1
1
2
3
Use of an Unmanned Aerial Vehicle to assess4
recent surface elevation change of Storbreen in5
Norway6
7
W.W. IMMERZEEL
1,*
, P.D.A. KRAAIJENBRINK
1
, L.M. ANDREASSSEN
2
8
9
10
1
Utrecht University, Department of Physical Geography, PO Box 80115, Utrecht, The11
Netherlands, w.w.immerzeel@uu.nl12
2
Norwegian Water Resources and Energy Directorate, PO Box 5091, Majorstuen, NO-13
0301 Oslo, Norway14
15
*Corresponding author16
17
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
2
Abstract18
Routinely and accurate monitoring of the outlines and surface mass balance of glaciers is19
essential. In this study an unmanned aerial vehicle (UAV) was used in September 2015 on a20
mountain glacier (Storbreen) in Norway to map the glacier outline, snow line and to derive a21
digital elevation model (DEM) of the glacier surface. The generated DEM has a relatively22
high accuracy with maximum horizontal RMSE of 0.36 m vertical RMSE of 0.44 m and the23
Structure for Motion algorithm also proved to be suitable under low contrast, high saturation24
fully snow covered conditions. A well distributed set of markers, measured by GPS, was25
required to generate a high quality DEM under the yielding conditions. The final UAV DEM26
was compared to a laser based DEM of 2009 and the annual geodetic mass balance between27
2015 and 2009 was estimated to be between -0.71 ± 0.1 m w.e. and -0.75 m ± 0.1 w.e., which28
is in good agreement with the glaciological mass balance of -0.80 m ± 0.18 w.e. a-1
. An29
analysis of the glacier outlines reveal that the glacier has lost 1.2% of its surface area between30
2009 and 2015. These findings confirm the strong mass loss and retreat of continental glaciers31
in southern Norway.32
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
3
1. Introduction33
The glaciers in Norway are, as elsewhere in the world, characterized by a reduction in area34
and a general loss of mass in particular since 2000 (Andreassen et al., 2016; Winsvold et al.,35
2014).36
Climate change is likely to play a major role and understanding the underlying37
mechanisms are of key importance for both water resources assessments and projections for38
future sea level rise. There is therefore a great need for systematic and accurate observations39
with a high spatial detail of glacier mass balances to further advance our understanding of the40
climate – glacier system.41
The surface mass balance of a glacier can be assessed using in-situ point observations42
(glaciological method) or by repeated surveys using remote sensing (geodetic method). Both43
methods are independent of each other, yet the differences between them are often substantial44
(Cogley, 2009) and a homogenisation procedure is required to make a reliable comparison45
(Zemp et al., 2013). The glaciological method uses point observations of ablation and46
accumulation (stakes, probings and snow pits), which are spatially interpolated to derive an47
overall glacier surface mass balance. The geodetic method is based on multi-temporal surveys48
of surface elevations and can be derived from several sources such as laserscanning (LIDAR),49
aerial photos or satellite imagery. Surveys can be terrestrial, airborne or from space. The50
glacier mass balance is quantified by differencing digital elevation models (DEMs) from51
different years and by converting volume change to mass change using a density conversion.52
Both the glaciological and geodetic approaches are imperfect as the glaciological method53
suffers from measurement errors and it is often complicated to measure a sufficiently large54
number of points on a glacier to derive an accurate interpolated spatial surface mass balance.55
The geodetic approach often relies on relatively coarse resolution DEMs and significant56
uncertainties stem from viewing angles, co-registration, density assumption, the glacier areas57
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
4
and masks and inaccurate DEMs in steep terrain (Kääb et al., 2015; Magnússon et al., 2015;58
Nuth and Kääb, 2011; Pellicciotti et al., 2015; Rolstad et al., 2009). A comparative study for59
10 glaciers in Norway with long-term series of glaciological and geodetic mass balance60
revealed that the discrepancy between the methods was larger than the estimated uncertainty61
for 7 out of 21 periods studied (Andreassen et al., 2016). The study stresses the need for62
independent geodetic survey as a way to validate field observations.63
Unmanned Aerial Vehicles (UAVs) are not yet commonly used in glaciology, but have a64
large potential for deriving accurate high resolution DEMs, quantifying surface velocity,65
thermal mapping and classification of glacier surface features among others (Bhardwaj et al.,66
2016; Immerzeel et al., 2014; Kraaijenbrink et al., 2016a, 2016b; Ryan et al., 2015; Vincent et67
al., 2016; Westoby et al., 2016). In the Nepalese Himalayas, for example, a fixed wing UAV68
was used to derive a pre- and post-monsoon DEM with the aim to investigate the surface69
lowering and dynamics of a debris covered glacier (Immerzeel et al., 2014). The study70
showed that it is feasible to derive accurate DEMs at 20 cm resolution for an area of several71
km2
with a high accuracy (~25 cm both vertically and horizontally) even in complex terrain.72
In subsequent studies UAVs were used to automatically derive seasonal flow velocities73
(Kraaijenbrink et al., 2016a) and to perform an object based classification of supra glacial74
lakes and ice cliffs (Kraaijenbrink et al., 2016c). The high resolution and accuracy in75
combination with the on-demand employability potentially provides the opportunity for76
frequent studies of smaller glaciers.77
The software used for generating the UAV based DEMs is based on the structure for78
motion (SfM) algorithm (Westoby et al., 2012) and it largely relies on automatic matching of79
features between overlapping images. Therefore, it is likely that DEM generation for snow80
surfaces and debris free glaciers with limited contrast is challenging and this has not yet been81
systematically tested. However, recent SfM studies under different conditions (prairies,82
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
5
exposed mountain tops, steep slopes) using different platforms (manned aircraft, copters,83
fixed wing UAVs) have shown that the accuracy in mapping snow depth in alpine terrain is in84
the order of 10 cm (Harder et al., 2016; Jagt et al., 2015; De Michele et al., 2016; Nolan et al.,85
2015). This offers the opportunity to routinely monitory snow and ice surfaces.86
In this study, a fixed wing UAV was used in combination with differential Global87
Navigation Satellite System (dGNNS) measurements on the glacier Storbreen in southern88
Norway to test whether it is feasible to use the SfM alghoritm to derive an accurate DEM for89
the snow cover accumulation area and the debris-free ice of the glacier. We compare the UAV90
results with a LIDAR based DEM from 2009 to quantify spatial changes in the surface mass91
balance and compare this with the glaciological mass balance data. Finally, we discuss the92
potential of UAVs in the routinely monitoring of the mass balance of glaciers.93
2. Study area94
This study focuses on the mountain glacier Storbreen in the Jotunheimen region in95
southern Norway (Figure 1). Storbreen (61°36’ N, 8°8’ E) has the longest record of observed96
mass balance in Norway, measurements started in 1949. Storbreen has been surveyed97
repeatedly in the past (Andreassen, 1999; Liestøl, 1967) and the latest survey is from 200998
(Andreassen et al., 2016). According to the 2009 survey, Storbreen has an altitude range from99
1400 to 2102 m a.s.l. and an average slope of 14° (Andreassen et al., 2016). The glacier has100
northeastern exposure, Storbreen can be characterized as a short valley or a composite cirque101
glacier (Liestøl, 1967). The area of the Storbreen has steadily decreased from 7.2 km2
at the102
end of the Little Ice Age (~1750) to 6.0 km2
in 1940 to 5.4 km2
in 1997 to 5.1 km2
in 2009103
(Andreassen, 1999; Andreassen et al., 2016). Reanalysis of the glaciological and geodetic104
mass balance series of Storbreen for three periods (1968-1984, 1984-1997 and 1997-2009)105
showed no statistical difference between the geodetic and glaciological mass balance.106
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
6
However, the differences were substantial for the 1984-1997 period (Andreassen et al., 2016).107
Their results show that mass balances for Storbreen have been predominantly negative since108
1968, except for a transient mass surplus over 1989-1995 (Andreassen et al., 2005; 2016),109
3. Methods and data110
3.1. Field surveys111
The latest full mapping of the surface elevation of Storbreen was made in 2009 and is used in112
this study together with the new UAV survey of 2015. In the following, we describe the 2009113
and 2015 campaigns.114
LIDAR survey 2009115
The data from 2009 consist of aerial photos taken on 14 September and airborne LIDAR116
elevation data of the glacier acquired on 17 October by Blom Geomatics. The time lag for the117
photography and the laser scanning is due to technical problems with the laser scanner on118
September 14. Aerial photos were also taken 17 October, but then the glacier was completely119
snow covered. The surveys were used to produce orthophotos with a 0.4 m resolution and a120
point elevation cloud with an average point density of about 1.7 points m-2
.The glacier outline121
was digitised manually from the orthophotos from 17 September.122
UAV survey 2015123
Storbreen was surveyed by an eBee (www.sensefly.com) UAV in 2015 on 9 and 10124
September 2015 in fair weather conditions. The UAV was used in six separate flights to cover125
most of the glacier surface (Fig. 2, Table 2). On September 9 it was launched from the upper126
part of the glacier and landed nearby on the snow, while on September 10 this was performed127
at the more easily accessible lower terminus. The UAV was set to take photographs with128
about 70% overlap. For each flight, the UAV followed waypoints of a predefined flight plan129
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
7
using its built-in GPS. Each waypoint’s altitude was determined in the accompanying130
software by combining a desired ground resolution setting with base elevation data. This131
helps to achieve a relatively constant UAV altitude with respect to the glacier relief and thus a132
relatively constant ground resolution. Desired resolutions were set to 6 to 10 cm per pixel,133
depending on the flight. The camera mounted in the UAV was a 16 megapixel Canon IXUS134
127 HS compact camera with customized firmware. The lens was set to its widest setting to135
reduce the number of required photos and flight time. In total the UAV acquired 915 separate136
images and covered an area of ~7.5 km2
.137
To put the data in a real world coordinate system 31 markers were distributed over138
accessible parts of Storbreen (Figure 2) before the UAV survey. The markers were made from139
1.0 by 0.8 m pieces of garbage bags in black (when positioned on ice or snow on the glacier)140
or blue (when positioned on debris or rock outcrops on the glacier or rock outside the glacier).141
To be used as ground control points (GCPs), each marker’s approximate centre was measured142
using a Global Navigation Satellite System (GNSS) rover from Topcon GR3 with an143
estimated accuracy of 0.1 m in x,y and z.144
There were no dedicated independent markers laid out for the determination of the145
geodetic accuracy of processed UAV data, as the number of markers were thought to be146
sufficient to leave a few redundant for ground control. For an additional independent147
indication of the output accuracy, surface profiles of the glacier (Figure 2) were measured148
using GNSS. This was performed by strapping the GNSS receiver to a backpack and149
recording its coordinates kinematically at a 1 second interval while traversing the glacier.150
Antenna height was measured on multiple occasions while standing still, but variation of the151
antenna height caused by movement of the person was not measured. We estimate the152
induced error is in the order of 0.3 m. The position of the profile was visually tracked and153
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
8
sampled from the DEM by identifying the xy locations of the footsteps in the snow on the154
orthomosaic.155
Glaciological and meteorological surveys156
During the field campaign in September 2015 stake length was measured for the 9 stake157
locations present on Storbreen and their position was recorded using GNSS. Snow covered158
most of the glacier and only the lower terminus and some exposed parts were snow free. The159
depth of the remaining snow was measured at 31 locations on the glacier using a snow probe160
and the density of the remaining winter snow was measured at three locations (Figure 2). The161
snow depth data was interpolated to a gridded continuous surface using ordinary kriging.162
3.2. UAV data processing163
The 915 images from the UAV survey were processed into a 3D model using Structure164
from Motion (Szelisky, 2011) in the software package Agisoft Photoscan Professional version165
1.2.4 (Agisoft LLC, 2014). Several previous studies have used this approach successfully to166
derive high quality digital elevation models (DEM) (Immerzeel et al., 2014; Kraaijenbrink et167
al., 2016a; Lucieer et al., 2013; Westoby et al., 2012) . The precise workflow and settings168
used to process the data are equal to those presented by Kraaijenbrink et al. (2016a, 2016b).169
To achieve optimal geodetic accuracy of the output product all but one of the markers were170
used as ground control during processing (Figure 2). One marker was discarded because of an171
erroneous GNSS measurement, most likely due to poor satellite coverage.172
The generated 3D point cloud was post processed in CloudCompare (Lague et al., 2013).173
The glacier surface was smooth at the time of the survey because of the remaining snow174
covering most of the glacier. A moderate local outlier filter was therefore applied to remove175
some erroneous irregularities in the cloud unrelated to the actual relief. The filtered cloud was176
gridded to a 0.5 m resolution DEM for accuracy assessment. For comparison with the LIDAR177
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
9
dataset from 2009 the point cloud was subsampled to a minimum point distance of 0.5 m.178
Additionally, the 3D information from the original point cloud was used in Agisoft to179
generate an orthomosaic, i.e. a stitched raster of orthorectified input images, with a resolution180
of 0.1 m.181
3.3. UAV product accuracy182
The geodetic accuracy of the SfM DEM was assessed using two methods: (1) cross-check183
with the GCP coordinates and (2) comparison with the independent GNSS profile. For the184
first method, the horizontal errors the difference between the measured marker coordinates185
and their digitized centres on the orthomosaic were measured. Vertical differences were186
subsequently quantified by comparing the output DEM elevation with the measured elevation187
after correcting for the horizontal error. As the GCPs are also used in the SfM DEM188
generation processes, the differences are a measure of the accuracy of the SfM optimization,189
whereas the comparison with the GNSS profiles is indicative for the spatial accuracy.190
For comparison with the GNSS profile 150 samples were randomly selected from the total191
of 23,164 points. Candidates were only points that lie further than 100 m from the nearest192
ground control point and that were recorded while walking steadily. At the points the193
horizontal error was estimated by measuring the shortest distance perpendicular to the centre194
of the footprint trail as observed on the orthomosaic, given that the trail was visible at the195
point under consideration. The difference between the recorded elevation and the DEM196
elevation at the point was used as indication of vertical error at all points. Horizontal error197
was here not corrected for, as horizontal errors could not be estimated for all points and as the198
differences between corrected and uncorrected vertical errors were generally negligible on the199
smooth glacier surface.200
In addition to the geodetic accuracies, local noise in the point cloud was estimated as such201
noise may affect the accuracy of gridded statistics, e.g. the cloud-derived DEM. The relatively202
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
10
smooth, snow-laden glacier surface was assumed to have little local variation, and high local203
variation in the point cloud was therefore utilized as noise indicator. To estimate it, linear204
models were fitted to the subsampled cloud on a 4 by 4 m grid for the entire glacier. The root205
mean square error (RMSE) of the residuals was then calculated to provide the noise estimate.206
The off-glacier areas are not smooth and consequently the method was not applied there.207
The importance of the number of markers and their distribution on the quality of the208
generated DEM was evaluated by reprocessing the UAV imagery using 5, 10 and 20209
randomly selected markers from the total set of 30 markers. The generated DEMs were210
compared to the original DEM and the differences were analysed. The analysis was done in211
those parts of the glacier that are characterized by an RMSE < 0.20 m and which are more212
than 100 meters away from the glacier margin to avoid undesired effects due to steep slopes.213
3.4. Glacier delineation and terminus retreat214
A vector outline of the glacier surface was constructed from the orthomosaic using object-215
based image analysis (Blaschke, 2010). The three-band orthomosaic was resampled to 1 m216
resolution and used as input for multiresolution segmentation in eCognition Developer 9.1.2217
(Trimble, 2015). From the resulting set of polygonal objects a training set of objects with 10218
samples for each of the classes ice, snow and other was produced. The training objects were219
used in a simple fuzzy neighbour classification (Trimble, 2015) as implemented in eCognition220
Developer. Object characteristics used for classification were the mean and standard deviation221
of the pixel values present within an object. The glacier outline was corrected by manual222
digitisation during a visual inspection of the classification, this was needed in particular at the223
southern tongue which was partly debris covered. In addition, the collected UAV imagery did224
not cover the entire glacier, some of the uppers parts of Storbreen were missing (Figure 2). To225
obtain a complete glacier outline, the upper part was replaced by the 2009 outline. Reductions226
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
11
in surface area of the glacier and terminus retreat over the past years were determined by227
comparing the 2015 and 2009 outlines.228
3.5. Elevation change229
To determine the elevation loss of Storbreen over the period 2009–2015 the post processed230
UAV point cloud from 2015 was compared with the LIDAR point cloud from 2009. Cloud-231
to-cloud differences were calculated in CloudCompare using the robust comparison algorithm232
Multiscale Model to Model Cloud Comparison (Lague et al., 2013), i.e. M3C2. The233
determined differences were gridded to a 2 m grid for further analysis.234
3.6. Geodetic mass balance235
Based on the surface elevation change between 17 October 2009 and 9-10 September 2015236
a geodetic mass balance was derived. First, the UAV DEM was masked out for areas with a237
RMSE > 0.2 m (Figure 7) or less than 100 meter from the glacier boundary to avoid any238
undesired effects due too steep slopes. Based on this mask (Figure 7), the average difference239
in surface elevation with the LIDAR DEM was computed. The mean elevation change for the240
area were then corrected for differences in the acquisition date and converted to mass using a241
density conversion (see results and discussion).242
The surface elevation difference was subsequently corrected for fresh snow, which was243
present in 2009. As a final step the average annual mass balance was computed over the244
masked area over the 6 year period and an error estimate that includes conservative error245
estimates for the LIDAR and UAV measurements, snow depth and snow and ice densities.246
4. Results and discussion247
The ortho-mosaic of the 2015 campaign shows that the glacier was still mostly snow248
covered and only the lower parts of both tongues and some parts in the ice fall were snow free249
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
12
(fig. 3a). The snow line at the time of the measurements was about 1570 m asl. which has not250
been so low at the ablation measurements since 1990 (Data: NVE). The snow depth251
measurements ranged between 0.45 and 1.83 m, with a mean of 1.31 m. Interpolation of the252
31 points gave a mean of 0.85 m over the snow covered parts (Figure 7).253
As the SfM workflow relies on matching features based on the Scale Invariant Feature254
Transform (SIFT) in overlapping pictures its application to pristine white snow surfaces could255
potentially be difficult. However, using the algorithm it was feasible to generate a sufficiently256
dense spare point cloud to derive a continuous ortho-mosaic and DEM for the entire glacier,257
including the pristine white upper part of the glacier. This can be attributed to a relatively258
high height above the surface, e.g. large area coverage per picture and the presence of small259
scale depressions and surface patterns due to melt and wind erosion on the snow covered260
glacier picked up by the SIFT algorithm. These findings are in line with previous studies261
which also do not report any major shortcoming as a result of over saturated pictures or lack262
of contrast over snow surfaces (Jagt et al., 2015; De Michele et al., 2016; Nolan et al., 2015).263
The accuracy of the generated UAV DEM was assessed in four different ways: (i) cross-264
check with the GCPs, (ii) comparison with the GNSS track over the glacier surface, (iii)265
comparison with stake observations of 2015 and (iv) assessment of elevation differences266
between the LIDAR and the UAV DEM outside the glacier. In panel A and B of Figure 4 the267
horizontal and vertical errors are shown for 30 GCPs. Since these GCPs were used in the268
processing they cannot be used as an independent validation, however the GCPs give269
information regarding locational errors resulting from the processing and ortho-rectification.270
The horizontal GCP RMSE was 0.28 m. and the vertical GCP RMSE was 0.22 m. In panels C271
and D the horizontal and vertical errors are plotted for independent points which were272
randomly selected from the DGPS track on the glacier. The horizontal and vertical track273
RMSEs are 0.36 m. and 0.44 m. respectively, which are slightly higher than for the GCPs. An274
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
13
additional source of error in the latter case is the fact that the antenna was carried in a275
backpack and its height and horizontal positions may vary while walking. It is also important276
to note that the vertical errors are well distributed around 0 for both the GCPs and the track277
and the mean vertical biases are 0.05 and 0.07 m. respectively.278
In 2015 the position and surface elevation of the stakes at the glacier surface were also279
measured using the GNSS one day before and during the same days that the flights were280
conducted (Figure 2). A comparison between the 2015 data for 10 stakes and the UAV DEM281
shows that the RMSE is 0.67 m, and if one outlier associated to a GNSS measurement error is282
omitted the RMSE is 0.41 m.283
The elevation difference between the 2009 LIDAR DEM and the 2015 UAV DEM in the284
off-glacier area was also compared as an independent check (Figure 7). The elevation285
difference based on the point cloud comparison also takes into account horizontal shifts286
between the point clouds. The average difference between both DEMs is -0.83 ± 0.78 m for287
the off-glacier area, suggesting that the 2009 DEM is consistently lower than the 2015 DEM.288
This is remarkable since the GCPs and track validation do not reveal a systematic bias. A289
possible explanation may be the presence of snow during the LIDAR campaign in 2009. The290
LIDAR DEM was acquired on 17 October 2009 and othophotos taken on September 14 and291
October 17 reveal that a snow layer had built up on the glacier and in the glacier forefield in292
this period. According to SeNorge, an operational snow model with 1×1 km resolution that293
uses gridded observations of daily temperature and precipitation as forcing (Saloranta, 2012)294
the snow depth is 29 cm for the off-glacier area. However, this is only an estimate of the snow295
depth. Previous studies have shown that seNorge may underestimate the precipitation at296
Storbreen and cannot describe the local accumulation characteristics in detail (Engelhardt et297
al, 2012). Temperature data from nearby met station Sognefjellshytta (1413 m. a.s.l.) reveal298
temperatures below freezing point from 28 September 2009 onwards and precipitation data299
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
14
from Bøverdalen (701 m a.s.l.) shows a total of 63 mm of snow between 28 September and 17300
October. An average snow depth of 29 cm may therefore be plausible. If accounting for the301
snow this would still imply a systematic difference of 59 cm between the DEMs outside the302
glacier. However, the difference may be explained by a larger error in the UAV DEM in the303
off-glacier area, because there are few GCPs here. In AgiSoft the sparse point cloud is304
geometrically corrected using the GCPs, however in areas at the margins of the region of305
interest without GCPs this may cause geometrical artefacts. In the off-glacier area there is306
indeed larger estimated error, there seems to be a slight north-south gradient in the error and307
that are areas where the RMSE reach values of 0.5 meter (Figure 7). Hence, the assessment of308
the off-glacier elevations reveals mostly an artefact of the UAV processing and has no bearing309
on the on-glacier accuracy.310
The experiments where the number of markers used in the AgiSoft processing is varied311
shows that both the distribution across the glacier and the total number has great bearing on312
the quality of the generated DEM (Figure 6). The average deviations from the reference DEM313
where all markers were used are -0.09±0.16 m, 0.02±0.44 m, -0.04±0.11 m for the 5, 10, 20314
marker experiments respectively. The 10 marker experiment reveals that the northern and315
southern parts show relative large deviations. This is caused by the fact that the 10 markers316
selected are all located on the central part of the glacier. For the 5 marker experiment the317
markers are more equally distributed across the glaciers and the deviations from the reference318
DEM is smaller than for the 10 marker experiment, even while the number of markers is319
halved. The 20 marker experiment shows the best result with only a small average difference320
and a narrow distribution of the differences. In this case the markers are also relatively well321
distributed across the glacier surface.322
The surface elevation difference map between 2009 and 2015 reveals that the glacier has323
lowered over the entire surveyed parts in this period. The lower tongue has lowered more than324
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
15
the upper part of the glacier. The lower tongue has lowered between 8 and 10 meters over the325
6 years, whereas the upper part of the glacier has lowered between 3 to 6 meters, except for a326
small area in the northwestern part of the upper glacier, which is an exposed gully of about 20327
meter depth and it may be subject to a microclimate, windy conditions and/or a specific328
radiation budget. The termini of both tongues show the largest elevation change up to 15329
meters in the northern terminus to 11 meters on the southern terminus.330
The surface elevation changes between October 2009 and September 2015 were used to331
derive a geodetic mass balance for Storbreen. Only the area within the 2009 extent were used332
and low accuracy parts (with an UAV RMSE > 0.20 m and more than 100 meter away from333
the glacier margin to avoid effects of shading and steep slopes) were excluded in the analysis.334
The total unmasked area is 3.88 km2
(77% of the total glacier area) (Figure 7). The average335
elevation change within this area was -5.30 m. To convert the mean elevation difference in336
meters to a geodetic mass balance in m w.e. and compare with the glaciological mass balance337
calculated in this period, one must account for ablation and accumulation between the338
glaciological and the geodetic surveys and make a density conversion (e.g. (Zemp et al.,339
2013)) The survey for the 2009 DEM was 17 October, a month later than the ablation340
measurements. The snow cover at this day is part of the 2009/2010 mass balance, and results341
in a higher surface elevation of the 2009 DEM. In 2015, the UAV survey was at the same342
time as the ablation measurements. The remaining snow is part of the glaciological mass343
balance for the mass balance year 2014/2015 and thus the comparison period 2009-2015.344
However, the remaining snow has a lower density than that of ice. In geodetic calculations, it345
is a common approach to assume an unchanged density profile from the surface to the firn–ice346
transition following Sorge’s law (Bader, 1954). A density conversion factor of 850±60 kg m-3
347
has been shown to be appropriate for a wide range of conditions (Huss, 2013). However, for348
shorter periods the density conversion factor can vary significantly.349
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
16
As mentioned, for the 2009 DEM snow depths we used the SeNorge model simulations350
giving snow depths of 0.29 m.351
Correcting for a mean snow layer of 0.29 m gives a dH of -5.01 m. Assuming Sorges law352
and applying a density conversion factor of 850 ± 100 kg m-3
results in a geodetic mass353
balance of -4.26±0.60 m w.e. Alternatively, accounting for the lower density of the remaining354
snow in 2015 gives a slightly more negative balance of -4.47 ± 0.60 m w.e. The mean balance355
over the six years is then -0.71 ± 0.1 m w.e. and -0.75 m ± 0.1 w.e. respectively.356
The cumulative glaciological mass balance over the six balance years from 2009/10 to357
2014/2015 is -4.8 m ± 1.1 w.e. or -0.80 m ± 0.18 w.e. a-1
(Kjøllmoen et al., 2016).358
The new 2015 survey reveal a significant retreat of the terminus since 2009. The southern359
terminus has retreated around 50 m, whereas the northern termini has retreated about 100 m.360
NVE’s length change measurements, conducted on the southern tongue, show a retreat of 49361
m from 2009 to 2014, or 9.8 m/a for the five years. In 2015, the southern terminus was snow362
covered and the front position was therefore not measured. GNSS survey of the southern363
terminus on 18 September 2014 (the terminus was snow free when measured) show that the364
tongue was at nearly the same position in 2014 and 2015, and thus the retreat of the southern365
tongue occurred from 2009 to 2014, the northern tongue was snow free and likely retreated366
also in 2015. The total glacier area reduction was 0.06 km2
, a 1.2 % reduction of the 2009367
area.368
5. Conclusions369
In this study a UAV was used on a mountain glacier in Norway to evaluate its potential for370
mapping and quantifying the surface mass balance. The UAV results were compared to a371
LIDAR dataset and to the glaciological mass balance and the accuracy of the UAV DEM was372
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
17
assessed using markers and tracks on the glacier, stake measurements and by conducting373
experiments with varying numbers of markers used in the UAV image processing.374
It is concluded that UAVs are an attractive alternative or complementary tool to375
“traditional” methods such as aerial photography, LIDAR and satellite imagery. The analysis376
shows that the accuracy of the generated DEM is relatively high and sufficient to quantify the377
surface mass balance. Key advantages over traditional methods are (i) that the UAV can be378
used at an optimal time under the best possible weather and light conditions, (ii) that the379
resolution of the output is very high and (iii) that a DEM and an ortho-mosaic are acquired380
simultaneously. Disadvantages are that the accuracy may not be high enough for annual381
campaigns and that the surveys must be accompanied by a number of markers on the surface382
that requires additional fieldwork. Furthermore, the UAV may not be covering the entire383
glacier if a suitable launching spot within a horizontal distance of ~2 km and a vertical384
distance of ~500 meter is unavailable.385
The 2015 campaign on Storbreen revealed that the SfM alghoritm also performs well on386
glacier covered in snow. As long as there are small surface patterns due to wind erosion and387
melt and the flight altitude is relative large to ensure sufficient variation within a single388
image, it is feasible to derive an accurate surface DEM also for snow covered surfaces. This389
provides the opportunity to use UAVs in the annual mapping of glaciers under varying390
conditions. It is recommended to repeat the campaign under low snow conditions (with a391
higher transient snow line) to assess whether accuracy of the DEM improves as a result of392
better contrast and more texture on the glacier surface.393
The analysis shows that the use of markers measured by GNSS is essential to derive an394
accurate ortho-mosaic and DEM from the UAV imagery. The number and the distribution of395
the markers are important for the accuracy of the final products. It is essential to distribute the396
markers evenly across the area of interest and to have at least markers at the margins of the397
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
18
area of interest. An average of about 6 markers / km2
equally distributed over the area of398
interest seems to provides accurate results in this case, yet its applicability elsewhere depends399
on the terrain morphology, flight altitude, light and surface conditions. GNSS measurements400
covering the entire glacier are however labor intensive and it is recommended to acquire401
GNSS measurement outside the glacier area in relative flat areas which are easily identifiable,402
stable and will not experience surface elevation changes, e.g. center points of large boulders.403
These points can be used in subsequent UAV missions as virtual markers in the image404
processing if GNSS campaigns are not feasible.405
The ortho-mosaic generated from the UAV imagery in combination with OBIA provides a406
suitable approach for the semi-automated mapping of the snow line and the terminus position.407
However, a manual inspection and digitization is needed to correct for debris covered parts of408
the glacier.409
6. Acknowledgements410
The authors thank Ånund Kvambekk, NVE, for field assistance and Bjarne Kjøllmoen,411
NVE, for dGNSS processing. This project has received funding from the European Research412
Council (ERC) under the European Union’s Horizon 2020 research and innovation program413
(grant agreement No 676819).414
7. References415
Agisoft LLC: Agisoft PhotoScan User Manual Professional Edition, Version 1.1, Agisoft,416
St. Petersburg, Russia., 2014.417
Andreassen, L. M.: Comparing traditional mass balance measurements with long-term418
volume change extracted from topographical maps: a case study of Storbreen glacier in419
Jotunheimen, Norway, for the period 1940–1997, Geogr. Ann., 81A(4), 467–476, 1999.420
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
19
Andreassen, L. M., Elvehøy, H., Kjøllmoen, B. and Engeset, R. V.: Reanalysis of long-421
term series of glaciological and geodetic mass balance for 10 Norwegian glaciers, Cryosph.,422
10(2), 535–552, doi:10.5194/tc-10-535-2016, 2016.423
Bader, H.: Sorge’s law of densification of snow on high polar glaciers, J. Glaciol., 2(15),424
319–323, 1954.425
Bhardwaj, A., Sam, L., Martín-torres, F. J. and Kumar, R.: Remote Sensing of426
Environment UAVs as remote sensing platform in glaciology : Present applications and future427
prospects, Remote Sens. Environ., 175, 196–204, doi:10.1016/j.rse.2015.12.029, 2016.428
Blaschke, T.: Object based image analysis for remote sensing, ISPRS J. Photogramm.429
Remote Sens., 65(1), 2–16, doi:10.1016/j.isprsjprs.2009.06.004, 2010.430
Cogley, J. G.: Geodetic and direct mass-balance measurements : comparison and joint431
analysis, Ann. Glaciol., 50(50), 96–100, 2009.432
Harder, P., Schirmer, M., Pomeroy, J. and Helgason, W.: Accuracy of Snow Depth433
Estimation in Mountain and Prairie Environments By an Unmanned Aerial Vehicle, Cryosph.434
Discuss., (February), 1–22, doi:10.5194/tc-2016-9, 2016.435
Huss, M.: Density assumptions for converting geodetic glacier volume change to mass436
change, Cryosph., 7(4), 877–887, doi:DOI 10.5194/tc-7-877-2013, 2013.437
Immerzeel, W. W., Kraaijenbrink, P. D. A., Shea, J. M., Shrestha, A. B., Pellicciotti, F.,438
Bierkens, M. F. P. and de Jong, S. M.: High-resolution monitoring of Himalayan glacier439
dynamics using unmanned aerial vehicles, Remote Sens. Environ., 150, 93–103,440
doi:10.1016/j.rse.2014.04.025, 2014.441
Jagt, B., Lucieer, A., Wallace, L., Turner, D. and Durand, M.: Snow Depth Retrieval with442
UAS Using Photogrammetric Techniques, Geosciences, 5(3), 264–285,443
doi:10.3390/geosciences5030264, 2015.444
Kääb, a., Treichler, D., Nuth, C. and Berthier, E.: Brief Communication: Contending445
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
20
estimates of 2003 - 2008 glacier mass balance over the Pamir–Karakoram–Himalaya,446
Cryosph., 9(2), 557–564, doi:10.5194/tc-9-557-2015, 2015.447
Kjøllmoen, B., Andreassen, L. M., Elvehøy, H., Jackson, M. and Giesen, R. H.:448
Glaciological investigations in Norway 2011-2015, Oslo., 2016.449
Kraaijenbrink, P., Meijer, S. W., Shea, J. M., Jong, S. M. D. E. and Immerzeel, W. W.:450
Seasonal surface velocities of a Himalayan glacier derived by automated correlation of451
unmanned aerial vehicle imagery, Ann. Glaciol., 57, 103–113, 2016a.452
Kraaijenbrink, P. D. A., Immerzeel, W. W., Pellicciotti, F. and Shea, J. M.: Distribution453
and characteristics of surface features on a debris-covered glacier, an object-based analysis,454
Remote Sens. Environ. (in Prep., 2016b.455
Kraaijenbrink, P. D. A., Shea, J. M., Pellicciotti, F., De Jong, S. M. and Immerzeel, W. W.:456
Object-based analysis of unmanned aerial vehicle imagery to map and characterise surface457
features on a debris-covered glacier, Remote Sens. Environ., 186, 581–595,458
doi:10.1016/j.rse.2016.09.013, 2016c.459
Lague, D., Brodu, N. and Leroux, J.: Accurate 3D comparison of complex topography with460
terrestrial laser scanner: Application to the Rangitikei canyon (N-Z), ISPRS J. Photogramm.461
Remote Sens., 82(February 2013), 10–26, doi:10.1016/j.isprsjprs.2013.04.009, 2013.462
Liestøl, O.: Storbreen Glacier in Jotunheimen, Norway, Nor. Polarinst. Skr., 141, 5–62,463
1967.464
Lucieer, A., Jong, S. M. and Turner, D.: Mapping landslide displacements using Structure465
from Motion (SfM) and image correlation of multi-temporal UAV photography, Prog. Phys.466
Geogr., 38(1), 97–116, doi:10.1177/0309133313515293, 2013.467
Magnússon, E., Belart, J. M. C., Pálsson, F., Ágústsson, H. and Crochet, P.: Geodetic mass468
balance record with rigorous uncertainty estimates deduced from aerial photographs and469
LiDAR data – case study from Drangajökull ice cap, NW-Iceland, Cryosph. Discuss., 9(5),470
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
21
4733–4785, doi:10.5194/tcd-9-4733-2015, 2015.471
De Michele, C., Avanzi, F., Passoni, D., Barzaghi, R., Pinto, L., Dosso, P., Ghezzi, A.,472
Gianatti, R. and Vedova, G. Della: Using a fixed-wing UAS to map snow depth distribution:473
An evaluation at peak accumulation, Cryosphere, 10(2), 511–522, doi:10.5194/tc-10-511-474
2016, 2016.475
Nolan, M., Larsen, C. and Sturm, M.: Mapping snow depth from manned aircraft on476
landscape scales at centimeter resolution using structure-from-motion photogrammetry,477
Cryosphere, 9(4), 1445–1463, doi:10.5194/tc-9-1445-2015, 2015.478
Nuth, C. and Kääb, a.: Co-registration and bias corrections of satellite elevation data sets479
for quantifying glacier thickness change, Cryosph., 5(1), 271–290, doi:10.5194/tc-5-271-480
2011, 2011.481
Pellicciotti, F., Stephan, C., Miles, E., Immerzeel, W. W. and Bolch, T.: Mass balance482
changes of the debris-covered glaciers in the Langtang Himal in Nepal between 1974 and483
1999, J. Glaciol., 61, 373–386, 2015.484
Rolstad, C., Haug, T. and Denby, B.: Spatially integrated geodetic glacier mass balance485
and its uncertainty based on geostatistical analysis: Application to the western Svartisen ice486
cap, Norway, J. Glaciol., 55(192), 666–680, doi:10.3189/002214309789470950, 2009.487
Ryan, J. C., Hubbard, a. L., Box, J. E., Todd, J., Christoffersen, P., Carr, J. R., Holt, T. O.488
and Snooke, N.: UAV photogrammetry and structure from motion to assess calving dynamics489
at Store Glacier, a large outlet draining the Greenland ice sheet, Cryosph., 9(1), 1–11,490
doi:10.5194/tc-9-1-2015, 2015.491
Saloranta, T. M.: Simulating snow maps for Norway: Description and statistical evaluation492
of the seNorge snow model, Cryosphere, 6(6), 1323–1337, doi:10.5194/tc-6-1323-2012,493
2012.494
Szelisky, R.: Computer Vision: Algorithms and Applications, edited by D. Gries and F.495
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
22
Schneider, Springer-Verlag, London., 2011.496
Trimble: eCognition Developer User Guide, Version 9.1.2, Trimble Documentation,497
Munich, Germany., 2015.498
Vincent, C., Wagnon, P., Shea, J. M., Immerzeel, W. W., Kraaijenbrink, P. D. A.,499
Shrestha, D., Soruco, A., Arnaud, Y., Brun, F., Berthier, E. and Sherpa, S.: Reduced melt on500
debris-covered glaciers: investigations from Changri Nup Glacier, Nepal, Cryosph. Discuss.,501
1–28, 2016.502
Westoby, M. J., Brasington, J., Glasser, N. F., Hambrey, M. J. and Reynolds, J. M.:503
“Structure-from-Motion” photogrammetry: A low-cost, effective tool for geoscience504
applications, Geomorphology, 179, 300–314, doi:10.1016/j.geomorph.2012.08.021, 2012.505
Westoby, M. J., Dunning, S. A., Woodward, J., Hein, a. S., Marrero, S. M., Winter, K. and506
Sugden, D. E.: Interannual surface evolution of an Antarctic blue-ice moraine using multi-507
temporal DEMs, Earth Surf. Dyn., 4, 515–529, doi:10.5194/esurf-4-515-2016, 2016.508
Winsvold, S. H., Andreassen, L. M. and Kienholz, C.: Glacier area and length changes in509
Norway from repeat inventories, Cryosphere, 8(5), 1885–1903, doi:10.5194/tc-8-1885-2014,510
2014.511
Zemp, M., Thibert, E., Huss, M., Stumm, D., Denby, C. R., Nuth, C., Nussbaumer, S. U.512
and Moholdt, G.: The Cryosphere Reanalysing glacier mass balance measurement series, ,513
1227–1245, doi:10.5194/tc-7-1227-2013, 2013.514
515
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
23
Tables
Table 1 Overview of the flight details
Flight Date Start
time
End
time
Duration Images Altitude
(m)
Area
(km2
)
Comments
1 09 Sep 15 10:09 10:27 0:18 168 359 1.50 -
2 09 Sep 15 11:03 11:24 0:21 210 251 1.99 -
3 09 Sep 15 12:02 12:08 0:06 70 238 0.86 Camera battery malfunction
4 09 Sep 15 12:49 12:56 0:07 76 230 0.64 Camera battery malfunction
5 10 Sep 15 10:19 10:35 0:16 160 318 1.28 Launched from terminus
6 10 Sep 15 11:10 11:29 0:19 231 245 1.23 Launched from terminus
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
24
Figures
Figure 1 Location map of Storbreen in Norway. (source of contour lines: http://data.kartverket.no/download/content/n250kartdata-utm33-hele-landet-fgdb)
The glacier extent is from the 2009 survey.
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
25
Figure 2 Overview of the flight tracks, GCPs, photo locations, GNSS points and launch site (panel A) and ablation stakes,
snow depth probes and snow density pits (panel B)
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
26
Figure 3 The ortho-mosaic and DEM of Storbreen from the 2015 UAV campaign at 0.25 m resolution
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
27
Figure 4 Errors between the SfM derived orthomosaic (horizontal) and elevation model (vertical), and the GNSS
measurements of the 30 GCPs (a, b) and the independent points selected randomly from the DGPS track (c (n=41), d
(n=150)). Random points were selected only from the parts of the GNSS track further than 50 meter from a GCP and where
the rover was moving
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
28
Figure 5 Spatial RMSE of the 2015 UAV DEM based on residual analysis of linear correlation of dense point cloud within 2
meter bins.
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
29
Figure 6 Impact of the number and distribution of markers used in the Agisoft processing. The black dots are markers used in
the processing, whereas white dots are not used. All plots show elevation differences relative to case where all 30 markers
were used. The boxplot shows the distribution of elevation differences for the three experiments (5 (panel A), 10 (panel B)
and 20 (panel C markers respectively).
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
30
Figure 7 Elevation changes between 2009 LIDAR and 2015 UAV determined by cloud to cloud comparison in
CloudCompare. The interpolated snow depth is shown as contour lines.
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.
31
Figure 8 The retreat of the Storbreen glacier between 1984 and 2015 shown on the 2015 orthophoto. Outlines are from the
glacier maps, except 2014, which is a GNSS survey of part of the southern terminus.
The Cryosphere Discuss., doi:10.5194/tc-2016-292, 2017
Manuscript under review for journal The Cryosphere
Published: 5 January 2017
c Author(s) 2017. CC-BY 3.0 License.