Current global elevation data sets over the Greenland Ice Sheet are too low quality for robust glaciological applications. The most widely used DEM for glaciological studies was generated by Bamber et al. from mid 1990s satellite radar altimetery and aerial photogrammetry. A subsequent study (see Scambos and Haran, 2002) used AVHRR photoclinometry to enhance the Bamber DEM, hereafter referred to as the Photo-Enhanced Bamber (PEB) DEM; however, while this approach improved the effective spatial resolution and accuracy over the relatively flat interior of the ice sheet, errors over regions with higher relief such as the margins and periphery remained equivalent to the Bamber DEM.

The GIMP DEM achieves enhanced resolution and accuracy by merging the PEB DEM with high-quality photogrammetric topography for the margins and periphery and registering all land elevations to ICESat/GLAS laser altimetry data. Registration in this context refers to the process of correcting elevations in the merged DEM based on the average of multiple nearby ICESat/GLAS elevations acquired between 2003-2009. This approach ignores temporal changes in ice elevation, which can exceed 100 meters near the fronts of some rapidly retreating glaciers, to focus on generating a continuous surface that approximates the mean elevation over the ICESat operational window.

The following sections describe how the input data were prepared and merged into the final DEM. The following figure illustrates this process:

Data Preparation

Input data were registered to ICESat/GLAS elevations using the GLAS/ICESat L2 Antarctic and Greenland Ice Sheet Altimetry Data, Version 33 data set (GLA12 V33¹). The GLA12 data were corrected for detector saturation and time-varying elevation biases estimated from apparent variations of the mean sea surface height. The time-varying bias correction should also account for offsets associated with campaign-to-campaign variations in the shape of the transmitted pulse (see Borsa et al., 2014). Poor-quality returns were removed using techniques that identify the best-quality returns based on parameters which describe the shape and amplitude of the returned laser pulse (see Shepherd et al., 2012). Elevations calculated in this manner should be accurate to better than 0.1 m, or two orders of magnitude smaller than the expected uncertainty in the DEM.

¹GLA12 V33 was subsequently replaced with GLA12 V34. See ICESat/GLAS | Description of Data Releases for details.

Photo-Enhanced Bamber (PEB) DEM

The PEB DEM was regridded from a spherical Lambert azimuthal projection posted at approximately 627 meters to the WGS84/NSIDC Sea Ice Polar Stereographic North projection (EPSG:3413)¹. The regridded data were up-sampled to 30 meters using bilinear interpolation and then registered to ICESat/GLAS using an iterative, 3-D conformal transformation that yields residuals between the DEM surface and the ICESat point cloud with a normal distribution and a mean of 0 (see Noh and Howat, 2014). The registration was performed on 25 km x 25 km tiles with 5 km of overlap. The registered tiles were then mosaicked using linear distance-weighted edge feathering.

The registered tiles have a total root mean square (RMS) error of ±21.8 m, nearly three times the error reported by Bamber et al. in 2001 and Scambos and Haran in 2002. This likely reflects more extensive sampling by the laser altimeter system on board ICESat compared with the airborne altimetry used in the earlier studies, especially over ice-free terrain where errors are much higher. RMS errors over the ice sheet interior are more consistent with the original Bamber and PEB DEMs.

¹For details about WGS84/NSIDC Sea Ice Polar Stereographic North (EPSG:3413), go to EPSG Geodetic Parameter Registry, select the "Retrieve by Code" tab, and search for code 3413.

ASTER GDEM

The Global Digital Elevation Model (GDEM) is produced by NASA and the Ministry of Economy, Trade, and Industry (METI) of Japan. The GDEM is created by average-stacking individual stereo-photogrammetric DEMs acquired by the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) between 2000 and 2010. The GDEM is posted at 30 m and distributed in 1° x 1° tiles in geographic projection. The GDEM metadata specifies the number of individual granules that were stacked to obtain each posted elevation but not which scenes were used. As such, the time period of elevation measurements cannot be determined directly.

GDEM data quality is poor over much of the ice sheet owing to low-contrast between adjacent surfaces on snow and ice. Additionally, artifacts are abundant over all terrains due to shadows, clouds, and errors made by the automated matching algorithm. Following reprojection and gridding to EPSG:3413, the investigators applied a pyramiding standard deviation filter in which the DEM was progressively smoothed to finer resolutions and differenced from the native-resolution DEM. Pixels with differences exceeding 2.5σ of the mean were discarded. The filter was applied separately to ice-covered and ice-free terrain using land classification masks because ice-covered terrain is substantially smoother than ice-free. Following automated filtering, remaining artifacts were manually masked on a hillshade image of the DEM. These procedures removed nearly all data from above 1600 m elevation, which coincides approximately with the average mass balance equilibrium line altitude. The final filtered and masked GDEM covers 30% of Greenland's total area and 92% and 19% of its total ice-free and ice-covered terrain, respectively.

SPIRIT DEM

Photogrammetrically derived DEMs over Greenland were produced from images acquired in 2007 and 2008 as part of the Satellite Pour l'Observation de la Terre (SPOT 5) stereoscopic survey of Polar Ice: Reference Images and Topographies (SPIRIT) program. The SPIRIT DEM is distributed in UTM projection referenced to the EGM96 Geoid and posted at 40 m. Two versions are available, processed with different correlation parameters, along with data quality and interpolation masks. Based on validation experiments using ICESat, SPIRIT DEMs offer slightly better precision and accuracy (< ±5 m) than ASTER DEMs (see Korona et al., 2009).

As advised by Korona et al., all available Version 2 SPIRT DEM products covering Greenland were obtained and interpolated pixels were masked out. The DEMs were then reprojected to EPSG:3413, up-sampled to 30 m, and filtered and masked using the same procedure as the GDEM. Each individual SPIRIT DEM was then registered to overlapping regions of the filtered GDEM using the previously described 3-D conformal transformation (see Noh and Howat, 2014). This provided a consistent registration between the SPIRIT and GDEM data sets to facilitate merging. Each individual SPIRIT DEM was then stacked into a single mosaic by taking the median elevation at each pixel. The resulting filtered SPIRIT mosaic covers 10% of Greenland's total area and 24% and 8% of its total ice-free and ice-covered terrain, respectively. The most continuous coverage is along the southwestern and southern coasts: approximately 50% of the land and ice area is covered in each tile, which accounts for most of the land and ice area below 1500 m elevation.

Data Merging

Following registration and stacking, the SPIRIT DEM mosaic was differenced from the GDEM and the differences were extrapolated across the grid using an inverse-distance interpolation. The extrapolated difference map was then added to the SPIRIT DEM. The GDEM and corrected SPIRIT DEMs were then merged at each pixel using the following conditions:

1. If there was a GDEM value but no SPIRIT value, the pixel was assigned the GDEM value;
2. If there was a SPIRIT value but no GDEM value, the pixel was assigned the corrected SPIRIT value;
3. If there were both GDEM and SPIRIT values and the pixel was over ice-free terrain, the pixel was assigned GDEM value. This is due to the GDEM's higher spatial resolution;
4. If there were both GDEM and SPIRIT values and the pixel was over ice, the pixel was assigned a weighted average of the GDEM and SPIRIT values (see Howat et al. 2014).

The merged GDEM and SPIRIT DEM (merged G-S) was then registered to the ICESat/GLAS point cloud using the 3-D conformal transformation described in the preceding sections (Noh and Howat, 2014).

To combine the merged G-S DEM and PEB DEM, the PEB DEM was first adjusted by differencing it from the merged G-S DEM and interpolating the differences across areas of no data in the merged G-S DEM. The difference was then added to the PEB DEM and the two DEMs were combined using the following rules at each pixel:

1. If there was a merged G-S DEM value, the pixel was assigned the merged G-S value;
2. If there was no merged G-S DEM value, the pixel was assigned the adjusted PEB DEM value.

Lastly, to ensure correct sea surface heights, ocean surfaces in the final DEM were masked out using the GIMP Ice & Ocean Mask and replaced with heights from the CLS11 Mean Sea Surface product from the Centre National d'Etudes Spatiales. The CNES CLS11 consists of mean sea surface heights computed from measurements acquired by several satellite radar altimeters during 1993-2009. For this project, the CLS11 was reprojected from its native 1/3° grid to EPSG:3413 and up-sampled to the 30 m GIMP grid using bilinear interpolation.

The final GIMP DEM thus provides an altimeter-registered, relief-enhanced version of the PEB DEM. The improvement is most pronounced over regions of high relief on the margin and periphery of the ice sheet. Notably, whereas outlet glaciers are not clearly defined in the PEB DEM, the GIMP DEM resolves outlet glacier termini and fjord walls in detail.

Error Sources

The overall root-mean-square (RMS) difference between GIMP DEM and ICESat elevations is ±9.1 m, which is less than half the RMS of the ICESat-registered PEB. The error on ice-free terrain (±18.3 m) is more than twice that of ice-covered terrain (±8.5 m), which is to be expected considering the higher relief at the ice-free margin. Note that an unknown fraction of this error can be attributed to differences between the geometries of the GLAS footprint (a typical diameter of 70 m) and pixels in the DEM, and that the impact of this difference increases with slope. Furthermore, some of the error over ice-covered terrain stems from temporal variations in surface elevation, ranging from 10ths of a meter over the ice sheet interior to 10s of meters over rapidly thinning outlet glaciers. In addition to ice thinning, crevasses and other surficial features advected by ice flow contribute unknown error. These errors should therefore be viewed as an upper bound for the true standard data error.

The largest errors lie in the most northern regions, where few high-resolution data exist and coverage is mostly from the PEB DEM. Errors exceeding ±20 m can also be found in areas of extreme relief such as the Geikie Peninsula (tiles 4,2 and 5,2), where gaps in high-resolution data exist over steep mountain glaciers and ice caps. RMS errors are smallest for the flattest surfaces, for example in the interior of the ice sheet, and increase with slope to a peak of ±24 m at 2° slope. This peak roughly corresponds to the equilibrium line of the ice sheet, and thus the boundary between the merged G-S and PEB DEMs. Errors in both DEMs result in spurious, step-like transitions when crossing boundaries between them. This effect yields a continuous zone of large errors running along the southeast margin, which is especially steep.

For additional information about errors in the final merged DEM, see Howat et al., 2014.