This DEM offers the cryospheric research community a new elevation data set over the Greenland Ice Sheet of sufficient quality for robust glaciological applications. It improves upon the MEaSURES Greenland Ice Mapping Project (GIMP) Digital Elevation Model, Version 1 by utilizing sub-meter resolution, commerical imagery. Imagery for this project was provided by the Polar Geospatial Center at the University of Minnesota. The underlying imagery acquired by DigitalGlobe, Inc.'s constellation of satellite imaging systems is available to researchers through the National Geospatial-Intelligence Agency's NextView License program.

Data Preparation

The DEMs in this data set were generated from in-track stereoscopic imagery in which both images of the stereo pair were collected minutes apart along the same orbital pass. Swath imagery was collected in strips, ranging from 12 km to 17 km wide and up to 300 km long, and segmented into scenes that overlap by approximately 20 percent. Individual, overlapping scenes were then paired for stereo processing. Using Surface Extraction from TIN-based Searchspace Minimization (SETSM) software (Noh and Howat, 2015), the investigators produced DEMs on High Performance Computing (HPC) systems at: the Ohio Supercomputer Center; the San Diego Supercomputer Center (Gordon); and the National Center for Supercomputing Applications (Blue Waters). Raster DEMs with a resolution of 8 m were extracted and filtered based on the density of point matches, which removes erroneous surfaces due to clouds, water, server shadowing, and other sources. The filtered scenes were then mosaicked back into DEM strips using the iterative slope regression method of Nuth and Kaab (2011). This work was supported by multiple Ohio State University and National Science Foundation allocations.

Quality Control

For quality control, all DEM strips were scored on a scale of 1 (best) through 5, and strips that scored a 4 or 5 were discarded. Strips with a score of 3—typically mostly clear images with some cloud obscuation—were manually masked to remove erroneous surfaces. Quality controlled strips were then mosaicked using the iterative slope regression method for DEM coregistration and registration to LiDAR altimetry (Nuth and Kaab, 2011). This method is same as the original GIMP DEM.

Two versions of the DEM are available. The "reg" DEM is an unsmoothed, reference DEM that provides best estimates of elevation and uncertainty for each pixel when the data were acquired. Date stamps and 1σ errors are provided as separate GeoTIFFs. reg is available as both a quarterly time series and as a single, merged best data mosaic. Because the edges of registered strips are not feathered or smoothed, surface discontinuities will be present in areas of rapid change.

The "fit" DEM offers users a smoothed, best continuous surface for visual display and for applications that require slope information such as supraglacial stream flow modeling. fit was produced by selecting adjoining strips that yielded the best alignment and then feathering strip edges through distance-weighted averaging. Once complete, the entire mosaic was then registered to all available Operation IceBridge Airborne Topographic Mapper data. This approach minimizes surface discontinuities in areas of change at the expense of elevation confidence. fit is provided as a single mosaic derived from all available DEM data. 1σ errors are stored in separate GeoTIFFs.

Mosaicking

The following sections outline the mosaicking procedures for each version of the DEM.

reg

All DEM strips were first individually registered to Operation IceBridge Airborne Topographic Mapper (ATM) and Land, Vegetation and Ice Sensor (LVIS) LiDAR point cloud data. All years of data (beginning in 1993) were used for bare rock areas; ice-covered areas were limited to data collected within 30 days of image acquisition, as classified by the GIMP Land Classification Products.

Once registered, strips were indexed by the quarter of the year in which they were acquired (see Table 1). For each quarterly mosaic, all registered strips from that quarter were added sequentially in order of ascending quality score and registration error, with the best data remaining. Strips with 1σ registration errors > 4 m were excluded and treated as unregistered.

Registered strips were added in “underprint” mode—only empty pixels are filled, without data averaging, edge feathering or smoothing—so that each pixel provides the best estimate of elevation on that date. Once all registered strips were added, unregistered strips from the same quarter were added, in order of quality rank and lowest coregistration error, to fill gaps via co-registration with registered data. Following coregistration, the edges of unregistered strips were warped using a linear weighted adjustment inward 500 pixels. The error at pixels in the unregistered strip is calculated as the square root of the sum of the squares of the coregistration error and the mean error of registered strips used as reference for coregistration.

Once all strips overlapping existing, registered data were added via coregistration, any remaining gaps were filled with available unregistered data. The strip with the most new data coverage was added first as an anchor point. Strips which overlap this anchor were then coregistered and added using the same procedure described above. Errors for these unregistered clusters are assigned a value of “inf.”

The single, best data reference mosaic was created from the quarterly time series mosaics by selecting the pixels with the lowest registration error. Once the initial mosaic was generated, unregistered ice-free areas in the quarterly mosaics were added via coregistration.

fit

To create fit, the mosaic was initialized by adding to the mosaic grid the quality-controlled, level 1 strip with the most spatial coverage. Next, the overlapping strip with the best coregistration fit (lowest residual after fit) was added if it contributed 100 or more pixels of new data to the mosaic. A linear, inverse-distance feathering was then applied to the overlapping region between the new and mosaicked data. Strips were added sequentially in this manner until no additional overlapping strips remained which met the 100 pixel requirement. If additional, non-overlapping strips were available, another cluster of coregistered data was initialized by again adding the highest quality-ranked strip with the greatest coverage and then adding overlapping data as before.

Once the mosaic was built, each cluster of coregistered data was registered to data from the Operation IceBridge Airborne Topographic Mapper (ATM) and the Land, Vegetation and Ice Sensor (LVIS) LiDAR point cloud. The 1σ error of the fit to the LiDAR point cloud is stored as a supplemental GeoTIFF file. Values of inf indicate that no registration data were available.

Following registration, GIMP ocean masks were applied and the DEM was merged with the GIMP Digital Elevation Model, Version 1 to fill any remaining data gaps. These data were merged by interpolating the difference between the two DEMs across gap boundaries and adding the difference to the GIMP DEM Version 1 value. The pixel with the missing data was then filled with the adjusted value.

Error Sources

The primary error sources are discussed in the preceding sections, namely discontinuities or a loss of certainty in the regions of change. In addition, discontinuities are apparent where gaps in fit were filled with lower-resolution GIM DEM, Version 1 data (refer to Figure 2). Note that these artifacts can be removed by masking out pixels with error values equal to inf in corresponding error GeoTIFFs.