reduced evidence of damage in spring, greater evidence in winter (you are close to evaluating this in your Figure 10). ALSO, importantly, was the algorithm affected by solar illumination angle? Low angle autumn images may well reveal more extensive areas exceeding the damage threshold. Further to this --- areas of extremely rough sastrugi might trigger the algorithm to map surface damage. You could look in some key areas of east Antarctica where meter-scale perennial sastrugi form as a test of the algorithm.

Minor comments
In several places, you refer to the remnant Larsen B ice shelf. Consider instead describing this as ‘Scar Inlet Ice Shelf’, as nearly all of the remnant Larsen B is within this ice-shelf-filled bay.
Line 52, cite Scambos et al. (2007) or Haran et al. (NSIDC) for the MOA mosaics. Note, you might also explore the LIMA mosaic (Bindschadler et al.) and RAMP (Jezek et al.) because these are uniformly processed they make a good template for your algorithm. More importantly – run your processing on the REMA mosaic. These different mosaics (esp LIMA and REMA) will establish better baselines for your surface damage processing time-series.
Line 133 – picky, but don’t say ‘austral summer’, say ‘sunlit season’ or similar.
Figure 7 – what is the scale of these image chips? More generally, it would be good to see a close-up view of a test area to see -exactly- what the algorithm is mapping.

Very minor note – I found the lack of spacing or indentation on the paragraphs and the references to be an impediment to quick reading and checking. I suggest adopting some kind of structure to facilitate a quicker scan.

McGrath, D., Steffen, K., Scambos, T., Rajaram, H., Casassa, G. and Lagos, J.L.R., 2012. Basal crevasses and associated surface crevassing on the Larsen C ice shelf, Antarctica, and their role in ice-shelf instability. Annals of glaciology, 53(60), pp.10-18.

Ted Scambos, Senior Research Scientist, ESOC / CIRES
University of Colorado, Boulder

Citation: https://doi.org/10.5194/essd-2026-414-RC2

Leyue Tang, Jonathan L. Bamber, Tian Li, and Gang Qiao

Supplement

https://doi.org/10.5194/essd-2026-414-supplement

Data sets

Surface Damage Dataset for Antarctic Ice Shelves 1999–2024 Leyue Tang, Jonathan L. Bamber, Tian Li, and Gang Qiao https://doi.org/10.5281/zenodo.20425951

Model code and software

Surface Damage Segmentation Leyue Tang https://github.com/tly-code/surface-damage-seg

Interactive computing environment

Jupyter notebooks for temporal variation analysis Leyue Tang https://github.com/tly-code/surface-damage-seg

Leyue Tang, Jonathan L. Bamber, Tian Li, and Gang Qiao

Viewed

Total article views: 199 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	Supplement	BibTeX	EndNote
129	63	7	199	18	9	21

HTML: 129
PDF: 63
XML: 7
Total: 199
Supplement: 18
BibTeX: 9
EndNote: 21

Views and downloads (calculated since 29 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	91	38	2	131
Aug 2026	38	25	5	68

Cumulative views and downloads (calculated since 29 Jun 2026)

Month	HTML	PDF	XML	Total
Jul 2026	91	38	2	131
Aug 2026	38	25	5	68

Viewed (geographical distribution)

Total article views: 150 (including HTML, PDF, and XML) Thereof 150 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 30 Aug 2026

Download

Preprint (2732 KB)
Metadata XML

Short summary

Many Antarctic ice shelves are thinning and retreating, threatening future sea-level rise. Surface damage indicates structural weakening, but long-term records are scarce. We created a multi-decadal surface damage record for nine representative Antarctic ice shelves using satellite imagery and deep learning. The dataset shows that Pine Island, Thwaites and Larsen B have experienced the most significant damage changes since 1999, providing valuable insights into ice shelf stability.


Total:	0
HTML:	0
PDF:	0
XML:	0