| 20 May 2026
A basin-scale mapping method for crevasse depth using ICESat-2: a case study on Greenland's Sermeq Kujalleq (Jakobshavn Isbræ)
Abstract. Surface crevasses across the Greenland Ice Sheet remain a major source of uncertainty in understanding mass loss processes, including calving, ice flow, and meltwater routing. Most previous studies have primarily focused on the two-dimensional characteristics of surface crevasses on the Greenland Ice Sheet, measurement of crevasse depth, especially from catchment to ice-sheet scale, is limited, thereby limiting our ability to quantify their impact on ice-sheet stability and surface hydrology. The ICESat-2 ATL03 data (0.7 m resolution) provide an unprecedented opportunity to measure crevasse depths, yet large-scale applications are hindered by challenges such as massive data volume and noises. Here, we develop an automated and efficient ICESat-2 method for crevasse depth estimation that rapidly identifies crevassed regions using a novel roughness index, retrieves signal photons via a similarity-based weighted density approach, and extracts crevasses for depth estimation using a local extrema method. A total of 18,775 crevasse locations were detected from 2,286 beams on Sermeq Kujalleq in 2019. The average crevasse depth is 7.20 ± 0.03 m, with depth maxima occurring at approximately 20 km and 10 km inland in the south and north ice streams, respectively. Crevasses in the northern ice stream are mainly distributed below 600 m a.s.l, whereas those in the southern ice stream predominantly occur below 900 m a.s.l. Compared to IceBridge ATM data, the RMSE of crevasse depths estimated from the ATL03 product is 0.97 m, 5.30 m lower than that from ATL06. In addition, ATL03-derived depths are approximately 28% and 30% deeper than those from ATL06 and ArcticDEM. This study enables metre-scale crevasses to be incorporated into large-scale analyses of ice dynamics and calving, and highlights the potential of ICESat-2 for large-scale crevasse depth estimation, providing valuable insights for global crevasse mapping.
Competing interests: At least one of the (co-)authors is a member of the editorial board of The Cryosphere.
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.
Chang and colleagues describe the development of a method for mapping crevasse depth from ICESat-2 imagery, applying the method to Sermeq Kujalleq (Jakobshavn Isbræ) in Greenland.
The method appears sound, but the application is not quite as novel as the manuscript makes it out to be, and comparison/contrast with peer literature is currently not in a detailed enough state for me to figure out where the true advances in this study lie. It is neither the first study to detect crevasses in ICESat-2 ATL03 data (but there is no comparison to alternative methods to show that the method advances the field); nor the first study to map crevasses over Sermeq Kujalleq (but there is no comparison with existing maps nor any science findings that advance the state of the literature). In this context, it is challenging to identify components of the current manuscript that satisfy the Cryosphere’s review criteria surrounding originality and significance. The paper likely requires either (i) a more explicit and rigorous comparison with existing methods in order to establish this method as a verifiable advance in processing quality or capability; or (ii) a stronger science component that apply the method to the study of the structural glaciology and/or dynamics of the study glacier.
Major Comments
The research gap identified by the paper is that “efficiently extracting crevasses and estimating their depths from the ATL03 product remains a non-trivial challenge”. There is, however, an established literature of extracting crevasse products from ATL03 data, of which the study identifies a couple (Herzfeld et al 2021; Chang et al 2023). Given this method represents an evolution of the method of Chang et al. (2023), the work of Herzfeld et al. (2021; and the uncited Herzfeld et al. 2017 https://doi.org/10.1109/TGRS.2016.2617323) is probably the most important comparitor, as the dda-ice algorithm is now established across a number of studies. The implication in this study is that the new method is “automated and rapid’’ (e.g. L548) and represents a step-change in processing capability that allows it to be applied across a large spatial scale (L100-105). However, given no information is provided about the actual processing times involved, nor how this compares to prior methods, I find that these statements are hard to justify and perhaps unfair to the prior work that has been done in this space. If these statements regarding the efficiency and prior processing limitations were to remain, I think it would be necessary to provide some sort of quantitative evidence regarding the improvement in processing capability cf. dda-ice. Certainly, given this method exists, it seems a little unusual to compare instead against ATL06 data in section 5.1, as ATL06 is *not* the null alternative given the existence of alternative ATL03 processing techniques. At the very least, the manuscript should spend longer comparing the new method to the dda-ice algorithm (and indeed, there are strong parallels between the two methods), identifying the differences in methodological choices between the two, and potential advantages/disadvantages of each.
A second claim is “this is the first study to map crevasse depths from ICESat-2 over the GrIS” (L528-529). Whilst this is true, there are two alternative studies that mapped crevasse depth at large scales, and more that map other crevasse components. The first is the study of Enderlin and Bartholomaus (2020; https://doi.org/10.5194/tc-14-4121-2020), who process OIB lidar processing to map >50,000 crevasses across 19 Greenlandic outlet glaciers. This study is likely worth discussing and contrasting, given the processing chain for OIB lidar displays strong parallels with processing for ATL03 data (i.e. extracting depths from a flight line of point returns). It provides alternative observations at Jakobshavn that should be compared directly. It also makes different assumptions about how to infer crevasse depth (discussed below). The second is the study of Chudley et al. (2025), which extracts crevasse depths from ArcticDEM data. This is mentioned in the text (L501), but in section 5.2 the compared crevasse depths are only identified manually. This seems a shame considering (i) using a prior automated method could provide a much larger sample size than the 16 in this study; and (ii) given an automated method likely performs worse than manual picks, the proposed method could actually be even better than comparator methods. There are also alternative methods/studies that map the *presence* of crevasses, including Van Wyk de Vries et al. (2021; https://doi.org/10.1017/jog.2023.3), and Khan et al. (2025, https://doi.org/10.31223/X5HM9P).
Given the data of Chudley et al. (2025) is public, and the methods of Chudley, Van Wyk de Vries, and Khan are all publicly available (https://github.com/trchudley/crevdem, https://github.com/MaxVWDV/Gabor-crevasse-detector, https://theghub.org/resources/crevassedetect/about), it is unusual to have no comparisons to any existing methods, which could help highlight the benefits of the method.
If the method is indeed better or more efficient than comparator methods, is it disappointing there has been no effort to share any easily reusable code via a GitHub repository or similar. Given that other approaches have shared their approach (see above), it is worth doing - especially as it would give the approach a clear advantage over dda-ice method, which is currently not shared.
A further note about the Section 5.3. The text claims that difference in depths are to do with photogrammetric error, but studies using high-resolution data have asserted that the apparent shallow depths of large (>10 m) crevasses are likely not an error but a real observation of infilled ice debris (Enderline and Bartholomaus, 2020; Chudley et al. 2021). Indeed, the comparitor profiles presented in Fig 13 appear, to me, broadly ‘within error’, and the average difference in depth of 1.40m seems reasonable considering the two datasets do not appear to be coregistered – or else probably explained by the higher resolution of the ATL03 (0.7 m cf. 2 m) rather than any photogrammetric error. (Elsewhere in the text it is instead said that ICESat-2 depths are “30% deeper” (L40;L560), but I cannot see where this figure comes from). Enderlin and Bartholomaus (2020) explicitly account for the debris infill in lidar data by making an assumption that crevasses follow V-shaped geometries, and estimating the ‘true depth’ of the crevasse by linearly projecting crevasse walls to an extrapolated intersection. Of course, any extrapolated intersection assumes that any linear data is exactly orthogonal to crevasse depth (which will be violated for ATL03 data) - but either way, it is likely best to acknowledge that the true crevasse depth – on the scale observed at Jakobshavn and similar fast-flowing outlets – cannot be determined in any optically-derived dataset.
Conclusion
Given my comments above regarding comparator papers from both the ICESat-2 and Greenlandic crevasse literature, I believe that this paper does not, in its current state, sufficiently communicate the “originality” not “significance” components of the Cryosphere review criteria. However, I think a clear pathway to this does exist, with a better focus on novelty either as a (i) a method paper or (ii) a science paper. If (i), it would be necessary to see a robust comparison to alternative approaches (especially dda-ice), which would help identify this as a quantitatively better method, either in terms of processing speed or accuracy. The existence of publicly available code (e.g. in a GitHub repository) would present a step-change from the dda-ice comparator method and help this method be distributed through the community. Through this route, I can see real impact in wider community use. If (ii), the dataset presented here should probably be paired with some aspect of scientific interrogation, either through some investigation of e.g. the relationship between depth and conventional predictors of crevasse propagation (following e.g. Enderlin and Bartholomaus, 2020), or by expanding the dataset to compare changes through time to prove the utility of the method (following e.g. Van Wyk de Vries et al. 2021). With the latter, there is real value in being able to ‘fill in’ the sparse observations by Chudley et al. (2025), or demonstrate temporal changes following the Enderlin and Bartholomaus (2020) statistics. This would evidence that the proposed method can provide invaluable temporal coverage to balance the spatial coverage of other datasets.