The Wrangell–St. Elias National Park Service (NPS) commissioned a photogrammetric dataset; the data were captured by air on 13 September 2018 over the debris-covered terminus of Kennicott Glacier, from its confluence with the Root Glacier to the toe. In total, 1691 images were taken using a Nikon D850 with a 24 mm prime lens at an average flying altitude of 940 m a.s.l. These images were processed using Agisoft Metashape to produce a digital elevation model (DEM) at 0.32 m resolution and an orthophoto at 0.16 m resolution. Four ground control points were surveyed at the McCarthy Airport (61.4377° N, 142.9027° W) to calibrate the final DEM product, resulting in a +0.13 m vertical adjustment to the ellipsoidal heights. The photogrammetric dataset covers 33.8 km2 (37.2 km2 when corrected for slope) of the total 54.8 km2 of debris-covered ice on Kennicott Glacier and extends up to an elevation of 830 m a.s.l. on the glacier surface (coverage mapped in Fig. a).

Past efforts at ice cliff mapping have included manual mapping, using visible imagery at times supplemented by slope maps , and automated mapping, using spectral curvature , with object and image segmentation algorithms applied to visible imagery , and slope thresholds using a DEM as input . While manual mapping is expected to provide the highest-fidelity data product, automatic mapping is more feasible for large areas such as at Kennicott Glacier. Automated mapping via visible/spectral image analysis has the advantage that it can be applied to a wide range of orbital/remote sensing datasets. Automated mapping via slope threshold has the advantage of relative simplicity and is not sensitive to variable lighting conditions or albedo of ice cliffs; furthermore, it is based on the foundational principle of ice cliffs' existence – slopes greater than the angle of repose have no debris layer. We developed our own automated mapping algorithm based on the slope threshold method.

Our automated ice cliff mapping algorithm uses slope calculated from the photogrammetric DEM as follows. (1) Grid cells with slope angles greater than >30° (representing areas where the angle of repose is exceeded) were isolated in a mask. (2) This mask was converted to a shapefile with contiguous collections of grid cells as individual shapes. (3) Gaps <10 m2 in size contained entirely within ice cliff shapes were filled. (4) We then calculated the steepest-slope value contained within each individual ice cliff shape; a threshold minimum steepest-slope value was tested as a method of reducing false positive identification of ice cliffs.

As a validation for our automated ice cliff mapping method, we performed manual mapping of ice cliffs via visual identification on a subset of the orthophoto. The debris-covered terminus region of the glacier was sliced into 500 m × 500 m grid cells, and 11 of these were selected for the mapping validation, sampling 8 % of the total orthophoto (Fig. b). Automatically identified cliffs that coincide with manually mapped cliffs were marked as “true positive”, while those which did not were marked as “false positive”. We also calculated the F1 score , which combines metrics of false positive and false negative identification: 1F1=2TP2TP+FP+FN, where TP, FP, and FN are the true positive, false positive, and false negative fractions, respectively. Our goal is to balance false positive with false negative identifications while maximizing true positives.

To prepare the DEM for hydrologic analysis, it was downsampled to 2 m resolution in order to reduce computation time. Supraglacial streams often undercut ice cliffs and are thus not visible in the top-down view of the photogrammetry dataset; debris mounds which the stream channel meanders around (via ice cliff undercut) may be seen as hydrologic obstructions. For this reason, sinks <3 m in depth were filled in the DEM; this value was chosen as a cutoff based upon the typical roughness of debris-covered ice within stream corridors observed in the field on Kennicott Glacier. Stream channels were then predicted over the modified DEM using the Channel Network tool in SAGA GIS; this tool calculates runoff catchment area for each pixel in the DEM and returns channels defined by contiguous pixels with catchment area greater than a set threshold, defined in this case as 2500 m2. The chosen threshold was somewhat arbitrary but represents a potential channel flow of ∼1200–2300 cm3 s-1 (assuming a mean sub-debris melt rate of 4–8 cm d-1 and ignoring ice cliffs). Note that these are predicted channels and may not necessarily be filled with meltwater in all cases.

The validation grid cells used in ice cliff mapping (Sect. ) were also used for a validation exercise of automated stream channel mapping. Each stream channel delineated by the flow algorithm was examined in the orthophoto and categorized according to the following scheme: 1 – channelized/flowing water identified; 2 – channel morphology identified (no water); 3 – no clear channel identified but plausible (textured debris; association with hydrological features such as ponds and channels; part of dendritic network); 4 – implausible/no hydrologic morphology identified; 5 – error (e.g., channel within a supraglacial pond) (examples in Fig. S1 in the Supplement).

We then determined the spatial proximity between supraglacial streams and ice cliffs at 2 m resolution. For each grid cell representing 2 m length of a supraglacial stream, the distance to the nearest ice cliff grid cell was calculated, and vice versa. Distances shorter than 2.83 m represent adjacent grid cells and thus direct ice cliff–stream interactions.

We developed an idealized geomorphic model of ice cliff backwasting in order to analyze the effect of streams on ice cliffs as recorded in the DEM. This model focuses on the relationship between ice cliffs and their “backwasting ramps”. We define a backwasting ramp as the slope on a debris-covered glacier surface generated by the ice cliff melting vertically and horizontally into the surface of the glacier. While backwasting ramps have not previously been identified in the literature, this model is similar in concept to that presented by for ice sails (positive relief bare ice features on some debris-covered glaciers).

Consider an ice cliff at an angle of θ that backwastes through a two-dimensional debris-covered glacier surface (Fig. a). Backwasting is driven by melt orthogonal to the ice cliff face. Thus, as the ice cliff melts at a rate M it produces vertical incision (Mcos⁡(θ)), as well as horizontal wasting (Msin⁡(θ)), and scours out a backwasting ramp at an angle of π/2 from the ice cliff and ϕ=π/2-θ from the horizontal. However, incision at the base of the ice cliff is not expected to be fully efficient, due to the effects of differential insolation-driven ablation and reburial by debris. The presence of streams may also aid in the efficiency of incision. We define the angle ψ between the ice cliff and its backwasting ramp. ψ and the ramp angle ϕ are theoretically indicative of the efficiency of ice cliff incision and survival.

We first sought validation for our model of ice cliff backwasting and ramp scouring by investigating a time series of ArcticDEM data strips . The ArcticDEM is constructed from stereo pairs of WorldView and GeoEye satellite images processed using “Surface Extraction by TIN-based Search space Minimization” to DEMs with 2 m horizontal resolution. For the years 2013 (15 July), 2014 (21 July), and 2017 (9 August), available DEM products covering Kennicott Glacier during the summer months of June–August were extracted using Google Earth Engine. These were then coregistered by minimizing the mean elevation difference between three patches of flat, stable, and lightly vegetated glaciofluvial or moraine deposits in each DEM mosaic, applying a simple elevation offset to do so. From our investigation of stable terrain, we estimate the uncertainty due to horizontal shift between the DEMs at 2–3 m (on the order of the horizontal resolution).

We then extracted topographic profiles for analysis of 100 ice cliffs and their associated backwasting ramps, using the 2018 high-resolution photogrammetry DEM (profile locations shown in Fig. b and c). We sought to investigate differences in cliff–ramp geometry between stream-associated cliffs and non-stream-associated cliffs in the context of our geomorphic model. Ice cliffs were therefore categorized as those associated with streams and those with no associated streams. The stream channel prediction results were used to aid in this categorization, but the presence of active stream action was confirmed in each case via visual identification of water flow or ice cliff undercutting visible in the orthophoto. Great care was taken in sampling the backwasting ramps, such that our sampling was not affected by other disturbances such as surface collapses, debris movement, or other ice cliffs. The average slopes of the ice cliff and ramp were then determined, along with the corresponding value of ψ between them.

To assess the impact of topographic profile selection on the results obtained using our geomorphic model, we selected two of the sampled ice cliffs for which we extracted 12–16 topographic profiles along their length. Each profile was independently analyzed using our model, and by comparing the results, we determined an uncertainty due to the profile selection of 3° for ice cliff and backwasting ramp angles and 5° for ψ (Sect. S2; Fig. S2).

We performed a series of discharge measurements to determine the ability of supraglacial streams to erode and transport debris. We selected a stream network exhibiting numerous examples of cliff–stream interactions for these measurements. We used the area–velocity method , and nine measurements were performed at five locations (locations shown in Fig. ). Experiments took place during afternoon hours across 3 d in 2021 representing hot, sunny (13, 17 July) and cool, cloudy (14 July) weather conditions. Note that because of the timing of the measurements, the results are most indicative of diurnal maxima of flow conditions. Stream measurement locations were selected in reaches with linear laminar flow and minimal turbulence. The length of sampled reaches varied between 1.5 and 8 m.

Two to three flow-transverse gates were defined for each discharge measurement, and the channel cross section for each gate was measured by hand using a measuring stick at 10 cm intervals. The area for each discharge measurement was then taken as the average of the area for the measured gates.

A neutrally buoyant mandarin orange was timed as it traveled in the stream between the flow-transverse gates. Individual measurements were removed from the analysis when the orange experienced undue turbulence or friction by interacting with eddies, the channel sidewall, or obstacles such as rocks in the stream bed. For each experiment, 5–15 velocity measurements were made and averaged. Full information on individual reaches is summarized in Table S1 in the Supplement.

Average velocity across the gates is estimated at 0.9× the value of the measured velocity. This is consistent with a low-friction stream bed such as an artificial channel or in this case bare ice . The determined average velocity is then multiplied by the measured gate area to estimate discharge.

Without applying minimum steepest-slope criteria, 78 % of ice cliffs were identified as true positive, while 22 % were false positive (steep sections of debris-covered ice). Applying minimum steepest-slope criteria reduces the number of false positives, while the false negative fraction begins to increase as the minimum steepest-slope threshold is increased (Fig. f). A threshold value of 51.8° yields an area-averaged balance between false positives and false negatives at 4 % each. The maximum true positive fraction and F1 scores occur at 51.7°, with values of 93 % and 0.96, respectively. This threshold value is also consistent with the distribution of values from manually mapped ice cliffs; the maximum slope observed on 90 % of ice cliff shapes is >50° (Fig. e). We selected a final value of 51.8° for the minimum steepest-slope threshold in our automated ice cliff mapping routine.

This threshold for the minimum steepest-slope value in each candidate ice cliff shape reduces the number of false identifications from steep debris-covered slopes or fully reburied ice cliffs (Fig. ). Some shallower cliffs are lost as a result, but these are a comparatively small fraction of the population. Because partially reburied ice cliffs are liberally mapped by the automated slope method, the total area compared to the manually mapped results is overestimated by 13 %.

The automated ice cliff detection routine identified 10 306 ice cliffs with a total combined area, corrected for slope, of 5.22 km2 (planform area 3.95 km2). Ice cliffs thus account for 14.0 % of the 37.2 km2 sampled debris-covered surface (11.7 % when ignoring slope in the area calculation). Median ice cliff surface area (true area) is 105 m2 (interquartile range, IQR, 45–290 m2). Median surface slope on individual ice cliffs ranged from 32–54° (Fig. S3a). Cliffs were observed at all aspects, with broad peaks in the population facing NE and SW (Fig. S3c).

A total of 5892 supraglacial stream channels were delineated, with lengths ranging from 4–1241 m and a total cumulative length of 51.9 km (average stream density 1.4 km km-2 for the study area). Average stream length was 88 m, with a standard deviation of 79 m and median of 65 m (Fig. S3d).

In total, 30 % of delineated stream channels were associated with identified active streamflow, 44 % were associated with channel morphology, 19 % were identified as plausible but without clear channel morphology, 5 % were identified as implausible, and 1 % was identified as artifacts that occurred within large supraglacial ponds (see examples in Fig. S1). Note that channels without observed active streamflow may be recently abandoned or they may simply exhibit flow which is below the detection limit dictated by the resolution of the orthophoto (several pixels; ∼1 m) or is frequently in shade/hidden by ice cliff undercuts. Thus, it is not possible to accurately estimate the proportion of active vs. inactive/abandoned stream channels from these results.

These validation results indicate 74 % true positive identification of supraglacial streams. With an additional 19% plausible identifications, we determine a possible 6 %–25 % overestimation of stream length/prevalence across the glacier surface.

Stream channel locations frequently coincide with the base of ice cliffs (Figs.  and ). In total, 33 % of ice cliffs are directly adjacent to streams, while 50 % are within 10 m of streams (Fig. ). Cliffs adjacent to supraglacial streams tend to be larger than those at a distance from streams; 69 % of the ice cliff area belongs to ice cliffs adjacent to streams. Conversely, 20 % of the total stream length (evaluated in 2 m grid cells/stream lengths) is directly adjacent to ice cliffs, while 48 % of the stream length is within 10 m distance from the nearest ice cliff. Stream channel networks connect together numerous ice cliffs with sub-linear, sinuous, and crescent planform morphologies. Stream course meanders are frequently associated with distinctive crescent-shaped ice cliffs on the outside of stream bends, reminiscent of cut banks on river channel systems in sedimentary environments. Examples of such ice cliffs are labeled in each panel of Fig.  with black arrows. Crescent ice cliffs are also frequently observed at a distance of some tens of meters from delineated stream channels (labeled in each panel of Fig.  with red arrows), implying a historical connection to the stream channel system.

We selected a very large (∼40 m tall) ice cliff in the stagnant ice zone of Kennicott Glacier (lowermost ∼2.5 km where glacier velocities measured in ITS_LIVE fall to values similar to noise floor observed outside glacier outlines ) to investigate ice cliff backwasting through time. This minimized the relative impact of surface change due to glacier flow. As shown in Fig. b and c, the ice cliff can be seen backwasting and scouring out a ramp, in this case at an angle of ψ=110°, consistent with our geomorphic model. This ice cliff is unique among the large ice cliffs in Kennicott Glacier's stagnant zone in that it has a supraglacial stream and debris-covered ramp at its base and not the surface of a supraglacial pond.

We sampled 100 cliff–ramp systems with roughly equal shares of cliffs with and without streams to investigate the impact of streams on the backwasting. Two example ice cliff–ramp profiles are shown in Fig. a, taken from the same ice cliff. Profile b–b′ is taken from a location where the cliff has no associated stream at its base; to the north of the profile, the cliff appears to be in the process of reburial under debris. Profile c–c′ is taken from a location where a supraglacial stream channel is predicted and corroborated by geomorphic evidence in the presence of an incisional canyon that approaches and intersects the ice cliff from the east, as well as the sharp, likely undercut, base of the cliff. Profile c–c′ exhibits a sharper ψ value than b–b′, due to both a steeper ice cliff and a much steeper backwasting ramp. The results show that ice cliffs with streams tend to have lower ψ values as a result of steeper ramp angles for ice cliffs with similar pitch (Fig. d and e), providing evidence that supraglacial streams enhance the efficiency of ice cliff incision.

The cross-sectional area of individual stream discharge measurement gates ranged from 630–2480 cm2, with stream widths between 57–150 cm and maximum depths up to 12–35 cm. The thalweg, or deepest part of the channel cross section, was in nearly all cases located nearer to the channel wall contiguous with the base of ice cliffs (Fig. c and S4–S7). This is consistent with lateral stream meander towards ice cliff undercuts. Mean measured velocities ranged from 0.60–1.13 m s-1, with the standard deviation in measured values ranging from 0.03–0.13 m s-1. The resultant flow rates were determined between 0.06–0.21 m3 s-1 (Fig. ).

The streams sampled in the discharge measurements tended to be nearly continuously ice-floored, with little to no sediment resting on the channel bottom. When observed, sediments tended to be small accumulations of fine gravels/coarse sands collected in hollows or as small, highly mobile bedforms (ripples) or larger rocks protruding above the surface of the stream. These larger static rocks were avoided in discharge measurement location selection. The streams themselves tended to be highly turbid, with low visibility beyond 5–10 cm.

In situ observations find supraglacial streams frequently at the base of ice cliffs, consistent with remote sensing results (Figs.  and ). Streams undercut the ice cliffs through lateral meandering and incision. Abandoned stream channels reveal up to ∼2.5 m horizontal and ∼1 m vertical incision (Fig. a and b). In the case of repeated incision by multiple episodes of stream meander action, up to 3.5 m total vertical incision was observed, which is a significant local contribution to the total area of ice cliffs (Fig. c).

There are also observed examples in which the action of supraglacial runoff contributes to ice cliff generation. Small rivulets of runoff, generated by the melt from ice cliffs and/or catchments of debris-covered ice, produce local incision and meanders as they travel through the supraglacial debris layer. This small-scale incision generates ice exposures on the order of centimeters to tens of centimeters in height which have the potential to grow into much larger-scale ice cliffs. In one dramatic example, a time-lapse camera used in fieldwork was knocked over by the generation and backwasting of an ice cliff that grew from 10 cm height cut by a rivulet to 1 m height and independent of the original runoff system within 38 d (Fig. ).

The impacts of two supraglacial stream avulsion events on ice cliffs were directly observed on Kennicott Glacier in the summers of 2021 and 2022 (Figs.  and ). These avulsions, i.e., rapid abandonments of the stream channel and formation of new channels, took place as a result of stream meanders in the same stream course system approaching each other from opposite directions in a manner similar to oxbow lake production in meandering river systems. As the stream meanders approach each other, the ice cliffs associated with the cut bank side of the meander meet, producing an ice fin with a sharp crest. Once the ice fin thins sufficiently, the upstream meander breaks through and the stream avulsion occurs, abandoning the prior stream path between those two points. Rapid streamflow at the avulsion point leads to accelerated collapse of the ice fin. In some cases, stream undercutting can contribute to mechanical collapse of ice cliffs and fins as well.

Channels abandoned by stream avulsion events initially reveal the extent of stream incision under ice cliffs. However, the shape of the overhang eventually melts out, and in many cases, the ice cliff begins to be reburied under debris. This was observed for south-facing ice cliffs following avulsion A2 on 15 July 2021 (Fig. ).

It is clear from the results of our DEM analysis that supraglacial streams and ice cliffs have a strong relationship in terms of spatial proximity and geomorphology. Linear to crescent-shaped cliffs are often observed with a supraglacial stream at their base or within tens of meters (Fig. ). Half of observed streams by length are within 10 m of ice cliffs, and 32.7 % of ice cliffs have supraglacial streams directly at their base (Fig. ). determined that 38.9 % of ice cliffs in High Mountain Asia are stream-influenced, i.e., defined as being within 40 m of supraglacial streams. If we applied this broad threshold to our analysis, 91 % of ice cliffs on Kennicott Glacier are stream-influenced. Conservatively, one-third to one-half of ice cliffs on Kennicott Glacier are stream-influenced. These cliffs typically follow the arc of supraglacial stream meanders on the outside of stream bends, analogous to cut banks in sedimentary river systems. This morphology is likely indicative of ice cliffs that have hydrological processes as their primary control.

Supraglacial streams act on the surface of the glacier through two processes, namely thermoerosional incision of the ice and the erosion/transport of supraglacial debris. showed that because heat transfer between streamflow and ice channel walls is dependent on turbulence in a similar fashion to sediment transport, supraglacial streams produce meanders with a very similar planform geometry to alluvial rivers. This lateral, as well as vertical thermoerosional, incision allows streams to carve new ice cliffs into the glacier surface (Fig. ), as well as undercut and enhance the incision of existing ice cliffs (Fig. ). Indeed, we have shown evidence from our geomorphic model that the presence of streams measurably enhances the incision of ice cliffs into the glacier surface, aiding in ice cliff survival (Fig. d). More pronounced lateral migration of streams under ice cliffs leads to more exposed ice and enhanced backwasting at the ice cliff base, as well as possible mechanical collapse (Fig. ). This is the mechanism by which stream channel meanders may control the morphology of ice cliffs, developing crescent-shaped cliffs on their outside bends.

Crescent cliffs adjacent to supraglacial stream systems often appear to be formerly associated with stream channel meanders in a manner reminiscent of oxbow lakes adjacent to sedimentary river systems . Observation of multiple stream avulsion events during 2021 fieldwork confirmed that stream abandonment of ice cliffs occurs due to both meander pinch-off or lateral migration (Figs.  and ). These streams leave behind ice cliffs with a morphology formerly influenced by stream action. Ice cliffs abandoned by streams lose the undercut morphology at their base and may be re-buried by debris, particularly if they are Equator-facing (Fig. ; ). This observation further underlines the importance of streams to ice cliff survival.

If the vertical incision component of a supraglacial stream is sufficiently greater than the average local melt rate of sub-debris ice and ice cliffs, it may lead to the production of a cut-and-closure channel whereby the stream enters the englacial hydrological system . This process becomes inevitable if the supraglacial stream reaches sufficient size. Most significant supraglacial streams observed during Kennicott Glacier fieldwork 2020–2023 eventually entered the glacier via a moulin or cut-and-closure channel. As debris thickness increases and sub-debris melt rates are reduced, relative stream incision increases, and cut-and-closure channels are more easily produced by smaller streams; this is one of the major reasons that supraglacial hydrological networks tend to be smaller on debris-covered glaciers compared to bare glaciers . This process may reduce the importance of streams to ice cliff morphology for glaciers with thicker debris. Meanwhile, the presence of englacial channels contributes to ice cliff formation upon their collapse, and the morphology of resultant ice cliffs will reflect that of the englacial stream course.

In sediment-bedded streams, the ability of the stream to erode (mobilize from rest), transport (carry within the flow), or deposit (drop out of the flow) clasts of a given size as a function of stream velocity has historically been described by a set of empirical relationships illustrated by the Hjulström curve (; Fig. S8). Comparison of our discharge measurement results to this diagram suggests the ability to erode debris clasts up to 1 cm in size and transport debris up to 10 cm in size.

However, because our supraglacial streams are generally bedded by ice and not sediment, the Hjulström diagram is likely to be ill-suited for the estimation of the debris erosion/mobilization threshold. If we assume a smooth ice-floored stream bed, we can predict the onset of sediment transport (erosion) with a simple force balance. We may consider the drag force Fdrag required to overcome the static frictional force Ffriction of a debris clast resting on the ice in a stream channel: 2Fdrag=ACwρwatervflow22,3Ffriction=μFN=μρrockVrockg, where A is the cross-sectional area of the rock orthogonal to flow direction, Cw is the drag coefficient, ρwater and ρrock are the densities of water and rock, vflow is the stream velocity, μ is the static coefficient of friction (0.05 for ice near 0 °C; ), FN is the normal force, Vrock is the rock volume, and g is acceleration due to gravity. At the critical drag force, Fdrag=Ffriction, we can determine the maximum size of a rock of a certain shape that is mobilized by the flow.

We assumed a rock density of 2700 kg m-3 and used a half-sphere (Cw=0.42) and a cube (Cw=1.05) as end-member shapes to calculate the diameter of debris clasts that our measured stream velocities are capable of mobilizing. We found our measured velocities (0.6–1.13 m s-1) capable of moving rocks up to 9–30 cm (half-sphere) and 14–51 cm (cube) in diameter. These values match or exceed our measured channel depths.

These results are consistent with the fact that little static sediment accumulation is observed in the measured stream channels. Where observed, static sediment size is typically the order of or larger than the stream depth or it is gravel and sand collected in hollows in the rough stream bed (our force balance assumes a smooth bed). Clasts larger than stream channel depth have also been observed in motion in ice-bedded streams.

This analysis thus supports that supraglacial streams are highly effective at removing debris from the base of ice cliffs, which may aid in preventing their reburial. The relative importance of debris erosion and transport in comparison to thermoerosional undercutting of the ice cliff face is unknown.

Because ice cliffs contribute significantly to melt runoff, there is additionally the opportunity for a feedback loop between ice cliffs and streams. More exposed ice cliff area leads to higher flow and incision rates with more power to create and maintain ice cliffs. Greater debris thickness leads to a higher differential melt between sub-debris and ice cliff melt rates, increasing the relative importance of such a feedback. At the same time, lower sub-debris melt rates lead to significantly reduced runoff from catchment areas without ice cliffs and a reduced ability of runoff to generate new cliffs.

Thicker supraglacial debris may slow the development of new stream networks by inhibiting runoff. At the same time, the relative impact of existing supraglacial streams will be increased due to reduced sub-debris melt rates. This will lead to greater impacts on ice cliffs for the same stream size, but larger streams are progressively lost to cut-and-closure conduit formation, as discussed in Sect. . Larger clasts are less effectively removed by streams from the base of ice cliffs; they may induce meandering, causing the stream to locally abandon the cliff or lead to deeper lateral undercutting. The presence of a sufficient volume of clasts larger than the stream can transport may lead to infilling of the stream channel and ice cliff reburial. Such abandoned stream channels, buried by debris, have been observed in the literature . Periods of higher flow due to exceptional melt, rainfall, and glacial outburst flooding could mobilize larger clasts and lead to accelerated reorganization of the supraglacial and englacial hydrological system.

While this study is focused on investigating the processes by which supraglacial streams control ice cliff formation and change, there are a number of other processes which also have major impacts on ice cliffs. Glacial hydrology more broadly can impact ice cliffs through other processes. The collapse of englacial conduits may lead to ice cliff production as a result of the exposure of former conduit walls . Supraglacial ponds often are rimmed in part by ice cliffs , with the pond playing a key role in altering ice cliff surface energy balance and undercutting through enhanced subaqueous melt rates . Ice cliffs may also be generated through the modification of crevasse walls. used geomorphic relationships to determine that supraglacial streams, crevasses, and supraglacial ponds are the primary controls on ice cliffs in High Mountain Asia, in order of decreasing importance.

On Kennicott Glacier, supraglacial streams are prevalent across the majority of the active portion of the debris-covered tongue. The importance of streams are diminished (1) where active crevassing occurs and runoff is thus sent rapidly to the englacial system via moulins or (2) where stagnant ice results in flattened surface topography and the dominance of supraglacial ponds. At Kennicott Glacier, the ice becomes stagnant ∼2.5 km from its terminus at an elevation of ∼500 m a.s.l. (, Fig. ); below this elevation, streams are rare, and ponds are dominant , while above this elevation, streams are more common (Fig. ).

We also observed evidence that local patterns in ice dynamics play a major role in ice cliff formation. Based on the ITS_LIVE velocity dataset derived from LANDSAT and other records , we observed two episodes of ice flow acceleration from 1988–1995 and 2018–2020. These occurred on the western half of the glacier from 2.5–7 km upvalley of the terminus (Fig. a; Supplementary Video 3). Velocities quadrupled in the area of interest for the most recent surge from 23 m yr-1 (2014) to 98 m yr-1 (2019) (they roughly tripled for the previous surge from 30 m yr-1 in 1986 to 95 m yr-1 in 1990). This “mini-surge” caused a minor localized advance and topographic bulge at the western glacier margin but had no impact on the glacier terminus. Field observations showed that the area coinciding with the strongest speed-up was heavily crevassed in 2019–2020 (Fig. d), with freshly exposed ice cliff faces. These new ice cliffs, originating as crevasse walls, can be identified by their morphology in our ice cliff inventory by their elongated linear morphology and tendency to occur in parallel sets (Fig. c).

Thus, for this area, ice dynamics supersede supraglacial hydrology as the dominant geomorphic driver of ice cliff formation and evolution. The crevassing and ice cliff formation related to the flow acceleration may be partly responsible for the higher percentage by area of ice cliff coverage in this study compared to that constrained in 2011 data by (14 % vs. 12 %). Debris-covered glaciers which experience surges may exhibit ice dynamics as the primary control on ice cliff formation. Interestingly, found that a surge which terminated upstream of debris cover on Urdok Glacier led to enhanced meltwater routing to the surface and thus enhanced ice cliff formation due to the stream water incision of cryo-valleys. Dynamic, hydrologic, and other drivers of ice cliff change may be further linked through other complex feedbacks.

A number of other studies have noted a close geospatial linkage between supraglacial streams and ice cliffs on other glaciers around the world, including Zmuttgletscher , Trakarding Glacier , and glaciers in High Mountain Asia . Our work shows the processes by which streams contribute to ice cliff maintenance and production at Kennicott Glacier; these processes are likely universal and applicable to other debris-covered glaciers such as those listed above.