Brittle fracture occurs mechanically in three modes, which are commonly denoted as I, II and III . Mode I represents cracks opening in the direction of the applied tensile stress, while modes II and III represent cracks caused by in-plane and out-of-plane shear stresses respectively . Additionally, due to the viscoplastic properties of ice, ductile processes can augment these brittle failure modes and produce complicated crevasse patterns. Mode-I failure tends to result in parallel, sharp-sided surface crevasses or rifts clearly visible in the Sentinel-1 backscatter signal and basal crevasses which can result in visible large-scale depressions in the surface , especially if subject to subsequent ductile deformation (“necking”) above the crack tip . However, it is difficult to discern whether such depressions are an indicator of basal crevasses or other processes such as channelised melting of the sub-shelf, so we focus our methods on the detection of surface crevasses in the knowledge that the sharpest-looking basal crevasses (i.e. where there are large-magnitude intensity gradients perpendicular to the crevasse walls) will be detected as well. Shear failure, ubiquitous in the margins of fast-flowing ice streams and shelves, can result in macroscopic crevasses or rifts when severe. Often, however, it results in interacting networks of microfractures, particularly on grounded ice streams. Though regions of high backscatter can indicate that the ice is rougher in the shear margins than elsewhere on the glacier, the “micro” nature of these fractures means that we cannot rely on seeing them in the surface signal, so we do not attempt to map this type of diffuse fracture.

In general, we restrict our attention to “sharp-sided” features that appear crevasse-like in isolation: largely surface fractures from mode-I and shear failure. We classify these features into two sets: type-A and type-B, based on their qualitatively different expressions in the backscatter data (Fig. ). Different methods are required for the extraction of these different classes given the disparity in their visual appearance, and the resulting datasets are useful for different purposes. Type-A features are large, multiple pixels in width, and are visible from many look angles of the satellite. Type-B features appear as fine, bright lines in the backscatter images. They can be pixel scale in width and are most visible when the horizontal component of the satellite acquisition angle is perpendicular to the crevasse walls. Hence, these features need to be recovered from data covering multiple acquisition angles. Happily, in many places, the Sentinel-1 acquisition tracks overlap obliquely and, often, at near-perpendicular angles (Fig. c and d). Broadly, these two categories distinguish crevasses on floating and grounded ice: type-A features include ice shelf surface crevasses, rifts and any basal crevasses that cause narrow surface depressions, while type-B features include grounded surface crevasses and, to a far lesser extent, narrow ice shelf surface crevasses bridged by snow.

We developed two neural networks and additional filtering techniques to identify these separate features from individual geocoded single-look complex (SLC) amplitude images, acquired using the interferometric wide-swath (IW) mode of Sentinel-1 and methods for constructing combined monthly mosaics of type-A and type-B crevasses. Figure  shows example type-A and type-B crevasses visible in Sentinel-1 SAR data over Crosson Ice Shelf. Figure  shows an overview of the procedures involved in constructing monthly fracture mosaics, such as that shown in Fig. .

We applied the method described above to every Sentinel-1 acquisition between January 2015 and July 2022 and produced composite crevasse maps covering the Antarctic Ice Sheet for each month. Figure  shows an example map made using Sentinel-1 acquisitions from April 2022. The results show crevasses to be a feature of a large and varied set of ice streams and shelves across the continent – from rifts in the shear margins of the fast-moving ice shelves of the Amundsen Sea (Fig. h) to the fine surface fractures fringing the grounded ice streams of the Amery Basin (Fig. d).

The type-A fractures identified by the network take a wide variety of forms and are exclusive to floating ice shelves. Of those observed, the brightest and most identifiable are large ice shelf rifts, such as Chasm-1 on Brunt Ice Shelf and those penetrating into the bulk of Shackleton and Larsen D ice shelves (highlighted with grey arrows in Fig. b, e, a respectively). Many fast-flowing ice shelves exhibit severe crevassing in shear margins that connect them to slower-flowing parts of the ice shelf, for example on Stancomb–Wills Ice Tongue, Fimbul Ice Shelf, Shackleton Ice Shelf and Pine Island Ice Shelf (highlighted with blue boxes in Fig. b, c, e, h respectively). We also see fractures resulting from the interaction of ice shelves and ice rises around the coastline – for example on Larsen D Ice Shelf (Fig. a).

Our observations show that type-B crevasses, though a less varied set of features, are as prevalent across the continent as type-A crevasses and occur in every major basin of the Antarctic Ice Sheet, generally in dense patches. Approximately 85 % of type-B crevassing appears on grounded ice. Our observations show that type-B crevasses are particularly dense and widespread in the ice streams of the Amundsen Sea Embayment (Fig. h). In a composite crevasse map of the grounded part of the ice sheet from June 2021, this region accounted for around 10 % of grounded crevasses despite covering only 3.5 % of the imaged surface area. These ice streams have undergone significant dynamic change over the last few decades, with ice flux across the grounding line increasing by 40 % to 100 % since the early 1990s . The large number of grounded crevasses observed in this region may be a consequence of the sudden increase in longitudinal strain rates which will accompany the observed speedup. The effect of crevasse fields like this could be to decrease the effective viscosity of the ice. Hence, fracturing can provide a mechanism for the sudden manifestation and persistence of ice dynamic changes, beyond that which can be accounted for by dynamic thinning. In many grounded areas, however, crevasse depth is likely to be only a small fraction of ice thickness due to large overburden pressures .

Elsewhere around Antarctica, surface crevasses on grounded ice appear sporadically in patches. In some instances, they contour the edges of ice streams, such as on Lambert Glacier in East Antarctica (white arrow, Fig. d). Elsewhere, the locations of patches of crevasses appear dependent on vertical shear stresses, for example mode-I crevasses forming due to abrupt changes in basal slip or mixed-mode crevasses caused by sharp changes in bed topography (white arrows, Fig. g). The sizes of these patches of crevassed ice are of the order of ∼10–100 km in the along-flow direction, showing that the crevasses can be healed when stress conditions change. Crevasse visibility is also influenced by spatial variability in the depth of drifting snow and, in regions for which image acquisition angles vary little, changing crevasse orientation. In future, we can hope to constrain these additional factors using additional sensors to better bound the regions in which type-B crevasses appear on grounded ice.

We finally note that, due to the shallow networks, the bootstrapping procedure, and the absence of training data seaward of the Antarctic coastline, the networks are sensitive to local textural deviations at, and beyond, the calving front. For example, leads and fractures in sea ice, dense ice mélange, iceberg boundaries and calving fronts exhibit strong signals in the fracture mosaics when unmasked.

Overall, we consider the Antarctic fracture maps we produce to be an accurate representation of crevasses across the continent that are linear on scales that are large compared to the resolution of the data, with a few exceptions. There are two components that inform this conclusion. Firstly, previous studies suggest that the majority of such crevasses are visible as type-A or type-B features in the Sentinel-1 SAR backscatter images . Secondly, the methods we have developed extract a majority of features in the backscatter images while highlighting a few features erroneously. The fracture maps covering Dronning Maud Land, the Amundsen Sea sector and the eastern Antarctic Peninsula (Fig. b, g and h respectively) show the lack of noise present in the monthly mosaics at different scales – indicating a high specificity for both type-A and type-B crevasses. In large part this can be attributed to the ability of the neural networks to deal with radar speckle and the efficacy of our parallel-structure filtering algorithm. These figures also show how mosaicking over monthly windows results in maps that do not display visible processing artefacts, such as the edges of SAR frames or boundaries between regions of different acquisition geometry. In part this is a consequence of using raw, un-normalised SAR data. This provides confidence that large features can be mapped across image acquisition boundaries and therefore provides a good representation of the ice sheet surface.

The few locations in which type-B features are present in the crevasse maps without obvious fractures visible in the backscatter images are in regions of steep topography, such as the ice rises and mountains around Crosson Ice Shelf (Fig. a3, b3, circled in green). As the topography of these regions is essentially unchanging over the time period of this study, these features could be dealt with by defining a mask based on the gradients of a digital elevation model or by retraining the neural networks with additional data covering such areas of steep terrain. However, as these regions are distinct from those with the kind of large-scale dynamic behaviour we are interested in, the misattribution of crevasses to these areas is incidental.

The specificity of crevasse detection is also low in regions where there is a large amount of surface meltwater – for example near the grounding line on Roi Baudouin Ice Shelf in eastern Dronning Maud Land (Fig. c1, d1, circled in blue). Values of type-A crevasse probability are high in regions of surface melt due to the presence of features that are not usually visible, such as sharp contrasts in backscatter at the boundaries of melt ponds, fluvial features originating from the persistent pooling and flow of surface water (e.g. near the grounding line of Amery Ice Shelf – Fig. c) as well as a greater sharpness of shallow surface depressions due to the decrease in microwave penetration depth. In static crevasse maps such regions appear to have a greater amount of crevassing compared to nearby areas with a comparable surface structure. However, there are only a few locations for which this is a persistent issue (e.g. Amery and George VI ice shelves), with other locations only experiencing melt events in the austral summer (e.g. eastern Dronning Maud Land). However, the effect of meltwater, even if intermittent, means caution is required when comparing crevasse maps covering ice shelves known to be affected by surface meltwater at different locations or times. If surface melt in Antarctica increases in the future in response to a changing climate, then methods will need to be developed to remove these features – or conversely to isolate and use the information as a proxy dataset for meltwater extent.

The method appears sensitive to the great majority of crevasse-like features visible in the backscatter images across the continent. Unsurprisingly, it is most sensitive to sharp-sided ice shelf surface crevasses and rifts. However, some smooth features – likely to be basal crevasses – are under-represented in the type-A maps, for example those near the calving front of Dotson Ice Shelf and those forming a track along the central–western part of Amery Ice Shelf (shown by grey boxes in Fig. h and d respectively), which appear only faintly.

The type-B mapping identifies the bulk of crevasses visible in the images. The exception is at the boundary of patches of grounded crevassing where the edges are sometimes cut off by the low-resolution filtering. There are also crevasse fields that include small regions, of the order of 5 to 10 km in width, in which crevasses likely exist but are less visible in the backscatter images (e.g. on the Thwaites ice stream – Fig. a2, b2, circled in white). These regions are blank in the crevasse maps, despite the fact that crevasses very likely propagate through them, linking the identified crevasses on either side.

To further assess the performance of our type-A and type-B fracture maps and to quantify statements regarding the specificity and sensitivity of our crevasse maps, we compared the monthly mosaics with three manually annotated Sentinel-1 IW SAR frames, created independently, without reference to the fracture mosaics. The European Space Agency provides in-house references for the relative orbit and frame numbers for the image acquisitions which we shall use to distinguish the SAR images here and in the rest of the article. We will refer to the relative orbit number as “path” and abbreviate the path and frame that identifies a specific acquisition as “PF”. The three SAR images chosen cover Crosson and Dotson ice shelves and their tributary glaciers (PF 7/913, dated 7 June 2021), Fimbul Ice Shelf (PF 2/931, dated 15 August 2018) and the grounded Fimbul Glacier (PF 2/925, dated 15 August 2018). We annotated these in their entirety in order to help facilitate future intercomparison with other methods. We selected these frames to represent a challenge to the method, as the full variety of crevasse features is present together with large regions of steep topography and persistent surface melt. For type-A features, we annotated individual crevasses at the pixel level in single SAR images corresponding to the chosen frame. However, for type-B features, it is all but impossible to pick out individual crevasses by hand. Instead, we delineated the boundaries of crevasse fields, which we then compared with versions of the type-B mosaics that were smoothed and thresholded to produce binary maps of contiguous crevasse fields. For the case of the SAR image covering Crosson Ice Shelf, we combined annotations from the given frame and another with a relatively oblique acquisition angle (PF 10/855, dated 7 June 2021) to reflect the different crevasses that were visible from the different angles. Given the continuous output and class imbalance inherent in the type-A mosaics, we produced receiver–operator characteristic (ROC) curves for the evaluation of the type-A maps, while we report confusion matrices for the evaluation of the type-B processing. Results are shown in Fig. .

The ROC curves quantify the sensitivity or specificity of the type-A mosaics over the full range of threshold values [0,1]. The areas under the curves for the frames covering Crosson, Dotson and Fimbul ice shelves respectively are 0.93 and 0.91, showing high discriminatory power (Fig. g and h). This is despite the failure of the network to extract a large portion of the likely basal crevasse impressions on Dotson Ice Shelf (Fig. d and e), those west of the Fimbul western shear margin (Fig. a and b), and the misattribution of flow features on the central trunk of Fimbul Shelf and fluvial features in its eastern shear margin as crevasses (Fig. a and b). The intersections of the annotated and predicted crevasse fields show a great deal of overlap in the crevasse fields on Fimbul (Fig. c), Pope, Smith and Kohler (Fig. f) glaciers, with the largest fields sharing the most overlap. There are a number of smaller features in the type-B mosaics that do not appear in the annotations – especially in the case of Fimbul Glacier and surroundings. This is where mountains and steep rocky features have been identified by the model as crevasses. However, the confusion matrices (Fig. i) show that these areas cover under 5 % of the total “undamaged” ice according to the manual annotations.

After establishing the above method of reliably extracting fracture-location data, we moved onto the natural question of whether it is possible to use this dataset to measure changes to “damage” through time. At the outset it was unclear whether the 7.5-year study period was long enough for real structural change to evolve in response to ice dynamic processes in Antarctica or whether the impact of radar interactions with the ice surface properties may cause crevasse mapping to vary too much in response to environmental factors unrelated to crevasse evolution, such as weather-induced surface melt. It is certainly clear that variability in the properties of the ice surface renders the direct comparison of fracture maps at different times an unreliable method of assessing change.

The method we introduce here measures trends in the local density of fractures using the full time series of fracture maps, producing a scalar “fracture density change” value that can be used to compare different locations.

A brief analysis of the fracture maps shows minor changes to crevassing on the grounded ice sheet over the last 7.5 years. We therefore focused on ice shelves, where crevasse patterns change on shorter timescales due to elevated flow speeds, interaction with rapidly changing ocean conditions and higher-amplitude loading–unloading cycles. More specifically, we considered the ice shelves of the Amundsen Sea Embayment, given that a large amount of the ice dynamic change in Antarctica was observed in that region between 2015 and 2022.

We define fracture density as the fracture maps integrated over an area of interest and use this as a heuristic measure of how crevassed that region is. We looked for linear trends in this parameter over the observational period. To analyse the spatial pattern of change, we first defined a fixed 2.5km×2.5km grid of points over the ice shelves. Time series of fracture density and backscatter standard deviation were extracted from daily mosaics over a 10km×10km square buffer around each point before the fracture density change and error were found. Given that we were not interested in the development of individual crevasses, we did not apply Lagrangian corrections to the crevasse maps.

As mentioned above, we had to account for the dependency of our fracture maps on the surface properties of the ice before calculating trends in the time series. For example, seasonal changes to crevasse visibility due to changing firn water content or thickness of the snowpack can dominate the signal over changes in crevasse length or concentration . It was necessary, therefore, to separate the parts of the signal due to real changes in the crevasse pattern and those due to changes in the surface expression of the crevasses resulting from such unknown environmental factors. This is only possible given the large number of crevasse maps, and hence fracture density data points, in the time series given the short repeat period of the Sentinel-1 satellites. Firstly, we discarded measurements made between December and March each year, due to the prevalence of surface melt during those months. We then made use of the observation that the surface properties of the ice seem to be discernible from the local standard deviation of the backscatter signal. We used this as a proxy to define sets of dates for which ice surface conditions were similar and constructed an ensemble of fracture density time series for each region. By taking a weighted mean and standard deviation of the trends and multiplying by the time span, we defined an estimated fracture density change over the time period (Fig. a, b, f, j) and provided an uncertainty (Fig. c, j, k). See Appendix  for more details.

By applying the method defined above to daily fracture map mosaics, we derived an estimated change in fracture density for the ice shelves of the ASE (Fig. a). For the most part, our results show that notable changes in fracture density over the observational period were confined to Pine Island, Thwaites and Crosson ice shelves (Fig. b, f, j respectively), with change elsewhere attributable to modest calving events or terminus advance (Fig. a).

On Pine Island (Fig. b–e), a large region of fracture density change in the interval 0.08–0.3 shows a significant deterioration of the southern shear margin over the observational period , continuing a decades-long pattern of structural decline on this ice shelf . The intense fracture density change at the seaward-most part of the shear margin reflects its disintegration over the course of 2020 (Figs. b, b and c), completely decoupling the ice shelves of Pine Island Glacier and Piglet Glacier which, until 2018, flowed north-east into the southern shear margin of Pine Island Ice Shelf, rendering it unbuttressed and marine-terminating. We also see a general region of increased fracture density at the seaward limit of the ice shelf, which can be attributed to a series of propagating rifts forming downstream of an ephemeral grounding point near the centre of the ice shelf that led to calving events between 2015 and 2020 .

Moving around the coast of West Antarctica to the ice shelves of Thwaites Glacier (Fig. f–i), a more complicated picture emerges. On the Eastern Ice Shelf, the data suggest a highly variable pattern of structural change. A patch of decreasing fracture density (black circle, Fig. f) indicates where sharp-sided crevasses have travelled further onto the shelf, followed by smoother (likely shallower-surface or basal) crevasses (Fig. d). Around the pinning point to the north of the ice shelf, we see decreased fracturing around its western end and increased fracturing to the south and east. This is due to crevasses or rifts opening up in situ, perpendicular to the orientation of the pinning point on the main body of the shelf in response to increased shear stress in the latter half of the 2010s .

Thwaites Glacier's western ice tongue displays increased fracture density on the floating section of the main trunk closest to the grounding line and in the ice to the south of the eastern shear margin (black box, Fig. f). Fracturing on the central trunk has been increasing steadily over the observational period, while that on the eastern shear margin increases and decreases with acceleration of the main trunk . Changes to fracturing in the regions closest to the grounding line have been periodic on timescales of a few years (Fig. b and d), so our linear trends do not capture the behaviour of this region in full. The protruding ice tongue has been degrading steadily over the observational period . However, while our data show a large region of positive change in fracture density, negative trends in fracture density are shown on the eastern side of the ice tongue. This reflects a reduction in the density of the mélange separating the ice tongue from the Eastern Ice Shelf, where, at some point, our fracture maps stop recognising the spaces between icebergs as fractures.

Our observations show changes in fracture density on the main body of Crosson Ice Shelf and in its shear margins (Fig. j–m). On the main body, these are largely due to the advection of rifts towards the ocean, as can be seen by the stripes of increasing and decreasing fracture density (Fig. j), though fracturing has progressed and increased slightly in density on the shear margins. Finally, we note that, despite our observations showing fracture density change outside of the Amundsen Sea sector to be limited (Figs. a, b), some ice shelf crevasses which may be linked to interesting glacial processes – such as those perpendicular to the direction of ice flow that follow the sub-shelf melt channel in the western part of Dotson Ice Shelf – are not visible in our fracture maps (Figs. h, f). This is because they have a distinctive visual representation in comparison to type-A and type-B crevasses, which the two neural networks used in this study are not tuned to detect. Future studies should seek to evolve the existing method to identify and map additional classes of crevasses across Antarctica, which will be a useful dataset for assessing different glaciological conditions.

After measuring the spatial pattern of fracture density change in the Amundsen Sea sector, we investigated where these changes might be important for the dynamics of the glaciers. To evaluate how changes to the local stress distribution might impact the glacier at large, we employed the fractional difference between the second principal component of the vertically integrated viscous stress tensor e2 and the vertically integrated hydrostatic pressure as a local notion of “buttressing” ; Appendix : 1κ=1-e2×12ρ̃gh2-1, where ρ̃=ρi(1-ρiρw), ρi is the density of ice, ρw is the density of sea water, g is the acceleration due to gravity and h is the ice thickness.

By weighting the fracture density change by this buttressing number as it was at the start of the observational period, we assessed where observed changes to fracturing might have a meaningful impact on the ice dynamics (Fig. e, i, m). For example, where a major increase or decrease in fracture density had occurred in a buttressing region of an ice shelf, we expect to also observe a dynamic change. Our observations show that the large increase in fracturing in the southern shear margin occurred in a strongly buttressing part of the ice shelf (Fig. e). Similarly, those changes to crevassing on the Thwaites Eastern Ice Shelf are amplified around the offshore pinning point when buttressing is taken into consideration, while changes on the main body of Crosson Ice Shelf are diminished.

Our results show that our method of assessing structural change (Sect. ) produces data that reflect the qualitative changes that are known to have occurred (Fig. ), like the deterioration of the southern Pine Island Ice Shelf shear margin and the advection of rifts on Crosson Ice Shelf. The results also show limited change where they are not thought to have occurred, for example on Cosgrove Ice Shelf (Fig. a). This inspires confidence in the data as a whole, including the additional unreported results such as the degradation of parts of the shear margins of Crosson Ice Shelf and the limited changes on its central trunk.

Generally, the uncertainties associated with our estimates of fracture density change (Fig. c, g, k) are similar in magnitude to the results. Of the ice shelves on which fracture patterns have changed, we have the greatest confidence in the results covering the central part of the Thwaites Eastern Ice Shelf, the southern shear margin of Pine Island Ice Shelf, and the central body and northern shear margin of Crosson Ice Shelf.

The quantitative estimates of fracture density change are likely to be dependent on the size of the region over which the fracture densities were calculated and, to a lesser extent, on the resolution of the grid on which measurements were made. The parameters were chosen such that the grid spacing was not larger than the spatial extent of particular features of interest – such as the Pine Island Ice Shelf southern shear margin. We took care to ensure that the sizes of the regions over which fracture densities were measured were not smaller than the distances over which crevasses could have been advected during the observational period. Future studies should intercompare different fracture observation products and investigate the sensitivity of the results to parameter choices and change estimates.

The usefulness of these results also depends on the linearity of changes in fracture density over the time period. For the most part, the assumption of slowly varying crevasse patterns allows us to see the most important structural changes in the region in our data. However, there are cases where important structural changes occur on short timescales and are not accurately captured when considering linear trends over longer, decadal periods. For example, a nearly flat signal is observed in fracture density near the grounding line of the Thwaites Glacier Ice Tongue despite the large oscillatory changes observed over the last decade (Fig. b) . Similarly, the method is insufficient when regions undergo rapid fragmentation, for example during the disintegration of the seaward part of the Pine Island southern shear margin in 2020. Here, the rapid changes in backscatter standard deviation led to a section of the fracture density time series being automatically discarded despite the observed jump in fracture density reflecting a real event (Fig. b and c).

Finally, due to the way we construct an ensemble of trend estimates, the results are likely to be biased towards small and negative changes in fracture density. This is because there is a component of the backscatter standard deviation time series due to the changing fracture pattern on the ice surface, with a greater number of high-contrast fractures generally increasing the standard deviation. Though our observations show this effect to be small compared to that of changing firn water content and the dependence of fracture density time series on fracturing on the ice surface, future work should aim to quantify this component and remove its effect from estimates of fracture density change.

Though the main aim of this article is to present the methods we have developed and some of the data generated, there are some immediately interesting features in the fracture density change maps (Sect. ) that merit further discussion, specifically regarding the Pine Island and Thwaites Eastern ice shelves.

Firstly, it is likely that the fracturing in the shear margins of Pine Island Ice Shelf (Fig. ) are due to increased shear strain rates as the ice shelf accelerated over the last 2 decades and a thickness deficit in the shear zone partially maintained by channelised melting of the sub-shelf . By invoking a local measure of buttressing (Fig. ), we have shown this crevassing to likely be, in turn, dynamically important for upstream ice. The relationship between fracture development and ice speed change is highly non-linear as crevasses caused by elevated strain rates directly impact the flow field by changing the constitutive ice rheology. Hence, our observations suggest that this crevassing in the shear margin is likely to be required to fully explain recent changes to the dynamics of this ice shelf . Though convincingly attributed the majority of the recent speed change to the calving of large tabular icebergs, we believe the full picture of dynamic change on the shelf will include the degradation of the southern shear margin.

The buttressing number we chose to consider can be thought of as a local measure of how the ice shelf differs from the archetypal freely floating, one-dimensional case . However, the net buttressing of an ice shelf on the grounding line is a highly non-local effect that depends on many factors that alter the transverse and longitudinal transmission of stresses across the boundary. It is difficult to link the two notions in most cases. However, it is clear that a pattern of increased crevassing across the entire lateral shear margin of a confined ice shelf will decrease the net buttressing of the grounding line. Hence, in addition to the dynamics of the shelf itself, we believe the observed structural changes in the southern shear margin are also required to fully understand changes to grounding line flux over the last decade .

An analysis of satellite images over the Thwaites Glacier terminus shows the Eastern Ice Shelf loosening from the pinning point and a shear plane developing between the two . The fracture density change map we see is consistent with a picture of an ice shelf transitioning from one that is slow-moving and heavily pinned, where fracturing occurs due to compressive stresses originating from the pinning point, to a freely floating ice shelf with crevasses forming and propagating perpendicular to flow by the action of internal stresses. This can be seen by the greater amount of fracturing at the eastern calving front and more regular fracturing perpendicular to the ice flow direction. Weighting by the buttressing number shows a stripe of increased fracture density close to the pinning point to be an important feature of structural change (grey arrow, Fig. i). This closely mirrors a feature that was identified by using an ice sheet model to infer changes to material properties of the ice as the aforementioned slip plane developed.

The benefit of using crevasse maps to assess structural change is that they can be used to derive quantitative information beyond that which can be gained by looking directly at satellite images. Our method achieves this, providing a scalar measure of structural change; we discuss in Sect.  a subset of the ways in which such a dataset could be useful in future work. Fundamentally, the use of the data relies on “fracture density” being a meaningful measure of structural weakness. Though it is not possible to find an exact mapping between fracture density and, for example, the continuum mechanics notion of damage , this is likely to be a good assumption in most cases. Though the metric is degenerate with respect to the orientation and size or number of crevasses, it seems unlikely that crevassed regions would naturally evolve between these degenerate states in a way that changes the dynamics of the glacier. Additionally, there may be cases in which dynamically unimportant changes in crevassing lead to changes in fracture density. This is a consideration, for example, in regions dominated by uniform longitudinal stress such as near the calving fronts of wide ice shelves. Here, the net impact of a field of parallel crevasses of equal depth is equivalent over large distances to a single crevasse of that depth. However, such regions are not highly buttressed, so structural changes are less likely to matter to the dynamics of the glacier as a whole. For example, the fracture density changes at the terminus of Pine Island disappear when weighted by buttressing number (Fig. b and e). Finally, we note that, as our fracture maps primarily locate crevasse walls, the widening of crevasses, which could have conceivable dynamic implications, is not measured as a change in fracture density.

Though our method of measuring the evolution of ice shelf crevassing is capable of resolving severe structural changes in fast-changing regions, few places in Antarctica are changing as rapidly as the Amundsen Sea Embayment . The results of applying this technique to the Antarctic Ice Sheet as a whole would likely be dominated by noise. In some places, such as Crosson Ice Shelf (Fig. j), the method may not be particularly useful where change is dominated by the advection of existing crevasses, and “interesting” changes are subtle in comparison. For example, Fig.  shows fracture density changes on the Brunt–Stancomb–Wills (BSW) and Shackleton ice shelves. Here, the stripes of positive and negative trends in fracture density show advection on the BSW Ice Shelf and in the Shackleton shear margin, though the steady propagation of Chasm-1 , for example, is difficult to discern. In future, we could apply Lagrangian corrections to fracture maps before looking at fracture density change or using a method that does not involve averaging over large regions and monitoring the size of isolated rifts.

Finally, future studies should take place investigating alternative means of reducing the component of fracture density time series due to the dependence of crevasse visibility on surface conditions. Our method uses backscatter standard deviation to alias consistent time series and build an ensemble of trends for each location. This is based on the observation that the fracture density and backscatter standard deviation signals appear to respond in the same way to changing firn or ice surface properties. It ought to be possible, for example using a regional climate model or reanalysis data, to better constrain the most salient factors in the environmental contribution to the fracture density signal. These can then be isolated using additional datasets to perform the procedure outlined here more accurately.

The physics of material failure is not well integrated into many large-scale ice sheet models but, through its distributed influence on stress conditions, this omission may have a large impact on predictions of the evolution of the ice sheet. Both the crevasse maps and fracture-density change maps introduced in this article have conceivable use in numerical modelling studies. Firstly, static, pan-continental mosaics of crevasses can be assimilated into numerical models to better constrain initial control parameters for simulations. Specifically, models often require a control parameter specifying changes to the effective ice viscosity from that provided by the ice constitutive relation . Given that the presence of fractures is assumed to change the effective ice viscosity, the fracture data can be used to constrain an inverse problem aiming to infer such a field from ice speed data. On grounded ice, these maps could be particularly useful in reducing the underdeterminedness of inversions for both ice softness and basal friction.

It is also conceivable that the smooth, scalar fracture density change maps we produce could be used more directly in diagnostic modelling. In continuum mechanics, a scalar “damage” parameter is often used to represent the reduction in effective ice viscosity caused by the presence of fractures . The dynamic impact of structural change for glaciers of interest, e.g. Pine Island Glacier, could be investigated by using fracture-density-changed fields as a proxy for damage change.

Additionally, data on the location of crevasses can be used to inform our understanding of the physical processes which lead to crevasse development. In particular, the healing of grounded surface fractures suggests their presence is a function of instantaneous stress conditions rather than a complicated and intractable stress history. This suggests the data could be used to constrain models seeking damage as a function of stress invariants, using stress fields inferred from coincident ice velocity observations. At a more basic level, such data, in combination with accurate vertical stress profiles, could be used to study the fracture toughness of meteoric glacier ice in the interior of the Antarctic Ice Sheet.

We turn now to a discussion of synthetic aperture radar data as the base dataset for this study. We have shown that a large variety of crevasse-like features are visible in the SAR backscatter images acquired by Sentinel-1 (Fig. ) and, using parallel processing of the images for type-A and type-B features, that most of these can be automatically extracted in a way reliable enough to promote discussion of existing crevasse patterns and to measure important structural changes on ice shelves. However, there could be choices other than SAR as the base dataset for this work that have the desired spatial coverage, in particular, the use of optical satellite data or a high-resolution digital elevation model (DEM). The further need for a time series of fracture maps leaves optical data as the only appropriate alternative candidate, although this would preclude year-round monitoring during the polar night.

Many but not all of the features we are interested in can be seen in optical imagery of comparable resolution to our SAR data. Crucially, crevasses corresponding to type-B features can rarely be seen reliably in optical imagery, except from in the most high-resolution (<10 m) cases. Even then, they are often bridged by snow. Hence, the use of SAR data allows for the mapping of crevasses on grounded ice, which are essentially exclusively type-B. Additionally, optical imagery, in contrast to SAR data, is hampered by the presence of clouds and cannot acquire images at night or during large parts of the austral winter. SAR data are therefore preferable for generating a consistent, reliable time series which can be used to measure changes to crevassing over relatively short timescales. Of the SAR satellites available, we consider Sentinel-1 to be the best tool due to its pan-continental large-scale acquisition plan which acquires new images every 6 to 12 d. Though it does not have as fine a spatial resolution as some other SAR satellites, such as TerraSAR-X, its short repeat period and extensive spatial coverage allow a consistent time series to be generated that covers the whole Antarctic margin.

There are, however, certain disadvantages to the use of SAR. Some ice shelf crevasses appear smoother on the surface than in optical imagery. For example, the crevasses on Dotson Ice Shelf, which appear faintly in SAR backscatter images at 50 m resolution, can be seen more clearly in the MODIS mosaic of Antarctica (MOA) image over the same region . Due to layover and shadowing effects stemming from the fact that SAR images are reconstructed from range distances rather than incidence angles of received radiation, there are additional issues in geolocating crevasses. This is most important for type-B features, where the geocoding errors induced by these effects can be of the order of a crevasse width. However, these are likely to be small in comparison to errors induced by deviations in the surface height through time from those given by the digital elevation model. Most importantly, optical data often come with multi-band information such as different colours in the visible spectrum. We use single-band SAR data due to constraints on the amount of throughput data required to generate pan-continental mosaics. However, as discussed above, the type-A processing fails in the presence of melt ponds. Optical data, in which the water appears blue, or multi-band SAR with dual polarisation, would be useful input to a neural network trying to discern fractures from linear boundaries of these pools of water. Finally, the use of a DEM might allow for the simultaneous extraction of crevasse-location and crevasse-depth data, the latter of which can only be estimated from SAR data.

The detection of crevasses using maps of interferometric fringes , phase coherence or strain rate fields comes with additional valuable information on the “activity” of crevasses – where their presence induces discontinuities in certain properties of the ice (e.g. its flow speed). Similarly, the use of precision altimetry data from ICESat-2 provides some information on the depth of crevasses . The method presented in this work does not provide any such information but, as a consequence, provides more extensive maps of crevasses with much greater coverage. As such, these data are useful for large-scale analyses, studies in regions with low interferometric coherence or imperfect velocity observations, and studies where the activity of the crevasses is of secondary importance – for example as a source of surface-to-bed hydrological pathways. Additionally, a dataset of crevasse locations such as that provided here can be used in combination with such methods to learn more about the importance of different crevasses across the continent. For example, it would be simple to locate ICESat-2 observations coincident with identified crevasses to assess some measure of depth, with the understanding that a reliable estimate of true depth is unattainable in many cases due to the sub-resolution width of the crack tip, or to look for discontinuities in ice shelf flow speed at crevasse locations in coincident velocity observations.

The method of mapping crevasses presented in this work includes different processing chains for detecting the features we call type-A and type-B. This is necessary because of their different appearance in the backscatter images but has the added benefit of providing an independent dataset for each fracture type, which allows for almost independent analysis of crevasses on grounded and floating ice. In order to constrain models of fracture development on ice shelves, it would be useful to further partition the type-A crevasses into basal and surface crevasses. This could be done by tuning the existing type-A network with a small dataset of manually annotated basal and surface crevasses separately and running them in parallel. This would also help solve the current insensitivity of the data to those features which might correspond to basal crevasses.

Improvements to the accuracy of the crevasse mapping can in all likelihood be gained without a great deal of work with the use of SAR data at multiple resolutions. We have been able to present accurate crevasse maps with large spatiotemporal coverage, in part by limiting processing to the use of data at a single spatial resolution, with our choice of 50 m based on a trade-off between the detail at which crevasses can be seen in the data and the finite capacity for computational throughput at our disposal. However, a greater number of crevasses can be seen at 10 m resolution, beyond the limit of what can be achieved with Sentinel-1 backscatter data, which, for example, would allow for more accurate bounding of regions of type-B crevasses.

Our discussions (Sect. ) suggest improvements can be made to the detection of basal crevasses and in discriminating between crevasses and the boundaries of surface melt ponds using multi-band input. We believe the greatest improvements could be seen by combining input SAR data with any available coincident optical data. This could be used to bolster static maps, where time series are not required. Additionally, we note that the sensitivity of the crevasse maps is not enough to capture subtle changes to crevasse density and length. This can be tackled in two ways: increase the sensitivity or increase the timescale of processes of interest. A potential method for the first is to work directly with time-series data. If a neural network were designed to receive as input a sequence of images and to segment the central one, the persistence of features of interest could be learned by the model, leading to more consistent time series. To tackle the second problem, it is feasible to train the network to segment older satellite imagery, e.g. from RADARSat, ENIVSAT or ERS1 and ERS2. This would enable a more extensive analysis of the longer-term (decadal) patterns of change in crevassing over the Antarctic continent.

The crevasse maps presented in this work provide unified coverage over the extent of floating and grounded ice in Antarctica imaged by Sentinel-1. As far as the authors are aware, there are no existing methods with publicly available datasets or code for the large-scale detection of Antarctic surface crevasses on grounded ice. However, methods for crevasse detection on floating ice shelves do exist. We conduct a brief comparison, for a single Sentinel-1 image frame, between the results from the method presented in this work and two publicly available existing datasets and methods, namely those of and . We refer to these respectively as “L20” and “I23” and to the method presented in this article as “S23”. L20 used U-Net to extract ice shelf crevasses from optical data covering the AIS at 125 m resolution, while I23 used a method based on the normalised radon transform that can be applied to data from different sensors at different resolutions. We applied I23 to a Sentinel-1 backscatter image (Sentinel-1 PF 7/913, dated 7 June 2021) at 50 m resolution using a window size of 10 pixels and a normalisation range of -30 to 0 dB. These are compared with the June 2021 mosaic over the extent of the aforementioned SAR backscatter image. This location (covering the Crosson and Dotson ice shelves of the Amundsen Sea sector of West Antarctica) was chosen again for the variety of crevasse features in the image. Results are shown in Fig. . Satellite images from which the data were generated are shown in Fig. a and d, and the derived data are shown in panels b, c and e. The SAR image (a) corresponding to data shown in panels b and c is from 7 June 2021, while the MODIS image (d) from which e was derived is constructed from images dating between 2008 and 2009. The outputs corresponding to L20 and S23 are in the range [0,1], while the range for I23 is a parameter that is chosen based on the windows size and image resolution to tune the output to best represent fractures visible in the input image. In this case, we set the bounds to [0.01, 0.1]. We show in Fig.  a “fracture score” that uses the same colour map for the outputs of each method but with different bounds.

Overall, with bounds for the range of the I23 dataset to [0.01, 0.1], we see a relatively good agreement between S23 and I23. Both pick up the large-scale patterns of fracture on Crosson Ice Shelf and display relatively little on Dotson Ice Shelf (Fig. a–c), though the features in S23 have a higher contrast with lower background noise. The resolution of the I23 output is defined by the window size over which radon transforms are calculated and, with a choice of 10 pixels, S23 improves on this resolution by an order of magnitude, though the comparison might not be fair given the possibility of using I23 with different, higher-resolution sensors. The S23 output is dominated by type-A crevasses, though closer to the grounding line there are fields of type-B fractures in the input SAR backscatter image that can be seen in S23 but not in I23. Finally, we note that the time taken to process this SAR frame was significantly higher in the case of I23 than our method when running with an equal number of processes and the same CPU hardware.

The differences between S23 and L20 are partly due to the evolution of Crosson Ice Shelf in the period 2009 to 2021 (visible in Fig. a and d), such as the advection and extension of a large central rift in the ice shelf and the degradation of the northern shear margin that borders Bear Island. However, there are also fundamental differences between these datasets that will influence their future application. Firstly, there is a preference for L20 to extract wider, smoother features over sharp rifts or deep, disordered fractures in the shear margins, while the reverse is true for S23. This leads to a near reversal in the features that are detected between the two methods. On Crosson Ice Shelf, the large central rifts visible in Fig. d do not appear in L20 (e), while similar rifts are prominent features in panels a–c. Additionally, the flowband features in L20 (many of which appear not to be crevasses) do not appear in S23. On Dotson Ice Shelf, however, L20 contains a great deal of basal-crevasse-like features visible in panel d which appear only faintly if at all in S23.

Overall, considering ice shelves as well as grounded ice, we would advocate for the use of the data and method presented in this article or a combination of the data presented here and those of when a single fully comprehensive map of crevasse features is required. This is especially true for continent-wide studies as our data do not cover the majority of the Ronne–Filchner and Ross ice shelves, unlike those of . However, we acknowledge that the comparison presented here is largely qualitative, covers only a small area, and includes only three datasets. Hence, future work should aim to conduct a more exhaustive comparison of existing crevasse-mapping products over a large and varied set of times and ice shelves to explore more completely the benefits and disadvantages of different methods.