2.1. ICESat GLAS data analysis

Satellite altimeter data provide the best method for observing elevation changes in the large regions of the Earth’s ice sheets (Reference ZwallyZwally and others, 2002; Reference HerzfeldHerzfeld, 2004; Reference WinghamWingham and others, 2006, Reference Wingham, Wallis and Shepherd2009). Radar altimetry (e.g. ERS-1, ERS-2 and CryoSat) has the advantage of being able to penetrate clouds, which are frequent in the Arctic and Antarctic regions, whereas laser altimetry has the advantage of higher resolution but penetrates only thin clouds and is affected by thicker clouds and other atmospheric conditions. The GLAS system aboard ICESat has a footprint of 70 m and records point elevation every 173 m along-track (Reference ZwallyZwally and others, 2002; Reference Schutz, Zwally, Shuman, Hancock and DiMarzioSchutz and others, 2005). GLAS data provide the best available observations of ice-surface elevations through 2003-09. However, because of technical problems (loss of a laser, L1, early in the mission) the remaining lasers, L2 and L3, were only operated for two or three 33 day submissions per year to extend mission life. Track spacing depends on latitude (see Fig. 4 for the PIG region), nominal spacing is 22.93 km at 75° S for the 33 day missions, but actual track spacing may be two to three times that, because data losses are often caused by atmospheric conditions.

Fig. 4.

PIG and PIIS elevation differences during GLAS observations: (a) over 3 years, February 2006 (L3E) – October 2003 (L2A); (b) over 3 years, May 2006 (L3F)–October 2003 (L2A); (c) over 2 years, May 2005 (L3C)–October 2003 (L2A); (d) over 1 year, May 2006 (L3F)–May 2005 (L3C); and (e) between seasons, May 2006 (L3F)–February 2006 (L3E). (f–j) Same as (a–e) respectively, with GLAS ground tracks from both missions superimposed (earlier mission tracks in red, later mission tracks in black). Surface-elevation contours of the earlier year/time superimposed for reference. Note that the grounding line (hinge line) can be identified from the break in surface slope, as seen in the change in density of contours. UTM coordinates. Elevation change in meters.

The figure of 22.93 km represents the spacing of ascending (or descending) tracks along the circle of 75° latitude, as well as the nominal width of the diamond-shaped gaps. Nominal spacing for the originally planned 91 day repeat orbit is one-third of the spacing of the 33 day repeat orbits, 7.64 km at 75° latitude. The tracks repeat only approximately, which makes it difficult to derive Δh/Δt without knowledge of cross-track slope.

Because of the above-mentioned difficulties in data distribution, the approach we adopt is to interpolate each submission dataset first, then form differences of digital elevation models (DEMs) (Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others, 2008). We derived elevation and elevation-change maps from ERS-1 data and early GLAS data collected near the beginning of NASA’s ICESat mission. Results showed that GLAS allowed elevation-change measurement from satellite altimetry with unprecedented accuracy and that PIG retreated at an increasing rate. The first two objectives of this paper are to extend this time series and to present a spatio-temporal analysis of the continued evolution of elevation changes in PIG.

Reference Pritchard, Arthern, Vaughan and EdwardsPritchard and others (2009) used ICESat release 28 GLAS data from February 2003 to November 2007 to derive maps of Greenland and Antarctica, which show extensive thinning on the margins of the ice sheets. They fitted planar surfaces to parallel tracks that are close in space (<300 m) and time (<2 years) and interpolated. The method sacrifices temporal resolution to gain spatial resolution of elevation change. Reference Pritchard, Ligtenberg, Fricker, Vaughan, Van den Broeke and PadmanPritchard and others (2012) concluded that the primary cause for Antarctic ice sheet loss is basal melting of the ice shelves, due to interaction with warmer ocean water. Lack of buttressing, resulting from ice-shelf loss, causes acceleration of the ice streams. Wind forcing could also play a role. Their paper presents thickness changes for all of the major Antarctic ice shelves and neighboring grounded ice for 2003–08, using a time series of repeat-track satellite laser altimetry. Ice-shelf thinning is reported as up to 6.8 m a ^{– 1} in the Amundsen Sea sector. Maps in Reference Pritchard, Ligtenberg, Fricker, Vaughan, Van den Broeke and PadmanPritchard and others (2012) did not resolve the changes in PIG spatially or temporally. The first conclusion of the literature review is that a spatio-temporal analysis of elevation changes in PIG has not been presented previously. This motivates the derivation of a time series of maps from GLAS and ERS-1 for 1995–2007, which is the first objective.

2.2. Elevation and velocity changes from other datasets, and grounding line retreat

PIG is a focus area of glaciological research because changes predicted by authors 40 years ago are happening now. PIG has been losing elevation and increasing in velocity faster than predicted and at an accelerating pace. Reference Shepherd, Wingham, Mansley and CorrShepherd and others (2001) analyzed ERS-1 and ERS-2 radar altimetry and SAR interferometry to show that the grounded PIG thinned up to 1.6ma^–1 between 1992 and 1999, affecting 150 km of the inland glacier, and concluded that the thinning cannot be explained by short-term variability in accumulation and must result from glacier dynamics. The map they presented is a collection of colored dots at crossover points, which makes spatial analysis difficult. Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others (2008) derived much higher thinning rates for the time interval 1995-2003 (3.125 m a–¹ for 25 m difference, 6.25 ma–¹ for 50m, and 12.5 ma–¹ for 100m elevation difference, which depends on location), and surface lowering reached 50-100m in a ring segment 20-60 km from the 1995 hinge line (Fig. 3a; the hinge is identified by a break in surface gradient). Even with the lowest rate, thinning has increased considerably in 1995-2003 compared to 1992-99. It is not clear whether this process is continuing, which motivates the second objective of this paper.

Reference RignotRignot and others (2008) reported that ice-sheet loss increased 59% in 10 years in the Amundsen and Bellingshausen sea sectors of the Antarctic ice sheet, using 1992-2006 SAR interferometry along the margin of Antarctica to estimate total ice flux into the ocean.

Reference Wingham, Wallis and ShepherdWingham and others (2009) used ERS-2 and Envisat satellite radar altimeter data to examine spatial and temporal changes in the thinning rate of PIG in 1995-2006, applying a method of dual cycle crossovers. They provided maps of thinning rates in 1995 and 2006 and confirmed that thinning increased (e.g. to 6ma–¹ in 2006 just above the grounding line). The main trunk of PIG and tributaries thin, but the slow-moving ice in the adjacent interior appears to thicken slightly. Reference Wingham, Wallis and ShepherdWingham and others (2009) assumed that the thinning is caused by ice drawdown propagating inland. If thinning continues to increase at the rate they derived, then the entire PIG trunk may become afloat in 100 years. Their analysis is based solely on widely spaced elevation-change points, which is a limitation when trying to identify spatial patterns.

Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others (2008) determined a grounding line retreat of ~20 km between 1995 and 2003 from ERS-1 and GLAS data analysis. The grounding line is determined as the break in slope between the ice shelf and glacier, as seen in the density of contours that are superimposed in elevation and elevation-difference maps in Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others (2008) and in this paper for different years. The 1999 grounding line, identified from differential SAR interferometry, is found in Reference Rignot, Mouginot and ScheuchlRignot and others (2011), and the 2009 grounding line is sketched in Reference Joughin, Smith and HollandJoughin and others (2010).

Reference Park, Gourmelen, Shepherd, Kim, Vaughan and WinghamPark and others (2013) investigated the position of the hinge line of PIG, based on 1992-2011 interferometric SAR data from ERS-1 and ERS-2 and found accelerated thinning (0.53±0.15 ma–¹) at the terminus with comparable rates upstream, which is consistent with increased melting in a cavity below the ice shelf. They conclude that future contribution to sea-level rise may be higher than previously thought, because the glacier-ocean system is not in a state of equilibrium.

To determine the continuation of elevation change and grounding line retreat in PIG beyond 2003, we extend the time series to 2007. To study the evolution of these processes, we break the time series into annual and seasonal time steps, as much as GLAS data coverage allows (Sections 4 and 5).

2.3. Attribution of causes of the retreat

The changes in PIG may be caused by external forcing or internal forcing:

In an external-forcing scenario, break-up of part of the ice shelf in Pine Island Bay leads to loss of back-holding power that affects PIG. As a result, the glacier accelerates, first in the parts that are closest to the grounding line.

In an internal-forcing scenario, changes in the glacier or in its accumulation basin explain elevation and thickness changes. These may or may not affect the ice shelf. Changes in the glacier may have dynamical causes (acceleration) or climatic causes (changes in accumulation, or increase in surface melting). Changes in basal conditions may lead to changes in the glacier’s flow.

Despite the fact that the early debate on the future of change processes in the WAIS evolved around surging or slower glacial retreat, most more recent papers have attributed the observed changes to processes in the ice shelf. Reference Payne and VieliPayne and Vieli (2003) analyze how events in the ice shelf may lead to changes in the glacier and support an external-forcing hypothesis. Reference Rabus and LangRabus and Lang (2003) support an internal-forcing hypothesis and attribute observed break-off of large icebergs off the ice shelf to chance rather than an external mechanism.

Based on comparison of 1995 ERS-1 data and 2003 GLAS L2A data, Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others (2008) concluded that retreat of PIG is attributable to internal processes in the glacier. Comparison of a map based on 2003 GLAS data with a map derived from 1995 ERS-1 radar altimeter data clearly indicates mass loss in large parts of lower PIG and Pine Island basin, and far less change in the floating part of PIG tongue and the ice shelf in Pine Island Bay. The onlapping tongue of grounded ice seen on the PIIS, which was distinct in 1995 (Fig. 2a and b), has become shallower and narrower, while not shorter, in 2003. (The onlapping tongue is the tongue-shaped area above 40 m centered at 450000 E/ –8340000N in Figure 2a, identified by a break in slope seen in the density of contours; the tip of the arrow ‘Pine Island Bay’ in Figure 1 points at the same location.) The observations of elevation-change distribution indicate dynamic thinning, which is internally forced, as the most likely cause. One possible internal mechanism is a dynamic wave that propagates down-glacier. Acceleration in the glacier in 2003 was largest in the lower regions, where the observed thinning was greatest (Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others, 2008). The fact that surface lowering was largest, reaching 50–100 m, in a ring segment 20–60km from the 1995 hinge line (see Fig. 3a), while the total volume lost in the ice shelf was much smaller (maximally 25 m thinning over a smaller area), suggests an initiation of changes in the interior of the glacier. The internal dynamic changes might be a sign of a surge-type acceleration as expected by Reference HughesHughes (1973), or a different type of dynamic process.

According to Reference Rabus and LangRabus and Lang (2003), an acceleration with an amplitude of 12% of the PIG surface velocity extends 80 km up-glacier of the grounding line (in the determination of Reference Schmeltz, Rignot and MacAyealSchmeltz and others, 2002), which coincides with the area of surface lowering detected by our analysis (Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others, 2008). In the difference map (Fig. 3a), surface lowering extends at least to 700 m elevation.

In recent years, increasingly large changes have occurred in the PIIS. Reference VaughanVaughan and others (2012) observed basal melting under the PIIS, and related fracturing in the top of the ice shelf. PIG was grounded on a transverse ridge, but warm sea water has recently started to flow through a widening gap under the ice shelf, rapidly melting the thick ice of the newly formed upstream half of the ice shelf. Retreat of PIG was found to be unstable (Reference JenkinsJenkins and others, 2010). Data collected as part of NASA’s Operation IceBridge revealed the existence of a trough below the PIIS from the ice-shelf edge to the grounding line, which enables warm Circumpolar Deep Water to penetrate to the grounding line and lead to higher melt rates (Schodlock and others, 2012); a sensitivity study indicated that the cavity shape matters in this context. Reference ThomasThomas and others (2011) assumed that rapid thinning of the PIIS increases the likelihood of the ice shelf’s break-up and calculated the accelerating ice loss that may be expected to occur in this case, concluding that velocities in PIG may reach 10 km a–¹.

Reference Payne, Vieli, Shepherd, Wingham and RignotPayne and others (2004, Reference Payne, Holland, Shepherd, Rutt, Jenkins and Joughin2007) attributed recent thinning in PIG to interaction of the PIIS with warm ocean water, and subsequent retreat of the PIIS. Reference Pritchard, Ligtenberg, Fricker, Vaughan, Van den Broeke and PadmanPritchard and others (2012) attributed ice-sheet loss in Antarctica generally to basal melting of the ice shelves. It is generally known that loss of an ice shelf will cause loss of buttressing and in consequence upward propagation of acceleration into glaciers that flow into this ice shelf, followed by mass transfer down-glacier. While this process is happening in the system of PIG and the PIIS, the question is whether the entire acceleration observed in PIG may be explained by mass loss from the ice shelf. O. Gagliardini (unpublished information) derived a new result from modeling, stating that changes in the PIG ice shelf cannot be the sole cause of the observed surface lowering and accelerations in the glacier.

Another different perspective on the possible sources of changes in PIG is given in studies of subglacial geophysics (Reference VaughanVaughan and others, 2006; Reference Scott, Gudmundsson, Smith, Bingham, Pritchard and VaughanScott and others, 2009). Reference Scott, Gudmundsson, Smith, Bingham, Pritchard and VaughanScott and others (2009) find that the increased rate of acceleration in PIG is strongly coupled to changes in gravitational driving stress. Using measurements of ice thickness, gravimetry, seismics and GPS-based surface elevation, Reference Smith, Bentley, Bingham and JordanSmith and others (2012) determine subglacial erosion beneath PIG at 0.6 ± 0.3 m a–¹ at a location in the main trunk and conclude that subglacial erosion in a soft bed plays a significant role in ice dynamics. Subglacial topography of PIG is described in Reference VaughanVaughan and others (2006).

As this short literature review shows, there has not only been a wide range of data collection and analysis of the changes in PIG, there is also a wide range of interpretations of the changes in PIG. In summary, most authors favor a single process as an explanation for the retreat of PIG. The third objective of this paper is to relate the elevation changes that will be derived from the GLAS-data analysis to the geophysical processes that may have caused them.

4.1. GLAS instrumentation and ICESat mission parameters

In this subsection we provide a brief overview of ICESat GLAS instrumentation and data characteristics, as much as is needed to motivate the analysis methods. GLAS, the sole instrument aboard ICESat, is a next-generation space lidar. GLAS combines a precision surface lidar with a sensitive dual-wavelength cloud and aerosol lidar, emitting infrared and visible (green) laser pulses at 1064 nm and 532 nm wavelengths. The 1064 nm (infrared) laser channel is designed to measure surface altimetry and heights of dense clouds, and the 532 nm (green) lidar channel to measure the vertical distribution of clouds and aerosols. More precisely, the GLAS laser produces a 1064nm 40 Hz pulse for altimetry and lidar, and a Doppler crystal produces a 532 nm wavelength pulse, which yields a more sensitive determination of the vertical distribution of clouds and aerosols (Reference ZwallyZwally and others, 2002; Reference Schutz, Zwally, Shuman, Hancock and DiMarzioSchutz and others, 2005).

ICESat orbits at 600 km above the Earth’s surface. The footprint of GLAS is ~65 m in diameter, and the ‘spots’ illuminated on the Earth’s surface are separated by 172-173 m in the along-track direction. The echo pulse is accepted by a 1 m diameter telescope, directed to an analog detector, digitized by a 1 GHz sampler, along with a digitized record of the transmitted pulse. The digitized pulses constitute the waveform data, from which transmit and echo-receive time are determined. From the resultant two-way travel time, the range between the satellite and the Earth surface is calculated, and, using precision orbit determination based on GPS star tracking and post-processed improvement of the pointing accuracy of the laser, the elevation of a point on the surface is derived (Reference Schutz, Zwally, Shuman, Hancock and DiMarzioSchutz and others, 2005).

4.2. Processing of GLAS data to derive PIG DEMs

The ICESat mission was designed as a change-detection mission, so each submission was operated as a repeat-track mission. However, the spatial distribution of observations along ground tracks varies due to several factors. Clouds, aerosols and other atmospheric conditions cause frequent losses of altimeter data, especially in Arctic and Antarctic regions. Off-pointing maneuvers result in different track locations. The resultant ground-track locations are only close to repeat, but not exactly repeating, which complicates derivation of differences along-track without knowledge of local slopes. Therefore formation of differences directly from track data is not generally possible. Because of difficulties in the spatial data distribution, we calculate DEMs for each (sub)mission first and then form difference DEMs, on which the analysis of elevation changes is based. Finally spatial changes in the change patterns are analyzed.

Prior to geostatistical estimation of DEMs, the GLAS data are analyzed as follows (for a more detailed description of GLAS data processing, as well as ERS-1 data analysis, see Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others, 2008). GLAS data are recorded in 33 day submissions (L2A, L2B, etc.), and data products of different types are created by the NASA ICESat Project. Submissions are named after the lasers that were operated. Here we use products GLA06 (global elevation product), GLA05 (waveform product) and GLA12 (Antarctic and Greenland ice sheet elevation product). To match problems in recording L2A data, we derived a corrected dataset from a combination of data in GLA06 and GLA05 (a data point in GLA06 was rejected if the received gain value of the matching record in GLA05 exceeded 50). The submissions L2A (October/November 2003), L3C (May 2005), L3E (February 2006), L3F (May 2006) and L3I (October 2007) used in the following analysis have been selected, because the spatial coverage was good and because the energy levels in the lasers resulted in data of good quality. For these missions from L3C onward, we used GLA12 data (release 428). The release 428 data include a gain criterion that matches the one applied for L2A data in earlier work (Reference Herzfeld, McBride, Zwally and DiMarzioHerzfeld and others, 2008) to produce GLA12 data, with submission-specific gain-threshold values matching the decrease in laser power. The ICESat Project continues to reprocess GLAS data, implementing new corrections. However, elevation differences between releases are now very small compared to changes in PIG (centimeters or subcentimeters versus 25–150m elevation changes), hence reprocessing of the maps presented here is not necessary.

4.3. Geostatistical analysis

To calculate DEMs from GLAS data, the geostatistical interpolation/extrapolation method of ordinary kriging (OK) is applied, in the form adapted for analysis of geophysical trackline data, and, in particular, satellite altimeter data, described in detail in Reference HerzfeldHerzfeld (2004) (see also Reference HerzfeldHerzfeld, 1992; Reference Herzfeld, Lingle and LeeHerzfeld and others, 1993a, 1994, Reference Herzfeld, McBride, Zwally and DiMarzio2008, Reference Herzfeld, Wallin and Stachura2011). Kriging methods are methods of geostatistical estimation, commonly formulated in a probabilistic framework. For the ground-track patterns of GLAS data (Fig. 4) geostatistical estimation with OK performs an interpolation. The OK method proceeds by two steps: (1) variography and (2) interpolation. In step 1, an analysis of the spatial structure of the data is carried out, experimental variogram functions are calculated and then variogram models are derived, which in turn provide input values for step 2, the interpolation. The variogram model type and the variogram model parameters determine the spatial continuity or roughness of the resultant map. The estimation is based on a least-squares minimization with respect to an unbiasedness condition. For OK the unbiasedness condition is that the mathematical expectation is constant (in the neighborhood of an estimated point), which translates into the condition that the kriging weights sum up to one. For the analysis of PIG GLAS data, variograms are derived for different subareas: (1) Pine Island Glacier, tongue and ice shelf, (2) lower PIG, (3) middle PIG and (4) PIG basin. Maps are created using three variogram models and a number of kriging parameter combinations to select a single model to be used for all DEMs. A Gaussian variogram model with a nugget effect of 350 m² , a sill parameter of 3450 m² and a range of 6000m was used. The same variogram model is employed for all GLAS submission datasets and throughout the region to ensure that differences in elevations are only due to differences in the data and not to differences in the variogram model used. Studying the influence of the variogram on kriging results is not the objective of this paper but has been investigated in Reference Herzfeld, Eriksson and HolmlundHerzfeld and others (1993b).

The error is a combination of (1) the error in the data, which is affected by atmospheric conditions including cloud cover, ice-surface roughness, instrument- and laser-pointing errors and (2) errors associated with the interpolation method. Instrument- and laser-pointing errors can be neglected for GLAS data, compared to elevation changes for PIG. Elevation maps, and hence elevation-change maps, derived from satellite laser altimeter data are prone to errors in areas of high topographic relief. The color scheme in the elevation-change maps is chosen such that red/purple areas at both ends of the color scheme flag likely artifacts caused by topographic relief. It is known that altimeters lack accuracy over mountain ranges and ice-shelf edges (Reference Partington, Cudlip, McIntyre and King-HelePartington and others, 1987), while the accuracy of GLAS data has been determined as 2.1 cm over the interior of the Antarctic ice sheet, under clear-sky conditions and in areas of low slope (Reference ShumanShuman and others, 2006). For our study, this means that elevation changes over PIG and its ice shelf can be interpreted as geophysically caused, while accuracy over the adjacent Hudson Mountains (northeast corner of the map, northeast of the ice stream) is poor. Apparent elevation changes over Pine Island Bay may be due to changes in ocean altimetry or floating icebergs.

There are two possibilities for calculating the interpolation error of kriging: (1) the estimation standard deviation and (2) the numerical error. The estimation standard deviation (Reference HerzfeldHerzfeld, 1992) is a function of the variogram and hence increases with distance from the ground tracks, especially if a nearest-neighbor search is performed. The numerical error associated with propagating the data errors through the kriging equations depends on the noise levels and does not reflect the ground track patterns, as demonstrated in Reference Herzfeld, Lingle and LeeHerzfeld and others (1993a). However, the accuracy of a DEM and hence the quality of an elevation map can be improved by implementation of an alternative search routine.

The kriging software applied in the calculation of the DEMs used in the analysis in this paper includes a search algorithm specifically implemented for analysis of geophysical trackline data (part of the ‘advanced kriging’ algorithms of the author’s software). For grid nodes close to a ground track, the kriged elevation is based on points in a small neighborhood. If too few data points are found close to a grid node (as is typical for gaps between tracks), then a quadrant search is performed to ensure that an interpolation is carried out at the grid node controlled by observations taken in all directions that surround the grid node, rather than an extrapolation from data at the nearest track. This search algorithm serves to counteract the track pattern effect as much as possible. Close scrutiny of the elevation and elevation-difference maps that are overlain by track maps (Fig. 2b, 3b and 4b) shows that elevation features do not follow track patterns (Fig. 2b) and that local maxima and minima in elevation difference are found directly under the tracks as well as between them (Fig. 3b) for elevation changes over longer terms of 8–11 years. Short-termelevation change maps are somewhat influenced by track patterns (Fig. 4b), depending on track density, as absolute elevation changes are smaller and as regular tracks are missing due to signal absorptions in clouds.

In summary, by first calculating DEMs from altimeter data, elevation-change values can be derived for every point of a grid. Without gridding, elevation changes could only be derived at crossover points between ERS-1 tracks and GLAS tracks, or elevation-change analysis be limited to the time frame of a single satellite mission. A detailed analysis of the combined effects of satellite-altimeter-track spacing and kriging on the accuracy of resultant DEMs is carried out in Reference Herzfeld, Wallin and StachuraHerzfeld and others (2011) but is beyond the scope of this paper. The primary objective of this paper is the analysis of elevation changes for PIG over a long time frame which spans two satellite missions (1995–2007); this is geophysically meaningful after interpolation because the elevation changes are absolutely very large.