Speedup and fracturing of George VI Ice Shelf, Antarctic Peninsula

Abstract. George VI Ice Shelf (GVIIS) is located on the Antarctic Peninsula, a region where several ice shelves have undergone rapid breakup in response to atmospheric and oceanic warming. We use a combination of optical (Landsat), radar (ERS 1/2 SAR) and laser altimetry (GLAS) datasets to examine the response of GVIIS to environmental change and to offer an assessment on its future stability. The spatial and structural changes of GVIIS (ca. 1973 to ca. 2010) are mapped and surface velocities are calculated at different time periods (InSAR and optical feature tracking from 1989 to 2009) to document changes in the ice shelf's flow regime. Surface elevation changes are recorded between 2003 and 2008 using repeat track ICESat acquisitions. We note an increase in fracture extent and distribution at the south ice front, ice-shelf acceleration towards both the north and south ice fronts and spatially varied negative surface elevation change throughout, with greater variations observed towards the central and southern regions of the ice shelf. We propose that whilst GVIIS is in no imminent danger of collapse, it is vulnerable to ongoing atmospheric and oceanic warming and is more susceptible to breakup along its southern margin in ice preconditioned for further retreat.


Background
In recent years, several Antarctic Peninsula (AP) ice shelves have undergone dramatic and rapid retreat (Cook and Vaughan, 2010)    that highlighted the on-going cryospheric response to atmospheric  and oceanic (Martinson et al., 2008) warming. Ice-shelf stability on the AP has been linked to the southward migration of a critical atmospheric thermal boundary by a number of previous studies. Morris and Vaughan (2003) remarked that the timing of ice-shelf collapse events were closely linked to the 5 −9 • C mean-annual isotherm that, during 2000, stretched from the Wilkins Ice Shelf embayment, across Alexander Island and George VI Ice Shelf (GVIIS), before traversing along the AP to Jason Peninsula north of Larsen C Ice Shelf (Fig. 1). It has also been noted that ice shelves respond to variations in oceanic temperature and circulation (e.g. Shepherd et al., 2003;Holland et al., 2010;Bindschadler et al., 2011; Consequently, the rate at which grounded ice discharges into the ocean is increased and thus their contribution to global sea level is amplified. Despite the uncertainty surrounding the stability of the remaining AP ice shelves, few studies have considered their long-term structural and dynamic evolution, even though Mercer (1978) and Vieli et al. (2007) both recognised that glaciological changes oc- 15 curred well in advance of breakup phases. Here we use optical, radar and laser altimeter satellite remote sensing data to: (1) assess the spatial and structural evolution of GVIIS from ca. 1973 to ca. 2010, (2) calculate multi-annual flow speeds of the iceshelf surface from ca. 1989 to ca. 2009, and (3) analyse surface-elevation change from ICESat GLAS data between 2003 and 2008. We use these results to highlight glacio-shelf varies in thickness from 100 m at the northern ice front to 600 m in the central region, before thinning again towards the southern ice front (Talbot, 1988;Lucchitta and Rosanova, 1998;Smith et al., 2007).
The ice shelf catchment covers much of the eastern coast of Alexander Island and the western margin of Palmer Land, with 12 km 3 a −1 and 46 km 3 a −1 of ice estimated to 15 be flowing into GVIIS, respectively (Reynolds and Hambrey, 1988). Ice from Alexander Island only extends a few kilometres into the ice-shelf system (Lucchitta and Rosanova, 1998) as tributary glaciers are generally small (between 54 km 2 and 144 km 2 ; Humbert, 2007). Glaciers flowing from Palmer Land are typically larger and supply much of the ice to GVIIS (Humbert, 2007). This domination of inflow from Palmer Land pro- 20 duces stagnation points along the ice shelf created as ice flows towards the opposite grounding line, but diverges prior to reaching this point (Reynolds and Hambrey, 1988). Bentley et al. (2005Bentley et al. ( , 2011 and Roberts et al. (2008) both show that during the mid-Holocene, George VI Sound was absent of shelf ice, and thus the present day ice shelf is still considered vulnerable to environmental change despite its atypical dynamic con- 25 figuration.

Ice-shelf mass balance
Wintertime snowfall on GVIIS rarely lasts through the summer season due to high surface melt rates, particularly in the northern regions where extensive meltpools develop over an ice-shelf area of ∼ 5900 km 2 , each year (Wager, 1972;Ridley, 1993;Smith et al., 2007) which subsequently refreezes on the ice-shelf surface during the austral 5 winter (Reynolds, 1981). The ice shelf thus consists of largely consolidated ice, fed from inland glacier systems (Humbert, 2007). Ablation occurs almost entirely as a result of seasonal-frontal calving and basal melting (Pearson and Rose, 1983;Reynolds and Hambrey, 1988;Lennon et al., 1982), with Potter and Paren (1985) suggesting that the ice fronts of GVIIS advance period- 10 ically before calving along rifts that penetrate the entire depth of the ice shelf. Mercer (1978), Doake (1982), Lucchitta and Rosanova (1998), Smith et al. (2007) and Cook and Vaughan (2010) used a combination of historical accounts and satellite imagery to document the fluctuation of the ice-shelf margins. Between 1947Between and 2008Between , 1939 of ice was lost from the northern and southern ice fronts combined, with no significant 15 advance (Lucchitta and Rosanova, 1998;Cook and Vaughan, 2010). Despite these studies, there is little analysis of the spatial or temporal patterns of retreat over time.
GVIIS has high basal melt rates (e.g. Jenkins and Jacobs, 2008;Holland et al., 2010). Warm water intrusion from the Circumpolar Deep Water (CDW) current, originating from the south east Pacific basin, flows onto the continental shelf and extends 20 underneath the entire length of GVIIS, contributing significantly to basal melt (Potter et al., 1984;Potter and Paren, 1985;Talbot, 1988;Lucchitta and Rosanova, 1998;Holland et al., 2010). This process is thought to be linked to the strength of the Antarctic Circumpolar Current (ACC) and a cross-current bathymetric low (Klinck and Smith, 1993). The circulation involves dense saline CDW water advecting from Marguerite Bay 25 beneath GVIIS via its northern ice front. Upwelling of warmer water instigates basal melting which is first deflected westwards by the Coriolis force, and subsequently advected northwards, completing the cycle (Potter and Paran, 1985;Smith et al., 2007). The maximum observed oceanic temperature from within George VI Sound is +1.1 • C, 3 • C warmer than the freezing point at the base of the ice (Talbot, 1988 in order for GVIIS to remain in equilibrium; recent basal-melt calculations suggest spatially-averaged losses of 2.8 m a −1 , 4.1 m a −1 , 3.0 m a −1 and 6.0 m a −1 (Corr et al., 5 2002;Jenkins and Jacobs, 2008;Holland et al., 2010;Dinniman et al., 2012, respectively) revealing that the ice shelf is currently estimated to be in negative mass balance, correlating well with the sustained thinning rates reported by Fricker and Padman (2012) and Pritchard et al. (2012). Here, we assess surface elevation changes with respect to spatial, structural and dynamic configurations of the ice shelf to investigate 10 ice-shelf response to recent climatic changes. Reynolds (1981) and Reynolds and Hambrey (1988)

Structural and spatial assessment
Structural and spatial mapping was carried out in ArcMap 9.3 Geographical Information System (GIS) software following similar procedures to Glasser and Scambos (2008) and Braun et al. (2009)

Interferometric SAR
InSAR procedures were carried out in GAMMA Remote Sensing software using two ERS-1/2 24-h repeat image pairs over the northern ice front for ca. 1995; one image pair was used to construct an ascending-pass interferogram, with the other image 5 pair used to construct a descending-pass interferogram. Velocity fields of the ice-shelf surface were then resolved via a trigonometric approach between the ascending and descending interferograms, assuming that the surface dynamics had not altered between the different image pair acquisitions (28/29 October 1995, 15/16 February 1996, and that flow was in the horizontal plane. Twenty-four hour velocity fields were scaled 10 up to annual displacements for comparison with the manually-derived feature tracking data. Atmospheric effects and baseline estimation errors are likely to be the largest inherent contributors to uncertainty in the resulting data (Mohr et al., 2003) but are no greater than 5 m a −1 . Errors caused by vertical ice-shelf motion (tide, atmospheric pressure) are estimated to be no more than ± 30 m a −1 using the fringe-rate visibility at 15 the grounding zone and the incidence angles of the ascending and descending SAR passes.

Manual optical feature tracking
Following the work of Simmons and Rouse (1984) and Simmons (1986), manual feature tracking was used to calculate surface speeds of GVIIS. We opted for manual fea-20 ture tracking methods over automated alternatives due to the considerable variations in surface feature scale, the ability to map displacements between multiple coincident scenes, and the capability of tracking features through fine clouds or atmospheric haze. Optical datasets were selected for manual feature tracking as the visual quality of the imagery is far superior to that of SAR data. Furthermore, by using Landsat data the 25 temporal resolution was extended back to 1986, thus creating an approximate 25-yr  window to assess changes in flow speed. In total 40 Landsat scenes (see Supplement) were used over time periods of no more than three years between image pairs. Images were coregistered to within 1 pixel accuracy with mapping also carried out to an estimated accuracy of 1 pixel, thus resulting in a total maximum uncertainty of 2 pixels between image pairs 5 Surface features were selected based on their distinctness between image pairs and their distribution across the image. Generally, fractures and rifts were used, although pressure ridges were also tracked where they could be identified clearly in both images. A polyline was digitised from the feature's starting point on image 1 to the same point on image 2 with the total polyline length (i.e. feature displacement) calculated 10 and added to an attribute table within a GIS. The centre point for each polyline was resolved and similarly added to the attribute table as X and Y coordinates and later used as the interpolation point. The displacement measurements were then converted from absolute displacement to displacement in metres per annum. Normalising the measurements in this way permitted data collation between different image pairs. Next, 15 the normalised displacements were merged into a single shapefile, representative of total displacement in m a −1 , for that particular time period. The data points were subsequently interpolated using a Natural Neighbour algorithm, selected over alternative algorithms as it performs equally well with regularly and irregularly distributed data (Watson, 1992). 20 This manual feature tracking approach works particularly well where distinct structures are densely packed rather than over featureless terrain where data points are more sparsely distributed. However, we assume that any large variation of flow is represented by visible surface structures (e.g. shear zones, pressure buckling) and indeed changes in these structures over time (fracture development/propagation). We 25 therefore suggest that where speeds have been derived from interpolation between sparsely-distributed points, the true speed is within the boundaries of the stated uncertainties. 172 m along-track in 33-day campaigns two to three times per year. Since ice-shelf analysis requires accurate, tide-corrected elevation measurements, the GLAS data (GLA12, release 531) was "retided" by adding back the GOT99.2 tide correction (Fricker and Padman, 2006). This dataset was then converted to a WGS-84 ellipsoid and then 10 corrected for vertical tidal displacements using the more accurate CATS2008a model (Padman et al., 2002;King and Padman, 2005), the inverse barometer effect (IBE, Brunt et al., 2010), and inter-campaign biases (Siegfried et al., 2011). After applying corrections, we filtered data from campaigns affected by clouds by inspection of gain and return energy values on a track-by-track basis and resampled each track to an 15 ad-hoc reference track (Brunt et al., 2010), which allowed for repeat-track analysis.
Because we were only interested in elevation changes, the absolute accuracy of the GLAS data is inconsequential. We determined a region-specific precision of our GLAS data using crossover points over a larger dataset also including Bach and Stange ice shelves on the AP (Holt, 2012) and iteratively removing outliers (following the methods 20 of Brenner et al., 2007). The calculated single-shot precision of 0.153 m (1 standard deviation, n = 30) is slightly higher than the calculated precision of ICESat data from early campaigns (Shuman et al., 2006), which is expected as the precision over ice shelves accounts for instrumental uncertainty as well as uncertainty in the IBE, tide, and inter-campaign bias corrections. 25 Following the pre-processing of the GLAS data, each track was converted to a Polar Stereographic projection to match other datasets in this study, and then subset to the ice-shelf area in hydrostatic equilibrium using the extents calculated by Brunt 383 (2010) to ensure that any variations in surface elevation reflected changes in iceshelf thickness from either surface or basal accumulation/ablation. Next, GLAS data were further culled, removing any points potentially affected by the motion of ice-shelf surface undulations (i.e. fractures, rifts, pressure ridges, longitudinal structures) so that the resulting surface-elevation changes reflected a true elevation change rather than 5 a change in position of surface structure. This process was similarly undertaken by Horgan et al. (2011), although instead of manually removing data, here the structural maps for ca. 2003 and ca. 2010 were used to automatically select and remove the elevation points in ArcMap within 200 m of any distinct surface feature. Finally, surface-elevation changes were calculated between comparable campaigns 10 (i.e. on 12-monthly timescales or multiples of) to avoid elevation changes caused by short-term seasonal ablation and/or accumulation. The campaigns that offered the best spatial and temporal resolution for each individual track were selected, and whilst intercampaign elevations were not used in the final elevation-change assessments, a record of surface elevation changes were subsequently obtained that reflected a sub-decadal 15 change. Repeat surface-elevation measurements were acquired over 16 ICESat tracks from a total of 4323 points and are displayed as change in m a −1 . Datasets produced from these methods are summarised in Table 1, with a full list of satellite images used given in the Supplement.

Ice-front retreat
Between 1974 and 1979, approximately 820 km 2 of ice calved off the north ice front (Figs. 2b, 3). Further retreat was recorded along the Palmer Land grounding zone between 1979 and 1989, 1989 and 1996, and from 1996  with the centre of the ice front retreating 40.6 km into George VI Sound. Post 1979, the retreat rate of the north ice front fluctuated between −10 km 2 a −1 and −20 km 2 a −1 .
Furthermore, a concave profile was observed at the north ice front in each of the observation periods. At the south ice front, approximately 925 km 2 of ice was lost, with 182 km 2 of ice

Structural glaciology and structural evolution
Mapped structures and surface features (January 2010) are displayed in Fig. 2, with a full list of glaciological structures, identifying criteria and their significance provided 25 in Table 2. Here we describe the structural evolution of the north and south ice fronts only.

Structural evolution at the north ice front
During 1974, ice flowing from Palmer Land into GVIIS was heavily fractured and could be traced towards the north ice front, becoming increasingly chaotic (Fig. 4a, b). Brash ice was observed between some of the larger rifts, whereas in the central portion of the north ice front, large rifts were filled with smooth, uniform ice with some evidence 5 of open water running along their lengths. These central rifts could be traced back intermittently (at ∼ 4 km distances) to the northern boundary of Riley Glacier whilst decreasing in their overall length. Rifts that eventually formed the ice-front during 1979 and 1989 were also visible ∼ 25 km from the 1974 ice margin. Chaotic fracturing and rifting was also observed along the Palmer Land margin of the northern section during 1979, 2001 and 2010 (Fig. 4). During 2001, several distinct rifts had developed along the Alexander Island margin, the largest of which had propagated across the ice shelf for ∼ 20 km that later facilitated iceberg calving between 6 January 2010 and 29 January 2010. This event was captured by the Landsat ETM+ sensor, but detailed observation is limited due to poor data quality and thus not shown here. 15 Few fractures were observed along the north ice front during March 2010, with the exception of some small rifts at each of the grounding zones (Fig. 4h).

Structural evolution at the south ice front
During 1973 (Fig. 5a, b), ice feeding the centre of the south ice front was dominated by cross-cutting fracture traces, emanating from fractures and rifts lee-side of the Ek- 20 lund Islands further upstream. The largest rifts were filled with unconsolidated brash ice, with occasional small pools of open water. Fractures and rifts also formed along the English Coast margin immediately downstream of ice rises and headlands. These spread for up to 20 km into the ice shelf before becoming indistinct from the surrounding ice. 25 In both of these regions, fractures and rifts continued to propagate throughout the observation period assessed here. During 1986 (Fig. 5c, d), 1991 (Fig. 5e, (Fig. 5g, h) and 2010 (Fig. 5i, j) the area of fractured shelf ice increased towards a retreating ice front, with more rifts showing evidence of open water or unconsolidated ice along their length. Furthermore, an area of open water lee side of the Eklund Islands, first apparent in 1991 (Fig. 5f), increased in size from 1991 to 2010 with the ice bridge that connected SIF1 and SIF2 eventually breaking away during March 2010, leaving 5 two independent ice fronts terminating into the Ronne Entrance.

North ice front
The flow of the north ice front is principally controlled by Riley Glacier (Fig. 6). Between ca. 1989 and ca. 2009 the surface speed of ice derived from this particular glacier

South ice fronts
The south ice fronts (Fig. 7)  zone of GT04, GT05, GT06 and GT07 towards the ice front. During ca. 2009, the maximum recorded velocity in the centre of SIF2 was 796 m a −1 .

Surface-elevation changes
Surface elevation change in the northern section of GVIIS was calculated over three tracks with a total of 1017 repeat measurements examined (Fig. 8). A non-significant 5 (less than our uncertainty) negative change is observed in the data, with pockets of positive elevation change interspersed between areas of surface lowering; there are few distinct patterns in the dataset.
In the central section, six GLAS tracks cut across GVIIS, with 1657 repeat measurements analysed (Fig. 8). A complex pattern of surface-elevation changes is observed, 10 with positive changes noted at the input of Goodenough Glacier (track 0063) and towards the northern extents of tracks 0197 and 0263, near the input of Kirwen Inlet. Large negative elevation changes are recorded along track 0382, which dissects central region of the ice shelf. Tracks 1298 and 0316 both illustrate ice-shelf wide negative surface elevation changes. 15 Widespread negative surface-elevation change has also been calculated over 1649 repeat measurements in the southern section, interspersed with localised pockets of positive elevation changes (Fig. 8). Due to fracturing and rifting towards the south ice front, however, the GLAS dataset was heavily filtered, and therefore direct measurements of surface elevation changes are incomplete towards the English Coast and 20 around the Eklund Islands.

Grounding-zone retreat
The analysis of sequential satellite images highlighted a gradual retreat of the grounding zone between 1973 and 2010 at the southern extent of GVIIS along the English Coast of Palmer Land (Fig. 8). In total, 172 km 2 of ice-shelf area was affected by the re-Introduction and the western boundary of GT07. There is no evidence of grounding zone retreat elsewhere along either Palmer Land or Monteverdi Peninsula, although the reducing surface visibility of various ice rises also points to a retreating grounding zone around the Eklund Islands from which is it inferred that the southern region has experienced widespread thinning since at least 1973. features had gradually developed, forming at the grounding zone through flow-induced longitudinal extension and/or bending stresses caused by tidal motion, before propagating across the ice shelf towards the north ice front (Fig. 4b). Further rifting between 1996 and 2001 eventually led to a large calving event during January 2010, although rift development this time developed from the western margin near Alexander Island 5 rather than from Palmer Land. Most of the retreat at the north ice front was thus governed by preconditioned and longstanding active rifts, developed well in advance of actual calving. Indeed, the eventual 1979 and 1989 ice fronts were observed as rifts in 1974 Landsat MSS imagery (Fig. 4), thus emphasising the long-term development of the retreat process the north ice front. As of March 2010, no ice-shelf-wide rifts existed 10 in this region, thus no immediate large-scale calving is anticipated. The retreat recorded between 1974 and 1979 in the northern region occurred in an area devoid of significant glacier input and thus we suggest that the retreat rate of the north ice front is largely controlled by its dynamic configuration. Elsewhere on GVIIS, areas of fractured ice are observed between coalescing flow units (Fig. 2). These areas 15 tend to exhibit longitudinal extension at the grounding zone, causing initial fracturing, with stress regimes becoming more compressive with increased distance from the iceshelf margin due to the coalescing of ice-flow units. Thus the switch from an initial tensile regime to a compressive stress regime within a limited spatial extent restricts fracture and rift propagation and indeed prompts the sealing of fractures. North of Ri-20 ley Glacier, coalescence with a second tributary glacier did not occur, thus ice did not undergo longitudinal compression after initial fracturing. It is proposed that as a result of this dynamic configuration, fractures emanating from Riley Glacier at the grounding zone are able to spread transversely across the ice shelf as it flows towards the northern ice front with little or no resistance, thus preconditioning the northern ice front for TCD 7,2013 Speedup and fracturing of George VI Ice Shelf, Antarctic Peninsula

Southern extent of GVIIS
At the south ice fronts, Smith et al. (2007) indicate a continued retreat since 1947, with substantial ice loss between ca. 1967 and 1973. From observations in this study, only a small portion of the south ice front (adjacent to Monteverdi Peninsula) was found to be advancing, yet even this area appears to show a repeated advance/calving regime, 5 with further calving anticipated along the ice-front rift that has progressively developed since ca. 1996 (Fig. 5).
The overwhelming pattern of the south ice front is that of steady retreat, concentrated in the centre of the ice front. It is inferred from visual assessment of 1973 Landsat imagery that this region has a lower surface elevation lee-side of these ice rises. Thick-10 ness calculation from Radar Altimetry (RA) elevation data (e.g. Griggs and Bamber, 2011) confirm this. To the north and west of the Eklund Islands the ice shelf is thicker, but heavily fractured.
The presence of the Eklund Islands towards the southern ice front is therefore a critical component in the retreat characteristics of GVIIS. First, their presence causes 15 regular fracturing and rifting of the ice shelf as it flows around (or over) such features, due to a shift from high-compressional stresses upstream to tensile stresses downstream. Second, the composition of shelf-ice, lee-side of ice rises and ice rumples has elsewhere been shown to have a higher concentration of warmer, marine-derived ice, accreted in basal cavities and incorporated into the ice shelf through the resealing of 20 fractures with flow (Fricker et al., 2001;Khazendar and Jenkins, 2003). Whilst considered to be less brittle than meteoric ice (Lui and Miller, 1979;Jansen et al., 2010), the warmer (Vieli et al., 2007), marine-derived ice is more susceptible to oceanographic variations (Fricker et al., 2001). Third, the Eklund Islands have a profound effect on regional surface speeds and thus the supply of ice to this region. Feature-tracking 25 measurements from each of the observation periods reveals substantially slower flow speeds down-ice of the Eklund Islands than those observed outside of this region; the ice rises/rumples essentially act as a buttress to flow. Furthermore, during 1973, the  Doake et al.'s (1998) criteria for irreversible retreat, displaying a largely concave profile. A combination of these factors made large portions of the south ice front pre-conditioned for iceberg calving and continued retreat as observed between 1973 and January 2010. Towards the central and southern extents of GVIIS, a stronger surface-lowering sig-5 nal is measured from which widespread ice-shelf thinning is inferred. In the central region in particular, the ice-shelf draft of GVIIS reaches much greater depths than at any other point in George VI Sound, and thus it may be subjected to warmer waters that tend to exist closer to the sea bed (Jenkins and Jacobs, 2008;Holland et al., 2010). Thus, the thickest parts of GVIIS are subjected to high rates of basal melt and lower 10 rates of basal accretion that results in a net loss of ice through a vertical column. It appears that an increase in the rate of retreat of the south margin between 1991 and 1996 immediately follows a strong vertical mixing within the Bellingshausen Sea between 1989 and 1992 (Holland et al., 2010). This pattern in also reflected in the timing of ice front retreat of Bach and Stange Ice Shelves (Holt, 2012) that promotes the idea 15 that short-term, intermittent oceanic variation can impact on the glaciological conditions of ice shelves. Subsequently, a warming Ronne Entrance portion of the Bellingshausen Sea coupled with the inferred high concentrations of marine-derived ice lee-side of the Eklund Islands and a structurally weak ice shelf preconditioned the south ice fronts for further retreat. Therefore, we suggest that the south ice front in particular responds to 20 changes in ice thickness caused by oceanic temperature variation.  Glasser et al., 2011;Berthier et al., 2012), and whilst Riley Glacier still contributes to the mass of GVIIS, the continued retreat of the north ice front has effectively removed a buttress, reducing back-stresses and subsequently permitted an increase in surface velocity at the northern margin. Further upstream, flow speeds of Skinner, Chapman, GT12 and Transition Glaciers do not show an increase or decrease over time, and thus 5 further illustrate the dominance and controlling nature of Riley Glacier on the dynamic regime of the north ice front. At the south ice front we propose that the removal of buttressing ice between 1973 and 2010 led to increased longitudinal extension through the reduction of back-stresses within the ice shelf. This is clearly reflected in feature tracking measurements between 10 ca. 1989, ca. 2002 and ca. 2009, with ever-increasing surface speeds between the Eklund Islands and De Atley Island over time. Indeed, between the Eklund Island and De Atley Island, localised flow speed had almost doubled from ∼ 380 ± 30 m a −1 to ∼ 780 ± 15 m a −1 . These increases appear to be driven by two different processes: (1) increased extensional pulling from the ice front; and (2) acceleration of GT07 from its 15 grounding zone. There is a clear link between the increase in flow speed over time and the increased distribution and extent of surface fractures, rifts and fracture traces that make this particular area structurally and dynamically weak. Between Monteverdi Peninsula and the Eklund Islands, however, an advance of the ice front is driven by tributary glaciers some 150 km back upstream. We measure no 20 significant (greater than our uncertainty) variation in flow speeds over time, and as a result, the area between the Eklund Islands and Monteverdi Peninsula appears to be more stable and less responsive to observed environmental and glaciological changes. 25 The retreat history of the north ice front since ca. 1940 alludes to periodic large-scale breakup of heavily-fractured ice that stretches for some distance back upstream from the ice front. The multidecadal recurrence interval over which fractures develop and propagate allows the anticipation of ice-loss extent well in advance of actual calving. Thus based on these historical observations, no immediate large-scale calving is expected from the north ice front, as few fractures or rifts currently exist. Iceberg calving will most likely be governed by regular, but discreet calving at the ice front. 5 Our data illustrate that the structural regime at the north ice front is governed by the dynamics of individual tributary glaciers that flow from Palmer Land. It has been shown that where there is a lack of longitudinal and transverse compression (due to the absence of flow-unit confluence), fractures and rifts are capable of propagating across the ice shelf, along which large-scale calving is initiated. The short bursts of rapid retreat are linked to the breakup of comparatively "less-active" zones of shelf ice, with retreat rate much less where the ice front is directly supplied by fast-flowing "active" ice. The latter situation currently exists (Fig. 6), with iceberg calving and retreat rate governed by the dynamics of Riley Glacier. The current flow configuration south of Riley Glacier has a comparatively slower, less-active dynamic regime absent of significant glacier 15 input. We propose that this area is currently exposed to high back-stresses by the presence of Riley Glacier, limiting longitudinal extension that ultimately restricts fracturing. Continued recession of the north ice front (over any timescale) would reduce the back-stresses upstream and encourage fracturing of shelf ice across an approximate area of 1200 km 2 towards Millett Glacier, potentially rendering this entire section weak 20 and susceptible to large scale retreat as observed between 1974 and 1979.

The north ice front
Surface melting on GVIIS in the northern region has been on-going since at least 1940 (Stephenson and Fleming, 1940) yet the compressive flow regime of the ice shelf has limited the effect of surface meltwater on its structural stability. Furthermore, the northern limit of surface meltwater has always been south of the ice front, but in recent 25 years it has been shown to expand towards a retreating northern margin. As a result, a combination of structural weakening (longitudinal extension), abundant meltwater and intensifying surface-melt brought on by warmer temperatures (e.g. Vaughan  (Torinesi et al., 2003), creates an environment susceptible to atmospheric warming. Recent modelling studies (e.g. Holland et al., 2010) have suggested that the rate of basal melt in the northern region increased from ∼ 1.8 m a −1 to ∼ 2.0 m a −1 since 1980. Whilst comparatively high relative to the other ice shelves on the west Antarctic 5 Peninsula, basal melt rate remains marginally less than the 2.1 m a −1 rate suggested by Potter et al. (1984) in order for the ice shelf to remain in equilibrium. The northern region of GVIIS is therefore not losing sufficient mass from its base to instigate widespread thinning. Whilst surface melting is evident for prolonged periods of the austral summer, the absence of sufficient surface drainage means that meltwater re-10 freezes on the surface of the ice shelf and therefore the actual mass lost is considered negligible. As a result, the northern section of GVIIS is perhaps less susceptible to ice-shelf thinning, although Fricker and Padman (2012) comment that surface-lowering terminated on Wilkins Ice Shelf approximately 8 yr prior to the two most recent breakup phases in 2008 and 2009, and thus a lack of vertical change may not necessarily point 15 to a stable ice shelf. The northern region, however, is susceptible to regular surface melting regardless of whether mass is lost from the ice shelf or not. The stabilising characteristic is its unusual dynamic regime that will ultimately control the spatial and temporal retreat/ breakup patterns. 20 The south ice fronts have seen by far the greatest change in recent decades as a response to climatic and oceanic variability. Prior to March 2010, this margin underwent sustained retreat, concentrated in the central portion that we inferred to be both thin, and have a high concentration of marine-accreted ice. This sustained ice loss (horizontal and vertical) has rendered large portions of the southern margins structurally weak 25 and prone to iceberg calving along preconditioned fractures and rifts. The removal of the central portion of the ice front consequently split it into two individual ice fronts, that even despite their close proximity, have distinctly different glaciological regimes. The negative surface-elevation changes observed to date is linked to increasing basal melting as a result of a warming ocean (Holland et al., 2010;Pritchard et al., 2012). Assuming that this warming will continue, the southern region of GVIIS is particularly vulnerable to heightened retreat as a result of increased basal melting and a reduction of in-situ basal accretion. 5 The stability of SIF1 is governed by its frontal geometry and thickness-driven velocity from further upstream, particularly towards Monteverdi Peninsula. This portion has historically shown prolonged periods of advance followed by large-scale calving, and indeed, calving is anticipated in the coming years along a well-developed rift. Even after this expected loss of ice, the ice front would remain relatively stable. Being convex 10 over much of its length, it is expected that it would also maintain a similar speed as it has done in previous decades as it is seemingly non-responsive to retreat patterns; its dynamics are driven by its thickness gradient between the ice front and its tributary grounding lines.

The south ice fronts
Towards the Eklund Islands, SIF1 is structurally weak. Large areas are heavily fractured and rifted and the recent retreat history illustrates that these are most susceptible to further ice loss, particularly as increasing basal melting and reduced basal accretion (Holland et al., 2010) will limit basal accumulation. Furthermore, the ice melange that fills the rifts is likely to reduce, and with it, any resistance to further propagation (Bassis et al., 2005). Within the next decade it is probable that the ice directly in front of these 20 ice rises will have disappeared, leaving the ice shelf pinned behind them. Ice feeding SIF2 has seen the greatest dynamic and structural changes and indeed is most likely to undergo further retreat. In particular, it remains vulnerable to oceanic warming and intermittent vertical mixing of CDW as noted by Holland et al. (2010). The positioning of several ice rises close to the grounding zone, in addition to those further 25 downstream, almost certainly create basal cavities, that not only results in thinner regions of shelf ice, but also encourage basal accretion of marine ice. As a result, ice feeding SIF2 is likely to have a high concentration of marine-derived ice that naturally results in these thinner areas being more susceptible to oceanic variation.  (Fig. 9), many being filled with an ice melange that elsewhere has been shown to limit further propagation (e.g. Bassis et al., 2005). Even within the last decade, the number of open rifts with visible water has increased, and as this region is outside of that which is subjected to abundant surface melting, it is clear that this is open sea water. A warming environ-5 ment (climatic and oceanic) has effectively begun to remove the stabilising ice melange in these rifts. As a result, future widening of these rifts is almost certain to occur as they progress through the ice shelf towards the front, increasing the rate of retreat and also increasing the likelihood of rapid breakup phases. 10 Our analysis of the glaciological evolution of GVIIS has revealed its vulnerability to ongoing environmental change in the AP. The north and south ice fronts occupy two very different environmental settings and thus the glaciological controls of, and responses of, ice-front retreat vary considerably. Analysis of optical, radar and laser altimeter remote sensing datasets has led to the following conclusions: 15 1. Spatial and temporal retreat patterns observed at the northern ice front between 1974-present were controlled by longstanding and widespread ice-shelf fractures. Rapid breakup (1974Rapid breakup ( -1979 was followed by steady retreat (1979-present) along rifts near the ice front, and coupled with a concave ice front profile; the northern region was arguably preconditioned for sustained retreat. At present, there are 20 few channel-wide fractures that have previously hastened calving from the north ice front, and thus we do not anticipate immediate large-scale retreat.

Summary and conclusions
2. The retreat rate recorded at the south ice fronts was steadier than the north ice front, although there is evidence of increased retreat rate following periods of warm-water upwelling in the Bellingshausen Sea. Iceberg calving at the south ice 25 front was focused in areas where the shelf was inferred to be thinner, was heavily 397 fractured, and where high concentrations of marine-accreted ice were inferred. We suggest that the southern ice front was also preconditioned for retreat between 1973 and 2010, although at present both SIF1 and SIF2 have more stable, convex ice-front geometries.
3. Surface velocities at the north and south ice fronts increased over time. In both cir- 5 cumstances this has been attributed to reduced back-stresses within the ice shelf that permitted enhanced longitudinal extension. Between the Eklund Islands and De Atley Island in particular, this occurred simultaneously with increased fracturing and rifting that has rendered this area structurally weak. The dynamics of the northern ice front are at present dominated by Riley Glacier that buttresses shelf ice further upstream; the longevity of Riley Glacier thus governs the glaciological responses of the rest of the ice-shelf system in the northern region.
4. In the central and southern regions of GVIIS widespread surface lowering was recorded. We link this to enhanced basal melting rather than surface ablation and thus conclude that the southern extents of GVIIS are more susceptible to on-15 going oceanic warming. Coupled with grounding line retreat, we infer significant ice-shelf thinning.
5. Rapid disintegration of GVIIS is not anticipated, however, our investigation has shown that significant glaciological changes have taken place south of the proposed atmospheric limit of ice-shelf viability, and that the ice-shelf system is more 20 vulnerable to oceanic warming than atmospheric change. TCD Sci., 7, 197-198, 1995 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Table 2. Ice-shelf features, identifying criteria and significance. Adapted from Glasser and Scambos (2008) and Glasser et al. (2009). See Fig. 2  Formed when ice is compressed perpendicularly (or approximately) to the orientation of the original fracture and/or rift.

Longitudinal surface structures
Long, linear structures aligned parallel with the principal flow direction. Typically < 1 km in width but often exceeding tens-to-hundreds km in length. Observed as dark and light lines caused by shaded relief. Originate from the confluence of two flow units and in regions of positive relief, at bed protuberances or in regions of high basal friction.
Generally indicate regions of faster ice flow and depict suture zones of different flow units. Cumulative length is due to the slow decay timescale relative to the time required for ice to travel a long distance.

Transverse structures
Term given to sub-linear surface features where no clear method of formation exists. Appear as dark lines on the ice-shelf surface.
May indicate degraded surface fractures or surface undulations caused by ice-shelf buckling under compressive stresses. Significance discussed on a case-by-case basis. Pressure ridges Succession of dark and light linear bands appearing perpendicular to principal flow direction. Often located near bedrock or ice-rises or between coalescing flow units Formed by ice straining vertically under longitudinal compressive stresses.

Grounding zone
Sudden break in surface slope or area of intense crevassing. Meltwater ponds tend to form at the grounding zone where there is a change of gradient.
Junction between grounded ice and floating ice. A dynamic zone often flexing with tidal amplitude.

Ice rises
Elevation of the ice-shelf surface with disturbance to ice flow indicated by pressure ridges on the stoss-side and/or crevasses in the lee of the ice rise.
Local bedrock high where the ice shelf is grounded.
Ice rumples Elevation of the ice-shelf surface with disturbance to ice flow indicated by crevasses.
Local bedrock high where the ice shelf is partially grounded, but ice flow continues. Surface meltwater Dark, flat areas on ice-shelf surface either as open or closed systems. May or may not form along pre-existing structural discontinuities.
Indicates surface ablation and can suggest ice-shelf surface slope orientation where flow direction is apparent. Assessing time-series of surface meltwater can indicate increasing/ decreasing atmospheric temperatures.

Ice dolines
Large sub-rounded surface hollows often filled with meltwater during the height of the melt season.
May indicate a link between the ice-shelf surface and subsurface. Considered to be located in thin and weak ice. Formation linked to abundant surface meltwater. 7,2013 Speedup and fracturing of George VI Ice Shelf, Antarctic Peninsula  The names of tributary glaciers were taken from the Antarctic Place-names Committee (http: //www.antarctica.ac.uk/apc/) except for those labelled "GT##" that were otherwise previously unnamed. Note the positioning of the −9 • C mean annual isotherm across Alexander Island (source: Morris and Vaughan, 2003  Increased fracturing between the Eklund Islands and De Atley Island is also linked to increasing ice-shelf flow speeds recorded between ca. 1989 and ca. 2010 (Fig. 7).