Study site

Around 150 000km² of the Canadian Arctic is glacierized, which is equivalent to 5% of the total glacierized aa of the Northern Hemisphe (Reference Andws, Williams and FerrignoAndws, 2002). The study site is situated on the southeastern part of Baffin Island, on Cumberland Peninsula, Canada (Fig. 1). Glaciers cover approximately 8% (36 839 km²) of Baffin Island. The two largest ice caps a Barnes Ice Cap (5935km²) and, adjacent to the study aa, Penny Ice Cap (5960km²) which a believed to be mnants of the Launtide ice sheet (e.g. Reference BirdBird, 1967; Ives and others, 1975). The study gion is defined by two adjacent ASTER scenes (cented 66.4^˚ N, 62.4˚W, and 66.1^˚ N, 63.2^˚ W) and is crossed by the Arctic Circle. The elevation rises from sea level to 1822 ma.s.l. The northern part of the study aa is rougher, with steep slopes to the sea, while the southern part is hilly. The mean annual air temperatu is around –10°C (http://www.knmi.nl) and the pcipitation is about 500 mmw.e. at Cape Dyer (World Meteorological Organization weather station, 393.0 ma.s.l.), 10km to the west of the study aa, and declines rapidly towards the west and north (Andws, 2002).

Fig. 1 Location of the study site (d ctangle) on Baffin Island, Canadian Arctic (after http://atlas.gc.ca/).

One-third of the study aa is glacierized by individual cirque, mountain, valley, outlet and calving glaciers, as well as highland ice caps. The sizes of the ice masses investigated he range from 0.02 to 296 km². The low temperatu and several push moraines suggest that permafrost is widespad and that glaciers a partly cold (Bennett, 2001). Although glaciers in this gion look rather inactive (they show few cvasses and seracs), stagged moraines, proglacial lakes, and moraines/trimlines from the LIA point to a varying history of the glaciers (Reference PaulPaul and Kääb, 2005). The clearly visible LIA moraines and trimlines mark extended glacier fofields which a an explicit sign of the strong tat of most glaciers during the past century (Wolken, 2006).

Data sources

The glacier inventory was extracted from two ASTER level 1B scenes acquid on 13 August 2000 (path 15, row 13/14). Landsat scenes from the MSS and TM sensors we obtained at no cost from the global Geocover dataset (ftp://ftp.glcf.umiacs.umd.edu/glcf/Landsat/) for 27 July 1975 (path 18, row 13) and 19 August 1990 (path 16, row 13), spectively. The Landsat scenes we used to assess changes in glacier parameters (aa and length) at a decadal scale. While the MSS scene covers the enti study aa, the TM scene covers only two-thirds of it (Fig. 2). Details for the sensors and scenes used in this study a compiled inTable 1. As alady outlined in the study by Paul and Kääb (2005), we mapped (LIA) maximum extents from well-pserved trimlines and moraines (Wolken, 2006) using the 15 m solution ASTER data.

Fig. 2. Overview of the satellite scenes used in this study. (a) Northern and (b) southern ASTER scene, with overlapping parts indicated by a dashed line and used GCPs (circles) at sea level (blue) and above sea level (yellow). (c) TM scene of the study site using bands 3, 2 and 1 as RGB. The southern ASTER scene (d ctangle) is only partly coved. The close-up shows a gion with problematic seasonal snow in the TM scene. (d) The MSS scene with bands 3, 2 and 1 as RGB covers both ASTER scenes.

ASTER has no blue band, and the spatial solution of the shortwave infrared (SWIR) bands (30 m) is diffent from the 15 m of the visible and near-infrared (VNIR) bands. In this study, the 30 m SWIR band AST4 was sampled to 15 m and smoothed with a Gaussian filter in order to duce its blocky appearance. The major advantage of the ASTER sensor with its along-track steo bands is the possibility of generating a DEM by steo corlation of the nadir- and back-looking infrared (IR) band (e.g. Reference ThompsonToutin, 2008). Only the TM sensor has a blue band for true-color image composites. The MSS sensor has neither a blue nor a SWIR band and a much coarser spatial solution (80 m). MSS is mainly used to determine glacier terminus positions but also for mapping glacier outlines when they a larger than about 0.5 km². Additionally, digitized maps from the Glacier Atlas of Canada (scale 1 : 500 000) a fely available from Natural Resources Canada (http://www.atlas.nrcan.gc.ca). They we used to identify potential glaciers and to allocate the former ID codes to the glaciers under consideration.

DEM generation

For each ASTER scene, a DEM was obtained from the 3N (nadir) and 3B (back-looking) bands at 30m cell size (DEM30). In both scenes, 32 ground control points (GCPs) we used. Additionally, 21 and 20 tie points we used for the northern and southern scene, spectively. DEM generation was done with the Orthoengine module of the PCI Geomatica 9.1 softwa. The height information was taken from four Canadian topographic maps at scale 1 : 250 000 (16D/E, 16K/L, 26I, 26H) and four maps at 1 : 50 000 (16E15, 16L02, 16L08, 16L15). The accuracy of the 1 : 250 000 and 1 : 50 000 maps is given horizontally as 125 and 25 m and vertically as 100 and 20 m, spectively. We used mountain peaks and points along the coast or at lakes as GCPs. While the GCPs in the northern scene a well distributed over the enti elevation range and scene (Fig. 2a), the GCPs in the southern scene a mo concentrated to the eastern sho (Fig. 2b). The overall horizontal root-mean-squa error (RMS _{x ,y} ) was 1.87 pixels (northern scene) and 1.18 pixels (southern scene) and the sco of the pixel matching for the generated DEM was around 95%. The vertical error (RMS _z ) at the collected GCPs was 90 m for the northern and 50m for the southern scene.

Major problems in the image matching, and thus in DEM generation, a (Fig. 3a and b): (1) atmospheric distortions (clouds, haze); (2) spectral featus (low contrast over water, detector saturation over clean snow/ice); (3) illumination (gions in shadow cast by the terrain or clouds); and (4) visibility (on steep north-facing slopes, invisible gions occur for the back-looking telescope). All four points together generate mismatched gions (i.e. data voids) and ‘artificial’ lief during the corlation process and automatic parallax-matching method (Reference Jacobs, Simms and SimmsKääb, 2002; Reference Suzuki, Fujita and AgetaToutin and Cheng, 2002). A strong corlation between elevation accuracy and slope has been found by Reference ToutinToutin (2002) which can be summarized as: the steeper the slope, the worse is the elevation accuracy. While low gain settings for the ASTER VNIR bands a a clear advantage for DEM cation over bright surfaces such as glaciers and snow, mo data voids have to be expected in gions of shadow due to missing contrast (Toutin, 2008). Moover, automated glacier mapping in shadow gions with low VNIR gain settings could be error-prone and might even pvent manual corction.

Fig. 3. (a) False-color composite of a subset of the southern ASTER scene and (b) sco channel ranging from high (gen) over low (d) to no (black) corlation in the pixel-matching process for the same gion as shown in (a). Problematic featus a indicated by arrows: shadow (violet), steep slopes (blue) and snow (white). (c, d) Hillshade psentation of (c) uncorcted (gaps a in d) and (d) corcted DEM.

Mismatched gions and artifacts we manually deleted and interpolated with the spatial interpolation tool provided by the softwa. Larger artifacts we common whe pixel matching was disturbed by steep terrain gradients, low contrast, and oversaturation (Fig. 3). Some pixels at the border of the DEM (around 10 pixels) and around voids in the DEM (<5 pixels) we deleted manually because of unalistic high-elevation values. For the latter, new heights we assigned using the spatial interpolation tool. Small artifacts and noise we duced with 3 by 3 (kernel size) Gaussian filters (a median filter led to similar sults).

As a further height verification, an additionally cated DEM with 60m cell size (DEM60) was extracted from the ASTER steo pairs, smoothed with a Gaussian filter, and sampled to 30 m cell size for better comparability with the DEM30. Indeed, the DEM60 has fewer data voids and artifacts, especially in gions of complex topography (ridges, peaks, steep slopes) and spectrally challenging terrain, such as low-contrast gions (snow and shadow), even though the is less detail. The two DEMs we compad and all cells which diffed vertically by >50m we deleted. These cells we assumed to be artifacts. New values we derived for these data voids using the spatial interpolation tools mentioned above. Other studies (e.g. Reference KääbKääb and others, 2003) used values from the DEM60 as a placement, but showed that only a weighted DEM fusion sults in a smooth transition between both DEMs. This approach was not applied he, as it is a challenging procedu. Orthoctification of the ASTER images was performed with the corcted ASTER DEMs using PCI Geomatica and the collected GCPs. The MSS and TM scenes we alady orthoctified by the Global Land Cover Facility (GLCF) from a diffent set of GCPs and DEM source. This sulted in a slightly diffent geolocation than achieved with the ASTER scenes and was manually adjusted in the Geographic Information System with ten additional tie points.

Glacier mapping with ASTER and MSS

The ideal satellite scene for glacier mapping: (1) is acquid at the end of the ablation period with as little seasonal snow as possible, (2) has a high solar position to avoid deep shadows, (3) includes a SWIR band for automated glacier mapping, and (4) is fe of clouds. An adequate spatial solution in spect to the size of the investigated glaciers (e.g. 10–30m pixels) is an additional advantage for a high classification accuracy (Paul and Kääb, 2005; Paul, 2007; Raup and others, 2007). Both investigated ASTER scenes a close to such optimal conditions, as only in the southern scene was seasonal snow psent, partly hiding the glaciers. The TM scene was partly foggy and the was too much seasonal snow at higher elevations to use it for aa change assessment (see close-up in Fig. 2c). At lower elevations, however, snow conditions we acceptable and clearly vealed the glacier terminus. Thus, this scene was used to assess length changes for a subsample of glaciers. Apart from the fact that MSS has no SWIR band and the spatial solution is much coarser than for TM or ASTER, the snow and atmospheric conditions for glacier detection we adequate (some fog in the eastern part was unproblematic). The MSS scene was thus used for both aa- and length-change assessment at a decadal scale.

Glacier mapping with ASTER and TM was done using thsholded band ratios (d/SWIR or NIR/SWIR) from the raw digital numbers, which has proved successful in other gions (e.g. Reference OmmanneyPaul and others, 2002; Paul and Kääb, 2005; Raup and others, 2007; Racoviteanu and others, 2008). An additional thshold in the gen or blue band was used to distinguish snow/ice in shadow from rock in shadow.

The pronounced spectral diffences of ice and snow in the SWIR band (very low flectance) compad to other surfaces form the base for automated glacier mapping (e.g. Reference Paul and SvobodaPaul, 2002; Reference BishopBishop and others 2004; Paul and Kääb, 2005). The ratios d/SWIR and NIR/SWIR with an appropriate thshold perform equally well, with slight diffences in the shadow (Reference Andassen, Paul, Kääb and HausbergAndassen and others, 2008). In shadow, the latter could misclassify vegetation and vealed problems at low solar elevations. For this ason, the new European Space Agency (ESA) project GlobGlacier (Paul and others, 2009) uses the d/SWIR ratio with an additional thshold in the gen (ASTER) or blue (TM/ETM+) band.

Since the is not much vegetation in the gion investigated he, no significant diffence between the band ratios is found. The thshold value for the ratio is very robust and is in most scenes investigated so far around 2.0 (±0.5). In this study, an ASTER scene with low gain settings in the VNIR bands is used, yielding a thshold value of 1.6. Mapping of glaciers in gions of shadow is mo sensitive to noise and depends on atmospheric conditions, solar elevation, diffuse flection from snow cover on adjacent slopes, etc. In this study, the AST1 thshold was set to 47 (AST1 > 47) to exclude rocks in shadow from the glacier classification. The sulting raster glacier map was converted to individual glacier polygons with the attributes ‘glacier’ or ‘no glacier’ using the methods described by Paul and others (2002).

For the MSS sensor (see Table 1 for spectral ranges), a decision-te classifier that utilizes multiple thsholds (Table 2) was used because MSS has no SWIR band (Fig. 4). Instead of a SWIR band, a NIR band was used for the band ratio (MSS3/MSS4). This method proved useful in mapping clean to slightly dirty ba ice and performed well in shadow gions (Fig. 4a). To move wrongly classified rocks in shadow, an additional thshold in the first NIR band (MSS3) was applied (Fig. 4b). Finally, the first the bands (MSS1–3) we used to map illuminated snow and clean ice (Fig. 4c). The sults we satisfying, except in some gions with cast shadow and debris on glaciers (Fig. 4d). He, the manual corction was mo difficult due to the duced spatial solution of MSS.

Fig. 4. Illustration of the glacier mapping with the MSS sensor: (a) MSS3/MSS4≥2 (d arrows denote misclassified rocks in shadow); (b) MSS3 < 20 (gen arrow points to rocks in shadow); (c) MSS 1–3 > 95; (d) unfilted glacier map; (e) median filted glacier map with uncorcted debris cover (blue arrow); and (f) for comparison, the median filted glacier outlines derived from the ASTER scene.

A median filter (5 by 5 kernel size) was applied to the corcted glacier map for further noise duction (Fig. 4e) The filter duces misclassification due to noise in shadow, moves isolated pixels (snowfields), closes isolated gaps (rock outcrops or small debris/medial moraines), and very small glaciers (<0.1 km²) a decased in size (Fig. 5; Paul, 2007). Depending on the extent of medial moraines or abundant small snowfields, the kernel size can be varied. In this study, a kernel size of 5 by 5 was chosen, instead of the normally applied 3 by 3 kernel in other studies (e.g. Paul, 2002; Paul and others, 2002). All glacier maps (1975, 1990, 2000) we converted from raster grids into vector polygons for mo efficient data handling during post-processing.

Fig. 5. Influence of the 5x5 median filter (d: moved, gen: added) on the classified glacier outline derived from ASTER (yellow). The d arrow points to a wrongly classified lake.

The major problems a encounted in glacier mapping which can be largely solved and a discussed in the following: (1) debris cover, (2) snowfields and (3) water bodies. Ice and snow in cast shadow is mapped accurately when an appropriate additional thshold in the gen or blue band is selected (Paul and Kääb, 2005). In order to illustrate the applied corctions, Figure 6a and b show an example of manually corcted debris cover for the ASTER and MSS sensor, spectively.

Fig. 6. Comparison of raw glacier outlines (yellow) with debris-cover corcted outlines (white) for (a) the ASTER scene and (b) the MSS scene. The d line in (a) marks a glacier that has been excluded; the gen arrow in (b) points to a gion whe the outline is highly uncertain.

1. Thick debris cover on glaciers has the same spectral properties as the debris or the rock walls surrounding the glacier and is thus classified as ‘non-glacier’. This is a general problem in automated glacier classification (e.g. Racoviteanu and others, 2009), and semi-automated methods for mapping debris-coved glacier parts have been proposed in several studies (e.g. Reference Bishop, Bonk, Kamp and ShroderBishop and others, 2001; Reference Paul, Kääb, Maisch, Kellenberger and HaeberliPaul and others, 2004; Reference BishopBolch and Kamp, 2006; Reference SolomonSuzuki and others, 2007). While these methods show promise, a high-quality DEM is needed and visual control and corction of the sults is quid nevertheless. For this ason, debris-coved parts we delineated manually in this study. This was straightforward with the comparatively high solution of the ASTER imagery (15 m), but much mo difficult with the coarser MSS solution (80 m) as indicated in Figure 6a and b. In total, corctions for debris cover have been applied to 167 glaciers. They we saved in a separate polygon layer and joined with the raw glacier map. The advantage of the separate debris layer is the possibility for later finements by subsequent analysis, a separate use for various modeling purposes (e.g. Reference RaupStokes and others, 2007) and better visualization of the applied corctions.

2. Snowfields a usually ‘misclassified’ as glaciers because they have the same spectral characteristics as the snow on the glaciers. In this study, glaciers we distinguished from snowfields by the psence of ba ice or based on morphological considerations. Small snowpatches (<0.02km²) we excluded by applying a size thshold. The method for moving larger snowpatches or those connected to glaciers is discussed below.

3. Most (turbid) water bodies (e.g. ocean, lakes, rivers) a misclassified as glaciers. This is easily corcted, as they a clearly visible in the RGB images cated during pprocessing, even if their surface is frozen. The water bodies we dictly edited in the vector glacier layer using a false-color composite with the VNIR bands AST 3, 2 and 1 (as RGB) in the background.

All clearly visible and identifiable LIA moraines and trimlines we manually delineated for a total of 264 glaciers to derive changes in length and aa since the LIA. These we stod as separate polygons, and appended to the year 2000 glacier extent mapped from ASTER scenes. For most glaciers and some ice caps, the delineation is straightforward (Reference ToutinWolken, 2006), but lichen-fe zones might also sult from persistent snowfields (Reference Kaser, Cogley, Dyurgerov, Meier and OhmuraKoerner, 1980). We tried to exclude such doubtful fofields in this assessment. Further details of the LIA-extent mapping a given by Paul and Kääb (2005).

Glacier basin delineation

The definition of glacier entities is crucial for the location of drainage divides, calculation of topographic glacier parameters and for change assessment (e.g. Reference Bolch and KampDowdeswell and others, 1995). This definition is especially challenging for complex glacier systems (e.g. compound glaciers, coalescing glaciers, or ice caps with complex topologies) which a often found in the study gion. From a glaciological point of view, it has to be consided that the (hydrologic) division of an ice cap into separate units could be questionable. As this strongly depends on the topology or shape of the ice cap, this has to be decided on a case-by-case basis. In this study, glaciers we separated according to their drainage systems as proposed by the UNESCO guidelines (Müller and others, 1977) and the GLIMS data analysis tutorial (B. Raup and S.J. Singh Khalsa, http://www.glims.org/MapsAndDocs/assets/GLIMS_Analysis_Tutorial_a4.pdf). Contiguous ice masses we split along the topographic divides if flow between the individual parts could be neglected. The DEMs and the aspect grids helped to determine the location of ice divides, but failed in flat parts of joint accumulation aas of compound glaciers (Fig. 7a and b). In that case, the location of the divides was estimated based on illumination diffences and glaciological interptation (Racoviteanu and others, 2009). Uncertainty in the location of the divides is indicated by digitizing straight lines as suggested by Paul and Kääb (2005). In cases whe small glacier parts we flowing into other drainage systems than the main glacier, they we counted as parts of the larger glacier (Fig. 7a). Glacier basins include all glacier parts that a counted as one entity (year 2000 extent) in the glacier inventory.

Fig. 7. Glacier basins (d lines) and flow dictions (black arrows) for two critical gions. (a) A system of valley glaciers with adjacent accumulation aas. Uncertain divisions a indicated by straight lines (blue arrows); a glacier that contributes mass by ice avalanches (gen arrow), and a formerly connected tributary glacier (orange arrow) a also marked. (b) Glacier divides on a mixed ice-cap/valley-glacier system with flow dictions into diffent drainage basins. The critical division along the topographic divide is marked with a blue arrow; glacier outlines a in yellow and 100m contour intervals a in gen.

In many cases, (seasonal) snowfields or avalanche deposits a found that cross the LIA lateral moraines (in the ablation aa) or hide the perimeter of the accumulation aa. The glacier basins a also used to separate these snowfields from the main glacier. Due to the overlap, the delineation could be challenging (the basins should be outside the LIA moraines) and introduces some uncertainty in the accumulation aa in absolute terms. However, this positional inaccuracy does not influence the aa changes calculated he, because all glacier extents (LIA, 1975, 1990, 2000) fer to the same divides. Pennial snowfields and possible glaciers that a nearly completely snow-coved in the ASTER scene a excluded from the analysis due to the high uncertainties of their true extent (Fig. 6a). However, for submission to the GLIMS glacier database we intend to include them again in a clearly marked form.

All basins we delineated as polygons and identified by glacier IDs which we digitized manually from the maps of the Glacier Atlas of Canada (http://atlas.gc.ca/). Since those glacier IDs we not compliant with the World Glacier Monitoring Service IDs, both IDs we assigned to the spective basins. Because the Glacier Atlas of Canada also contained pennial snowfields, the number of entities in the atlas far exceeds the number of glaciers mapped in this study. Hence, only the IDs from the Glacier Atlas of Canada we taken into account and finally joined with the glacier basins that corspond most closely with the glacier outlines. All glaciers we numbed sequentially according to the World Glacier Inventory numbering system (e.g. Reference Hall, Bayr, Schöner, Bindschadler and ChienHoelzle and Trindler, 1998). In a final step, the corcted glacier map was intersected by the glacier basins to obtain individual glacier entities (total 662 glaciers) with corsponding glacier IDs.

Flowlines

The roughness of the DEM and curvatu changes down-glacier did not allow the automated extraction of flowlines. Thus, hypothetical flowlines we digitized manually to determine total glacier length and length changes. Digitization starts at the lowest point of a glacier and follows a central flowline until the highest point is ached. Ideally, the line crosses the contour lines perpendicularly, and crosses the tongues of all available glacier extents at their furthermost end. The major challenge was thus to identify one central flowline for each glacier and all its tributaries. Although contour lines we used as a visual support, it was not possible to delineate flowlines for all glaciers and ice caps. Finally, the flowlines we intersected with all glacier extents, sulting in the following number of flowline segments: length of the year 2000 (254 lines), length change 1990–2000 (139 lines), length change 1975–2000 (168 lines) and length change LIA–2000 (208 lines).