Jökull - 01.01.2019, Qupperneq 109

Tussetschläger et al.
referenced based on the DEM from LMI and geo-
referenced orthophotos of 0.5 m resolution from the
private company Loftmyndir. Subsequently, at least
eight ground control points (mountain peaks, roofs of
houses, churches etc.) were selected in each area and
used for connecting the aerial images to the referenced
orthophotos. We used the aerial photos for valida-
tion purposes, by extracting snow patches based on a
manually identified threshold value of the reflectance,
which was kept equal for all photos. Orthophotos
(from Loftmyndir) with a spatial resolution of 0.5 m
and covering the study areas are available from the
year 2000 (Brimnesdalur, Kerling and Búrfellsdalur;
date unknown), eighth of August 2014 (Kerling and
Sakka) and nineteenth of August 2016 (Almenningar,
Úlfsdalir and Brimnesdalur). Areas of snow patch de-
rived from aerial images and orthophotos were com-
pared with areas of snow derived from the satellite
images. Furthermore, single snow patches were se-
lected and compared with snow patches in satellite
images. Another opportunity to validate identified
snow patches was through comparison with photos
taken in the field. In some areas, e.g. Búrfellsjök-
ull, glaciological mass balance has been determined
since 2006 (Brynjólfsson, 2018 and 2019) and photos
and valuable data about snow conditions are available
since then. If a picture has been taken on the same
date or close to the acquisition date of a satellite im-
age, it can be used for comparison with the snow patch
classification. If the time gap is longer, fieldwork data
can still be useful for mapping the development of
the snow distribution during a summer. In this study,
field work photos from the years 2015, 2016, 2017
and 2018 are used for comparison.
The locations of the identified snow patches were
analysed in connection with the prevailing tempera-
ture and precipitation in the area as well as the devel-
opment of the snow patches. In the research area there
are several weather stations, operated by the Icelandic
Meteorological Office (Figure 1). The weather sta-
tions in Siglufjörður, Ólafsfjörður, on Grímsey island,
the Öxnadalsheiði pass and Vaðlaheiði pass measure
temperature, dew point, wind speed and wind direc-
tion and the weather station on Tindaöxl mountain
measures temperature and snow depth. For precip-
itation we used data from the weather stations lo-
cated in Ólafsfjörður, Akureyri and Skeiðsfoss (Ice-
landic Met. Office). The stations measure precipita-
tion every 10 minutes and we used 1 hour mean val-
ues. Furthermore, we have used a MAAT raster with
1x1 km grid size, which is based on measurements
and interpolated with a DEM, to calculate tempera-
ture decrease by elevation (monthly average in tem-
perature decrease by elevation was used) (Björnsson
et al., 2007).
Perennial Snow Patch Classification
The semi-automatic processing procedure for PSPs
classification from optical satellite data is explained
in the following section. At first, the input data
(Sentinel-2 and Landsat-5/-7/-8 images) were down-
loaded and pre-processed. An atmospheric correc-
tion of the Sentinel-2 (Level-1C) images was applied
with the Sen2Cor tool implemented in the Sentinel
Application Platform (SNAP) (Muller-Wilm et al.,
2017). The Level-1C images with Top of Atmosphere
(TOA) values were atmospherically corrected, clas-
sified in different scenes and converted into images
with reflectance values (Level-2A products). The im-
ages were then resampled to an equal grid spacing
of 10 m. Furthermore, the spectral band 11 was
resampled from 20 m to 10 m and a cloud mask
was generated using an integrated tool in SNAP. The
Landsat images were pre-processed using the Semi-
Automatic Atmospheric Correction Plugin for QGIS
(Luca, 2017). This processor uses an image based
technique to perform a simple correction using DOS1
(Dark Object Subtraction 1) method. However, stud-
ies (e.g. Winsvold et al., 2016) have shown that an
atmospheric correction has only little influence on the
satellite images if the identified class is snow or ice
(against bare ground or vegetation).
The Normalized Difference Snow Index (NDSI)
was calculated using the ratio of the green band and
the Shortwave Infrared (SWIR) Band (see Equation
1, Dozier, 1989; Dozier and Painter, 2004):
NDSI=(Green – SWIR)/(Green + SWIR) (1)
In this approach a pixel is classified as snow if the
NDSI is greater than 0.3. For Sentinel-2 the blue band
(Band 2), which distinguishes snow in shadow, must
108 JÖKULL No. 69, 2019