CHAPTER 4
Glacier mapping and monitoring using
multispectral data
Andreas Kääb, Tobias Bolch, Kimberly Casey, Torborg Heid, Jeffrey S. Kargel,
Gregory J. Leonard, Frank Paul, and Bruce H. Raup
ABSTRACT
Multispectral satellite data represent the primary
data source for spaceborne glacier mapping and
monitoring, and remote-sensing studies have gener-
ated significant results regarding global glacier
observations and understandings. In this chapter
we provide an overview of the use of multispectral
data and the methods typically applied in glacier
studies. Besides multispectral techniques based on
the visible and near-infrared section and the short-
wave infrared section of the spectrum, we also
briefly discuss methods for analyzing thermal and
radar data, with special emphasis on the mapping of
debris-covered glacier ice. A further focus is on
spectral change detection techniques applied to
multitemporal data, with special attention to a
novel image-differencing technique. Then we pro-
vide an overview of satellite image–based measure-
ment of glacier flow. Finally, we offer a suggestion
for a new combination of glacier observations to be
made by both multispectral and radar/microwave
remote-sensing sensors.
4.1 INTRODUCTION
There are numerous challenges associated with
extracting accurate information from satellite
imagery for assessing glaciers and climate change.
Numerous methods exist, although the effective
use of such methods varies depending upon data
types and environmental conditions, in addition
to applications.
In general, remote sensing of glaciers is feasible
due to the large difference in snow and ice reflec-
tance in the visible (VIS) and near-infrared (NIR)
(combined: VNIR), and the shortwave infrared
(SWIR) regions of the spectrum. This relatively
unique shape of the spectral curve and spectral
differences (VIS versus SWIR) represents the basis
for most multispectral glacier-mapping applica-
tions. On the other hand, remote sensing of glaciers
is complicated by the fact that glacier surfaces other
than snow, firn, and clean ice (e.g., debris-covered
ice) can be spectrally very similar to rocks and
sediment on the surrounding landscape. In fact,
even water containing abundant suspended rock
particles can have a multispectral reflectance similar
to rock (see Figure 3.8 of this book by Furfaro et
al.). In radar images, backscatter contrast between
ice and snow, and other terrain can be small and
temporally unstable, so that the corresponding
image segmentation is usually not straightforward.
Finally, whereas many image analysis methods
result in sharp boundaries between different classes,
nature also contains gradual transitions that may
not be adequately characterized given the nature of
the data and/or the methods used (e.g., gradual
change of surface debris cover within a pixel).
These capabilities and problems characterize most
76 Glacier mapping and monitoring using multispectral data
methods that are used for glacier mapping and
monitoring.
This chapter provides an overview of various
approaches for mapping and monitoring glaciers
using space-based spectral data. We address the
information characteristics and accuracy associated
with information products. Specifically, we con-
sider:
. image segmentation and classification for map-
ping glacier outlines and glacier zones;
. spectral detection of changes in glacier boundaries
and zones over time using multitemporal data;
and
. geometric detection of ice displacements (ice flow)
using multitemporal data.
These three categories categorize the nature of
glacier observations from space. We do not address
geometric analysis of stereo data for derivation of
digital elevation models (DEMs), as this is found in
Chapter 5. Furthermore, our main focus is on the
VIS to SWIR regions of the electromagnetic spec-
trum. Thermal and microwave data and methods
are briefly covered, as are Synthetic Aperture
RADAR (SAR) and Interferometric SAR
(InSAR).
4.2 IMAGE PREPROCESSING
In most cases, satellite imagery for glacier mapping
and monitoring requires radiometric and geometric
preprocessing. Radiometric calibration transforms
at-sensor raw digital numbers to meaningful units
(i.e., radiance, reflectance) to account for sensor
and atmospheric conditions, while geometric pre-
processing addresses pixel locations and geometric
fidelity. Specifically, radiometric calibration can
include corrections for instrumental gain and offset,
illumination conditions, as well as atmospheric
absorption and scattering. Geometric preprocessing
can include geolocation, registration, and/or re-
projection of image geometries, in addition to
inter-sensor co-registration (e.g., between VNIR,
SWIR, and TIR sensors and optical paths). In
Chapter 6 of this book, Ramachandran et al. review
the standard types of preprocessing and available
image products for the ASTER project. Further
preprocessing may be needed, depending on the
applications, as we briefly discuss below.
4.2.1 Radiometric calibration
A first step in radiometric preprocessing is the cor-
rection of systematic differences and variations in
the sensitivity of detectors. In pushbroom imagery,
such as ASTER, these sensitivity differences pro-
duce along-track stripes. The necessary radiometric
correction coefficients (gain and offset, or higher
order polynomials) are provided with the raw image
data (ASTER Level 1A) or are already applied
(Level 1B and up). Scanning instruments, such as
Landsat, may have comparable effects resulting
from across-track scanning (Crippen 1989).
Subsequent steps of radiometric preprocessing
include atmospheric correction and topographic
normalization which represents the correction of
topographically induced illumination variations.
A number of techniques, however, can be applied
to satellite data without correcting for the effects of
the atmosphere or topography such as band ratios,
normalized differences, or normalizations within
image matching that avoid or reduce such effects.
It has been shown that errors in radiometric correc-
tions in complex mountains can be substantial and
may not necessarily improve uncorrected data as a
whole for mapping purposes (Paul 2001).
Multiangular or multitemporal applications may
also be affected by variation of the bidirectional
reflectance distribution (Fig. 4.1). Such effects are
difficult to correct, but can in some cases be avoided
by comparing data from similar acquisition pos-
itions and under similar illumination conditions.
4.2.2 Geometric preprocessing
Here, we do not cover band-to-band or sensor-to-
sensor co-registrations, which are usually done at
an early processing level of satellite data. After this,
the two most important types of geometric prepro-
cessing are transformation of the image data to a
map projection, and removal of topographic distor-
tions. Depending on the processing level at which
the base data are provided, image data may come in
raw swath geometry, georeferenced (data in raw
geometry, but transformation information pro-
vided to map geometry given), or geocoded (trans-
formed to map geometry using a reference surface,
usually the ellipsoid). Different data providers
usually have slightly different terminologies for
their processing levels, or hybrid levels might exist
(e.g., ASTER L1B, which is georeferenced and cor-
rected for Earth rotation during over-flight).
 Image preprocessing 77

Figure 4.1. (Left) Section of an ASTER 3N NIR image over Svalbard, near Longyearbyen. (Right) Normalized
difference index image between the orthorectified 3N and 3B images. Green colors indicate no difference between
both images, red and blue indicate differences in digital numbers. Differences are due to reflectance differences
under different viewing directions (i.e., effects from the bidirectional reflectance distribution function, BRDF, of the
targets). Certain landforms (glaciers and floodplains) can be especially well distinguished in the difference image,
indicating a strong BRDF effect for these surface types in the NIR. Topographic distortions and occlusions in the 3B
image also contribute to image differences (e.g., steep north-facing flanks, shadows). Figure can also be viewed as
Online Supplement 4.1.

Orthoprojection (or orthorectification) repre-
sents the removal of topographic distortions and
sometimes involves applying terrain correction to
swath geometry data. This processing step requires
a DEM. The quality of the resulting orthoimages
depends on
. the quality of the spacecraft position and attitude
information (i.e., the pointing vector);
. the quality of ground control points (GCPs) in
case they are used to define the image orientation,
or to refine the latter as obtained from the sensor’s
pointing data; and
. the quality of the DEM used.
It is important to understand error propagation
from vertical DEM errors to horizontal position
errors in the resulting orthoimage. For a point with
a cross-track distance R from the nadir track (or-
thogonal projection of the satellite path to the ellip-
soid), an elevation error
h causes a cross-track
offset of
r for a flying height above ground H
described by:
R

r ¼
h ð4:1Þ
H co-registration accuracy better than one pixel is
For scanners and pushbroom sensors the resulting
offset is in the across-track direction. For frame
cameras, for instance, R and
r are in the radial
direction from the image nadir point.
Whereas errors in map transformation can be
roughly solved by co-registration of data to a refer-
ence dataset, errors in the topographic correction
cannot be easily corrected due to the usually non-
systematic variations of DEM errors over a scene.
Stereo sensors such as ASTER offer the possibility
to closely examine errors in orthoprojection (Kääb
2005b). Higher-order geometric distortions may
exist in the image data (Leprince et al. 2007, Nuth
and Kääb 2011), but for many applications will
remain uncorrected.
For many glacier-mapping applications, final
geolocation accuracy of the orthorectified data of
about one pixel will be sufficient (i.e., 30 m for the
widely used Landsat TM/ETMþ data or ASTER
when including SWIR bands). Glacier changes to
be observed have to exceed this accuracy level,
which defines the time interval at which these
changes can be mapped at a statistically significant
level for a given rate in glacier boundary change.
For measuring glacier displacements using offset
tracking between repeat image data, sometimes a
desirable, in particular for small movements that
require subpixel precision image matching.
78 Glacier mapping and monitoring using multispectral data
4.3 MULTISPECTRAL METHODS
4.3.1 Spectral reflectance of
glacier surfaces
Spectral separability is determined by variations in
reflectance in different regions of the spectrum
based upon an n-dimensional feature space. Here
we focus on the VNIR and SWIR spectrum (Fig.
4.2). Representative measured reflectance curves are
given in Fig. 4.3. In Chapter 3 of this book by
Furfaro et al. the spectra of pure and mixed
materials are calculated to show more systematic-
ally the influences of ice grain size, clastic sediment
content in ice, water turbidity, and other variables.
The next five subsections describe the reflectance
properties of common glacier and near-glacier sur-
face components.
4.3.1.1 Snow
Fresh snow diffusely reflects up to 95% of incoming
VIS and about 50–80% of NIR radiation. In the
VIS spectrum, snow reflectance decreases with
particulate inclusions such as dust, soot, and biota
(Wiscombe and Warren 1980, Warren 1982, Hall et
al. 1989, 1990, Painter et al. 2001, Takeuchi 2009,
Casey et al. 2012). In the near-infrared, the
corresponding influence of dust contamination
decreases, and increasing snow grain size reduces
snow reflectance (Painter 2011). In the shortwave
infrared, the reflectance of snow is very low, with a
marked dependence on grain size, and a lesser in-
fluence from snow contamination (Dozier 1989,
Bourdelles and Fily 1993). Furfaro et al. (Chapter
Figure 4.2. Atmospheric transmission, sections of the optical and microwave spectrum, and spectral band widths
of Landsat ETMþ, ASTER, and active microwave sensor bands. UV: ultraviolet; VIS: visible; NIR: near-infrared;
SWIR: shortwave infrared; MIR: middle infrared; TIR: thermal infrared; P–K: radar bands.
3 in this book) provide a detailed radiative transfer–
based approach to the reflectance of snow and ice.
The large contrast between VIS and SWIR reflec-
tance is exploited for snow classification (Rott 1976,
Dozier 1989). Under similar emissivity factors, TIR
and passive microwave emission of snow and ice is,
in comparison with other materials, limited by the
fact that the surface temperature is at or below 0 C.
4.3.1.2 Ice
As a consequence of the reflectance properties of
snow, bare glacier ice has a lower reflectance than
snow in the VIS spectrum due to the accumulation
of optically active contaminants (Warren and
Brandt 2008). This effect increases with dirty glacier
ice (Qunzhu et al. 1983, Koelemeijer et al. 1993). In
the NIR, reflectance decreases as snow grain size
and particulate concentration increases. In addi-
tion, the presence of liquid water on the ice surface
might lead to reduced reflectivity in the NIR (Rott
1976; Paul 2004). For debris-covered ice, the spec-
tral signature of debris may prevail over the ice
signature depending on the percentage of debris-
covered surface area within a pixel (Casey et al.
2012). If a glacier pixel is covered with more than
about 80% debris, it cannot be differentiated from
surrounding pixels representing bedrock or perigla-
cial debris.
4.3.1.3 Rock
Unlike snow and ice, rock or debris-covered sur-
faces show a significantly different reflectance in the
VNIR, SWIR, and TIR. This fact usually allows for
good spectral separability of such surfaces against

Figure 4.3. Atmospheric transmission, locations of ASTER and Landsat bands, and typical reflectance curves for
glacier surfaces (left) and materials found around glaciers (right).
clean snow and ice. The highly variable reflectance
of different mineral and rock types in the VNIR,
SWIR, and TIR forms the base for lithology map-
ping from multi and hyperspectral imagery (Clark
1999; Sabins 1999; Rowan and Mars 2003, Volesky
et al. 2003; Casey et al. 2012). Knowledge of the
geology of a high-mountain area supports geomor-
phological and geomorphodynamical analyses
including erosion and sediment transport (Casey
et al. 2012). There are also direct applications for
glaciological investigations. Detecting the spatial
distribution and the type of surface debris on
glaciers, for instance, allows the identification of
material origin and transport paths (Kääb 2005a,
Casey et al. 2012), and therefore conclusions on
present and past dynamics can be drawn.
4.3.1.4 Water
The reflection of open water in the VIS is highly
variable depending on its turbidity. In the NIR and
SWIR, water strongly absorbs radiation, indepen-
dent of its turbidity. NIR and SWIR reflection of
water is similar to snow and ice, and can lead to
misclassification. Therefore, the use of the VIS
region is required to separate both categories (Hug-
gel et al. 2002, Paul 2004). The VIS and TIR regions
can be used to assess the turbidity and temperature
of glacial lakes (Wessels et al. 2002).
4.3.1.5 Vegetation
Much research on multispectral remote sensing of
vegetation is available (Liang 2004, Jensen 2007).
Vegetation mapping usually takes advantage of
high reflectivity in the NIR. In glaciological
Multispectral methods 79
research, the existence of vegetation itself might
be the most interesting result since it potentially
points to, for instance, comparably stable surfaces,
plant succession, or lack of surface abrasion. First-
order vegetation mapping may be applied to
exclude distinct areas from further analyses or
remove misclassifications, although in some rare
cases vegetation is known to grow in soils overlying
stagnant glacier ice.
4.3.2 Image classification approaches
The surface signature in the spectral domain is used
to describe and distinguish surface types and con-
ditions. Such classifications may be characterized
(Schowengerdt 2007) by such terms as: hard/soft
classification, manual/supervised/unsupervised
classification, parametric/nonparametric classifica-
tion, spatial/spectral segmentation, pixel/subpixel
classification, and multispectral/hyperspectral clas-
sification. The following six subsections describe the
characterizations for these various data classifica-
tion types and methods.
4.3.2.1 Hard and soft classification
Hard classification results in the assignment of only
one category (aka landcover class) to a pixel, pro-
viding sharp boundaries between classes. In soft
classification, likelihood or uncertainty values for
multiple classes can be assigned to each pixel. In
glacial environments, both discrete (hard) and fuzzy
(soft) transitions exist, indicating that both
approaches can be utilized. In nature, the bound-
aries between, for instance, glacier and rock, or
snow and bare soil are very sharp, while their
80 Glacier mapping and monitoring using multispectral data
detectability in imagery depends upon sensor spa-
tial resolution. The classification goal is to deter-
mine the boundary position as accurately as
possible. On the other hand, transitions between
stagnant ice and moraine, for instance, or gradually
decreasing vegetation coverage are smooth (i.e.,
gradational), so that a change in class likelihood
might better reflect environmental boundaries in
nature.
The choice of hard versus soft classification
strategies depends not only on the natural charac-
teristics of the boundary between two surface
categories, but also on the spectral, spatial, and
radiometric resolution of the sensor. A category
transition in nature might be discrete in one part
of the spectrum, but fuzzy in another region. If the
spatial resolution of a sensor (ground-projected
instantaneous field of view, GIFOV) is significantly
larger than the spatial variations of a land cover
class, the resulting pixel contains a mixed signal
of more than one ground category (mixed pixel),
independent of whether the category transitions are
discrete or fuzzy. Such mixed pixels can be classified
‘‘hard’’ (e.g., using thresholding) or ‘‘soft’’ (e.g.,
using category percentages).
4.3.2.2 Manual, supervised, and unsupervised
classification
Manual delineation is used not only for simple clas-
sifications, but also for accuracy assessments,
ground data reference acquisition, or completion
or correction of other classifications. Supervised
classification involves intensive computer training
to statistically represent spectral variations. Train-
ing areas of the categories to be mapped are oper-
ator-selected a priori in order to develop spectral
signatures used as the basis for classification. Separ-
ability analysis can be applied to test whether the
selected categories can be spectrally differentiated at
a statistically significant level. The spectral signa-
tures are then used to automatically segment the
entire image, usually based upon probabilities or
spectral ranges. Unsupervised classification repre-
sents automatic differentiation of spectral clusters
in feature space, based upon an initial selection of
the number of classes desired. The image is auto-
matically segmented based upon cluster statistical
separability. Depending on the numbers of clusters
selected, cluster classes may represent artificial cat-
egories or biophysical conditions of the landscape.
Unsupervised classification is, therefore, a data-dri-
ven method that requires careful evaluation of clas-
sification results; an expert user selects final classes,
in many cases requiring the merging of class groups
from the classified output. All three classification
approaches may be combined in a variety of ways.
4.3.2.3 Parametric and nonparametric
classification
The use of a parametric algorithm for classification
assumes a certain statistical distribution and con-
ditions associated with a spectral sample (e.g.,
normal distribution, homogeneity of variance).
The statistical distribution parameters (i.e., mean
and variance) are estimated from training samples
obtained from the image under analysis, or inferred
from other sources. An individual pixel is assigned
to a certain class according to its statistical
proximity to the class parameters (e.g., nearest
mean, or maximum likelihood). For nonparametric
algorithms, class membership is not based upon
parametric parameters (i.e., probabilities) but on
simple spectral ranges in the spectral space (paral-
lelepiped algorithm), or on Euclidean distance to
training pixels. Artificial neural networks (ANNs)
represent one form of artificial intelligence tech-
niques that can be used for classification and repre-
sent a nonparametric algorithm, where the decision
boundaries between the classes are determined
iteratively by minimizing an error criterion on the
training data given (Jensen 2005).
4.3.2.4 Spatial and spectral segmentation
The aforementioned classification approaches
primarily rely on the concept of spectral separ-
ability in feature space. Image segmentation can
also be accomplished using spatial information such
as texture and spatial topology which is based upon
contextual relationships between pixels or objects
(spatial segmentation). Edge detection and various
spatial feature extraction algorithms can be used to
characterize spectral variability and detect class
boundaries. Areas of the same class membership
may be aggregated by region-growing algorithms
where individual pixels are joined according to alge-
braic rules (e.g., based on spectral characteristics).
Some spatial algorithms can be applied to panchro-
matic imagery. The spectral and spatial approaches
can be combined for spatial–spectral segmentation.
More advanced approaches to image classification
that include spectral and spatial analysis represent
object-oriented image classification (Jensen 2005,
Schowengerdt 2007).

4.3.2.5 Subpixel classification
Quantifying the landscape features that contribute
to the spectral signature of a mixed pixel leads to
subpixel classification. With subpixel assessment, a
distinction must be made between spectral subpixel
resolution (which categories contribute, and how
much?) and geometric subpixel resolution (where
are the categories?). In principle, the subpixel
locations of the contributing categories cannot be
resolved. Corresponding assumptions might, how-
ever, be drawn including the spatial context of a
pixel or external knowledge, for instance, about
the typical characteristics of a boundary. Some kind
of geometric subpixel precision may be achieved by
post-classification editing of class boundaries that
were originally determined with pixel precision.
Simple approaches of that type consist of the inter-
polation or smoothing of classification results. For
instance, pixelwise (i.e., horizontally stepped) clas-
sification between glacier ice and adjacent bedrock
might be better represented by interpolating a
smooth horizontal curve separating both classes.
In practice, however, such a procedure is often
complicated by a number of classification problems,
and might require knowledge-based interpolation
algorithms; success of the approach depends on
the ratio between GIFOV and spatial resolution
at which the category boundaries should be
mapped, among other factors.
Linear unmixing tries to determine the fraction of
idealized pure signatures (endmembers) contrib-
uting to its actual spectral composition. End-
member signatures can be derived from extreme
pixels assumed to consist only of one end member
class (i.e., pure signature) or inferred from other
data (Klein et al. 1999, Painter et al. 2003). Fuzzy
set classification follows the opposite approach to
the linear mixing concept by allowing one pixel to
be a member of multiple categories with a certain
probability connected to each membership (Binaghi
et al. 1997).
4.3.2.6 Combinations and others
The classification approaches covered here can be
applied using a variable number of spectral bands.
Some techniques, however, are especially suited for
hyperspectral data (e.g., linear unmixing), and a
number of additional algorithms that are based
upon spectral matching can be used (Lillesand
and Kieffer 2000, Schowengerdt 2007).
Classification approaches can be combined or
applied sequentially. Instead of using only spectral
Multispectral methods 81
data, most classification algorithms can make use of
spectral features (e.g., band ratios, principal com-
ponents, or multitemporal data) or nonspectral
data including DEMs and geomorphometric
parameters (Brown et al. 1998). There exists no
entirely accurate classification methodology to con-
sistently characterize glacierized terrains. What
works well in one region, or using a particular
dataset, may not have the same success in a different
part of the cryosphere or by applying a different
data source.
4.3.3 Image-processing techniques
Here, we summarize selected classification proce-
dures in the optical spectrum that have already been
tested for glaciological applications, in particular
mountain glaciers (Sidjak and Wheate 1999, Paul
2001, 2002, Albert 2002).
4.3.3.1 Manual delineation
Manual delineation of panchromatic or multi-
spectral image features might be useful for highly
complex classifications where expert knowledge is
needed, for instance, for separating rock glaciers,
dead ice, debris-covered ice, and periglacial debris
from each other. An analyst is able to include
experience, knowledge, and complex logical rules
in the decision process, also relying on nonspectral
information (i.e., multidimensional data or knowl-
edge). Manual delineation is often needed to com-
plement and correct automated classification results
(Rott and Markl 1989, Hall et al. 1992, Williams et
al. 1997, Paul 2002, Andreassen et al. 2008), but
especially for older panchromatic data like air
photos, Hexagon or Corona reconnaissance satel-
lite images (Bolch et al. 2010b, Narama et al. 2010)
and those data without SWIR bands (e.g., Landsat
MSS, IKONOS, QuickBird, ALOS AVNIR, etc.).
4.3.3.2 False-color composites
False-color composites (FCCs) take advantage of
the differences in reflectance of landscape features
(Fig. 4.4). For example, in a Landsat ETMþ 5,4,3
RGB composite (red: channel 5; green: channel 4;
blue: channel 3), snow and ice are clearly differen-
tiated from clouds, debris, rock, or vegetation due
to FCC image color differences. FCCs can be used
to facilitate manual delineation or for automatic
classifications. They work well for clean ice and
snow (Della Ventura et al. 1983, Williams et al.
1991). Other similar approaches could also include
82 Glacier mapping and monitoring using multispectral data

Figure 4.4. Landsat TM color composites over Tordrillo Mountains, Alaska. (Left) TM channels 3,2,1 red–green–
blue true-color composite; (middle) 4,3,2 false-color composite; (right) 5,4,3 false-color composite. Figure can
also be viewed as Online Supplement 4.2.

Figure 4.5. IHS transform of the Landsat TM 3,2,1 composite of Figure 4.4. (Left) Intensity (brightness); (middle)
hue (color); (right) saturation (amount of white). Figure can also be viewed as Online Supplement 4.3.

the use of intensity–hue–saturation (IHS) trans-
forms or decorrelation stretching.
4.3.3.3 Calculation of reflectance
Derivation of the surface reflectance for each pixel
requires three steps (Bishop et al. 2004):
1. Calculation of at-sensor radiance from image
DN values. The DNs have to be transformed
into radiance at the sensor using the gain and
offset values of the calibration coefficients
(Markham and Barker 1985).
2. Atmospheric correction of radiance due to atmo-
spheric attenuation and additive path radiance
(Vermote et al. 1997).
3. Topographic correction to account for varying
illuminations caused by different solar zenith
angles, localized slope and slope azimuth effects,
cast shadows, and hemispherical topographic
shielding if relief is significant (Hugli and Frei
1983, Sandmeier and Itten 1997).
Comparing remote-sensing derived reflectance with
that obtained from in situ measurements, or with
theoretical predictions from spectral signatures,
allows improved surface interpretation from the
corrected image. This approach has been used
mostly to characterize snow and ice facies, or their
albedo, respectively (Hall et al. 1988, 1990, Gratton
et al. 1990, 1993, Koelemeijer et al. 1993, Winther
1993, Knap et al. 1999, Paul 2001, Paul 2004, Casey
et al. 2012).

Figure 4.6. From left to right the first three principal component images (PC1, PC2, PC3) from the Landsat scene
displayed in Fig. 4.4. All VIS, NIR, and SWIR bands were included in the principal component analysis. Figure can
also be viewed as Online Supplement 4.4.
4.3.3.4 Spectral transforms
1. IHS transformation: For visual interpretation
and further classification or processing, a trans-
formation of RGB images (i.e., three-band
imagery) into another color space may be useful
(Fig. 4.5). The most often applied color space in
that context is the IHS color space. The IHS
components of an RGB image are, for instance,
obtained from geometric projections of a color
vector in the RGB space (Schowengerdt 2007).
The IHS components of an image might be
manipulated (e.g., replacement of intensity com-
ponent by a higher resolution I component, or
stretching) and then inversely transformed to the
RGB space. Furthermore, the components I, H,
or S might be used in a classification individually
(Paul 2004).
2. Principal component analysis (PCA): Multispec-
tral image bands are often highly correlated due
to similar matter/energy interactions in different
spectral bands and topographic effects (e.g., cast
shadows). Principal component analysis extracts
the unique linear variation components from the
entire multispectral dataset and generates princi-
pal component (PC) images that do not exhibit
multicollinearity (Richards 1993, Schowengerdt
2007). The technique reduces information redun-
dancy and PC images can be used to facilitate
image interoperation and image classification
(Sidjak and Wheate 1999). Direct assignment
of PCA results to classes of glaciological interest
is, however, difficult (Fig. 4.6) (Paul 2004). PC
images can be useful for ice displacement meas-
urements based on offset tracking in order to
optimally exploit all intensity variations con-
Multispectral methods 83
tained in a multispectral dataset (Ahn and
Howat 2011). A disadvantage of this method is
that the results are empirical in nature depending
on landscape features and complexity.
3. Decorrelation stretching: This technique repre-
sents application of PC or IHS transforms to
reduce redundant information in multispectral
imagery (Gillespie et al. 1986; Schowengerdt
2007). It allows better visual exploitation of mul-
tispectral imagery, assists in manual delineation,
and helps with the preparation and selection of
imagery for classification approaches (Fig. 4.7).
Decorrelation stretching uses PC images, which
are first stretched to optimally fill the color space,
and then transformed into the RGB color space
(Fig. 4.7). Decorrelation stretches are available
as ASTER Level 2 image products.
4.3.3.5 Image algebra and segmentation
Different algebraic algorithms may be applied to
spectral bands, among which band ratios R, for
example, represented as:
DNi
Rij ¼ ; or
DNj
DNi  DNminðiÞ
Rij ¼
DNj  DNminðjÞ
and normalized difference indices (NDI, or modula-
tion ratios), for example, represented as:
Rij  1 DNi  DNj
NDIij ¼ ¼
Rij þ 1 DNi þ DNj
are the most widespread, where DNi and DNj are
digital numbers of two bands i and j that show high
84 Glacier mapping and monitoring using multispectral data
Figure 4.7. Decorrelation stretch of the Landsat scene in Fig. 4.6, computed as a histogram stretch of the PC1,
PC2, PC3 RGB false-color composite.
separability for the category to be classified with
respect to other categories in the image, and
DNmin are the minimum (i.e., darkest) DN values
of an individual band.
For band ratios, subtraction of the minimum DN
for each band applied reduces illumination effects
from atmospheric scattering (Crippen 1988). Simi-
larly, the results of normalized differences might be
improved by subtraction of band-specific minimum
DNs.
For ice and snow, bands in the VNIR and the
SWIR are usually used in the index (e.g., TM2,
TM3, or TM4 versus TM5, or ASTER 3 versus
ASTER 4; Paul et al. 2002, Kääb et al. 2003a, Bolch
and Kamp 2006). The normalized difference snow
index (NDSI; Dozier 1989, Hall et al. 1995b, Sidjak
and Wheate 1999) is defined as
green  SWIR
NDSI ¼
green þ SWIR
Water can be detected by the normalized difference
water index (NDWI; Huggel et al. 2002), similar to
the well-established normalized difference vegeta-
tion index (NDVI):
NIR  blue
NDWI ¼
NIR þ blue
Classification of water or vegetation might be useful
for eliminating misclassifications from the above
band ratios for ice and snow (Paul et al. 2002, Bolch
and Kamp 2006). Band ratios and NDIs partly
eliminate atmospheric and topographic influences
that affect both bands when applied in a similar
way (Holben and Justice 1981). For final segmenta-
tion or classification of the ratio or normalized
difference image, thresholds have to be chosen.
Image algebra and segmentation algorithms are
especially robust (Paul 2001). To gain better control
of the results, further terms might be added, poss-
ibly including additional bands.

Currently, VIS/SWIR or NIR/SWIR band ratios
are among the most often and most operationally
used methods for mapping clean ice areas. A thresh-
old is typically added to a visible channel to
improve algorithm performance in cast shadow.
The VIS/SWIR ratio misclassifies dark shadow as
glacier ice when using images with uncorrected path
radiance. This can, however, be removed with a
threshold in the blue band (e.g., TM1). When little
glacier ice is present in a scene the NIR/SWIR ratio
is preferable (Paul and Kääb 2005). A typical, ratio-
based glacier mask might therefore be:
DNVNIR
> k and DNVIS > l
DNSWIR
where
. DNVNIR are digital numbers in the VIS or NIR
channels, usually red or near-infrared (e.g., TM3
or TM4, or AST2 or AST3; the choice of red or
near-infrared depends on scene conditions);
. DNSWIR are digital numbers in a SWIR channel
(e.g.,TM5 or AST4);
. k is the threshold applied to the ratio in order to
obtain a binary glacier mask; typical values for k
are of the order of 1.5–2.5, but may be even
smaller or bigger and depend much on the indi-
vidual scene and imaging conditions;
. l is the threshold on a shortwave visible channel
(DNVIS ; e.g., TM1), with values that have to be
adjusted to actual scene and illumination con-
ditions (Paul et al. 2009). Instead of DNVIS the
intensity value of an IHS transform can be used
(Paul 2004).
4.3.3.6 Unsupervised classification
Unsupervised classification approaches tend to be
successful for relatively homogeneous land cover
terrain that exhibits few landscape features,
whereas problems often occur for complex terrain
with numerous features (Paul 2001, Paul 2004).
Unsupervised cluster classes have to be assigned
to landscape features, and this can be problematic
given the relationship between the number of
classes and the selected feature space that may or
may not result in good statistical separability.
Nevertheless, the unsupervised ISODATA classifi-
cation was applied by Aniya et al. (1996) for glacier
inventorying of the Southern Patagonia Ice Field
using Landsat TM bands 1, 4, and 5, and was also
used to classify debris-covered ice on glaciers
Multispectral methods 85
around Mt. Rainier using ASTER VNIR imagery
(Section 17.5.2.2).
4.3.3.7 Supervised classification
Supervised classification algorithms (e.g., maximum
likelihood, spectral angle mapper) often produce
accurate results for high-mountain terrain. Never-
theless, debris-covered ice mapping remains chal-
lenging. Reasonable results can be generated
using VIS, SWIR, and TIR spectral bands, as well
as spectral features such as band ratios, principal
components, normalized differences, or TIR-
derived silica weight (Bronge and Bronge 1999,
Sidjak and Wheate 1999, Casey et al. 2012). The
choice of training sample locations is especially
crucial for identifying heterogeneous terrain com-
mon in alpine environments. Usually, class spectral
signatures obtained from one scene cannot be util-
ized on other scenes, although numerous studies
have been conducted (Gratton et al. 1990, Binaghi
et al. 1993, Paul, 2001, Paul 2004). Klein et al.
(1999) used spectral unmixing for classification of
snow and ice, and Binaghi et al. (1997) used fuzzy
set theory. Brown et al. (1998) applied maximum
likelihood and artificial neural network (ANN)
classifications on a DEM of a glacierized landscape.
4.3.3.8 Artificial neural networks
The decision boundaries in an ANN classification
are not fixed in a deterministic way from spectral
training signatures, but are determined in an itera-
tive fashion by minimizing an error criterion (Scho-
wengerdt 2007). The use of an ANN classification is
not restricted to spectral multitemporal data, as
topographic information can also be used. For gla-
cial and periglacial terrain, ANN classifications
from data of a single domain (spectral or DEM)
have been tested (Brown et al. 1998; Bishop et al.
1999; Paul. 2004). The main potential of ANN
application in glaciology might, however, lie in
the integration of multidimensional data (Paul.
2004).
4.3.3.9 Combinations
Often, classification procedures can (or must) be
combined either by fusing different approaches or
by performing them sequentially. As mentioned for
supervised classification, spectral derivatives or
results of other preprocessing classifications may
be combined as input layers. Sidjak and Wheate
(1999), for instance, achieved good results for gla-
86 Glacier mapping and monitoring using multispectral data
cier mapping from a supervised maximum likeli-
hood classification using PCs 2–4 of TM bands
1–7, a TM4/TM5 band ratio, and the NDSI as
input. The most often used chain of sequential clas-
sification approaches is to complete and correct
automatic classifications manually. For instance,
such a procedure is still unavoidable for mapping
debris-covered ice. In another example of combined
classifications, band ratios for glacier mapping may
result in misclassifications for vegetation in shadow
and turbid water, which in turn may be eliminated
by applying an NDVI and an NDWI.
4.3.4 Postprocessing and GIS work flow
Spatial domain filtering may be used after prepro-
cessing and automatic segmentations and classifica-
tions. For example, a median filter applied to ratio
images is recommended (Paul 2002). The filtered
segmentation results, still in raster format, can then
be converted to vector format. A number of man-
ual, semiautomatic, or automatic steps can follow
in order to obtain glacier outlines from a clean ice/
snow mask (Kääb et al. 2002, Paul 2002). For
example:
1. Separation of contiguous ice/snow areas into
glaciers. This step can be automated to a high
degree based on hydrological basin modeling
using a DEM (Schiefer et al. 2008, Bolch et
al. 2010a).
2. Correction of outlines for debris-covered ice,
lakes (see above), and other misclassifications.
3. Given repeat glacier inventorying, existing out-
lines can significantly facilitate Step 2 (Bolch et
al. 2010a).
4. Intersection of glacier outlines with a DEM in
order to derive topographic glacier parameters
might be possible (Paul et al. 2009).
5. The generation of central flow lines and intersec-
tion with glacier outlines in order to derive
glacier parameters related to the flow line (e.g.,
length) (Paul 2002).
4.4 MAPPING DEBRIS-COVERED ICE
Mapping of debris-covered glacier surfaces consti-
tutes one of the most difficult aspects of global-scale
glacier mapping. It is important to note that debris-
covered glacier surfaces cannot be accurately de-
lineated in most cases, and that debris-covered sur-
faces need to be treated separately from clean ice
surfaces in climate-related analyses due to the prop-
erties of debris cover which can enhance or reduce
ablation. Transition from a debris-covered glacier
surface to debris-covered stagnant ice or to ice-cored
moraines is ever changing and most likely represents
an indeterminate boundary. Possible definitions of
this transition include:
. Ice content: definition of the glacier boundary
under debris might be deduced from the percen-
tage ice content at depth and its characteristics.
The determination of intra and subdebris ice is
often the domain for in situ geophysical surveys.
Subsurface ice content is not strictly visible at the
surface, but might be detectable at the surface via
a thermal signal, distinct topography, or in tem-
poral imagery.
. Ice ablation rates: the amount of ice ablation
might define the glacier boundary (Whalley
1979, Haeberli and Epifani 1986, Kääb et al.
1997). This ablation rate is not necessarily related
to ice content (e.g., it is zero for an ice-cored
moraine under permafrost conditions where the
supraglacial debris thickness exceeds the local
active layer thickness; thermokarst features and
related topography are one expression of these
ablation rates).
. The kinematics of a debris-mantled ice body and
the ice body’s dynamic connection to the main
glacier. A deformation threshold would have to be
set, which depends also on the measurement pre-
cision of the method used.
These factors might either be directly observed, or
in some combination at the surface (e.g., through
topography, debris lithology, or ice velocity as
detected in repeat imagery).
Furthermore, debris-covered glacier surfaces
should be treated separately from clean ice glacier
surfaces due to very different atmospheric forcing
(Scherler et al. 2011). Debris cover less than a few
centimeters in thickness (Östrem 1959, Fujii 1977,
Nakawo and Young 1982, Mattson et al. 1993,
Rana et al. 1997, Adhikary et al. 2000, Kayastha
et al. 2000, Nicholson and Benn 2006) increases
ablation rates, whereas thicker debris cover reduces
ablation rates. Where, however, glacier surface
roughness increases and supraglacial ponds develop
as a consequence of the differential melt pattern and
thermokarst processes induced by debris cover,
area-averaged ablation rates might well be of the
order of clean ice or perhaps even higher (Sakai
 Mapping debris-covered ice 87

Figure 4.8. North Iceland. (Left) ALOS PALSAR phase coherence between the data of a 3-year interferometric
pair. Light blue indicates low coherence, pink to yellow high coherence. (Right) SPOT Quicklook. The sharp
transitions from low to high coherence coincide at many places with the outlines of glaciers, debris-covered glaciers,
and rock glaciers. Figure can also be viewed as Online Supplement 4.5.

et al. 2000, Bolch et al. 2011, Kääb et al. 2012). So
far, attempts to map debris-covered glacier surfaces
have been based on topographic, spectral, and
radar-interferometric data (Shukla et al. 2010).
Due to glacial transport, the source of supra-
glacial debris is different from the rock and sedi-
ments that are directly adjacent to a given position
on a glacier. In some cases, lithological variations
may therefore exhibit differences between glacially
transported or transformed debris and rock/debris
from adjacent valley flanks (Kääb 2005b, Casey et
al. 2012). In most cases, however, VNIR and
SWIR-based spectral classification alone will not
be sufficient to map debris-covered glacier surfaces.
The idea that the thermal signal of debris and debris
on ice could be different has led to attempts to
exploit TIR bands (see Section 4.5).
A number of studies investigated mapping
debris-covered ice using geomorphometric param-
eters and analysis (Bishop et al. 2001, Bolch and
Kamp 2006). Other works combined information
types (e.g., VNIR/SWIR or TIRþ geomorpho-
metric parameters) (Paul. 2004, Bolch et al. 2007,
Shukla et al. 2010, Bhambri et al. 2011).
Another approach to mapping supraglacial
debris is to compare elevation models of different
times and view elevation change over time as an
indicator for ice ablation or ice dynamic processes
(Kääb 2005b). As yet, this approach is often found
lacking of available DEMs with sufficient resolution
and/or accuracy.
Ice kinematics can be directly measured using
image-matching methods (see Section 4.5) or radar
interferometry, thus potentially mapping the land-
scape units belonging to a dynamically defined
glacier system (Kääb 2005b; Bolch et al. 2008). In
an extension of this idea, the loss of radar phase
coherence over time indicates tiny ground move-
ments and may surprisingly clearly depict the
dynamic boundaries of a glacier. This approach
seems especially successful for L-band radar data
with comparably robust phase coherence for the
terrain surrounding the glaciers (Fig. 4.8) (Atwood
et al. 2010, Strozzi et al. 2010).
The aforementioned approaches provide useful
results, but also highlight limitations related to
data, glacier characteristics, and universal applic-
ability. Unfortunately, manual digitizing of
debris-covered glacier surfaces is still the most often
used method for delineating debris-covered glaciers,
88 Glacier mapping and monitoring using multispectral data

Figure 4.9. TIR images over Tordrillo Mountains, Alaska, during the same day. (Left) Daytime TIR from Landsat
ETM; (right) night-time TIR from ASTER. During the day (left) longwave emission is modulated by topography due
to incoming shortwave radiation. This effect is largely reduced in the night-time image (right) in which the elevation
dependence of longwave emission is more pronounced, in addition to emission contrasts due to surface type. The
differences between both datasets are also related to surface thermal inertia.

Figure 4.10. ASTER RGB composite (bands 4, 9, 10, NIR, SWIR, TIR) over Hispar Glacier in the Karakoram
Himalaya. Lithological differences are visible using this band combination, and allow interpretation of the dynamic
history of the glacier and the sources of supraglacial debris. For instance, bends in distinct longitudinal debris bands
point to ongoing or past surges. Figure can also be viewed as Online Supplement 4.6.

although errors and uncertainties have been widely
recognized.
4.5 THERMAL IMAGING
The TIR spectrum has not been rigorously
exploited for investigating mountain glaciers (Fig.
4.9). Given incoming direct shortwave radiation,
debris generally exhibits stronger thermal emission
than ice due to its lower albedo, providing a rela-
tively strong signal in the TIR bands (Warren and
Brandt 2008). Researchers have investigated to
what extent TIR data can be used to detect glacier
ice under loose or thin debris cover from respective
cooling of superimposed debris (Taschner and
Ranzi 2002; Kääb 2005b; Mihalcea et al. 2008,
Shukla et al. 2010, Casey et al. 2012).
Fig. 4.10 shows an example of how the inclusion
of TIR supports mapping of the geological compo-
sition of the surface. Taschner and Ranzi (2002)
show that differences in TIR emission exist between
a debris-covered glacier and its surroundings, and
these differences can facilitate glacier outline de-
lineation. Shukla et al. (2010) and Bhambri et al.
(2011) combined TIR data with other spectral data
and geomorphometric information for mapping
debris-covered ice. In general, mapping debris-
covered ice from thermal data has the shortcoming
that recorded thermal emission on such a surface is
not strictly dependent on the ice underneath, but
rather on a number of factors contributing to the
energy balance such as roughness, incoming short-
wave radiation, thermal emissivity of the surface
layer, and meteorological conditions. For example,
Kääb (2005b) and Casey et al. (2012) explore litho-
logical variations of supraglacial debris as detected
by emission variation, and how these results can be
exploited glaciologically. Thermal-based silica map-
ping may be used to indicate weathering and active
turnover of glacier debris (Casey et al., 2012). One
disadvantage of TIR data is the coarse spatial reso-
lution compared with VNIR and SWIR bands (i.e.,
ASTER 90 m, Landsat ETM 60 m). Furthermore,
the actual resolvability of thermal features and
extraction of reliable surface temperatures is subject
to considerably worse resolution than the formal
TIR pixel resolution (see treatment by Ramachan-
dran et al. in Chapter 6).
During the daytime, the longwave radiation
emitted from the terrain surface is to a large extent
dependent on shortwave incoming radiation (Mit-
taz et al. 2000, Hoelzle et al. 2001). Therefore,
Microwave/SAR methods 89
nocturnal or early morning TIR imagery may reveal
thermal differences between surface materials (Fig.
4.9).
Thermal data have also been used to map the
thermal resistance of debris cover (Suzuki et al.
2007). In large-scale applications, low-resolution
TIR data with high-frequency revisit times (in par-
ticular MODIS) are operationally used for mapping
melt extent and its temporal variations over large
ice bodies (Hall et al. 2006).
4.6 MICROWAVE/SAR METHODS
Active microwave radar methods can be used for
glacier mapping as well. In the microwave region of
the spectrum, the matter–energy interaction is a
function of wavelength, polarization, incidence
angle, surface roughness, microwave penetration
(volume scattering), and the complex dielectric con-
stant of the surface (Lillesand and Kieffer 2000).
The latter describes the reflectivity and conductivity
of terrain material to incoming energy. Reflectivity
depends among other influences on surface rough-
ness in relation to the applied wavelength and vol-
ume scattering. While centimeter-scale roughness
appears rough for the K-band (  2 cm), it
appears smooth in the L-band (  20 cm).Material
conductivity is mainly dependent on liquid water
content. A cold and dry snowpack surface may
not interact with the energy, such that the dominant
reflection component is from snow at depth,
whereas penetration into wet firn or ice is very small
and surface reflectivity is high (Marshall et al. 1995,
Kelley et al. 1997, Engeset and Ødegård 1999;
Nagler and Rott 2000, Zahnen et al. 2003, Warren
and Brandt 2008).
Small mountain lakes with smooth surfaces cause
specular reflection and a dark signature in SAR
imagery (Pietroniro and Leconte 2000). Polariza-
tion of emitted and received radar signals (H: hori-
zontal; V: vertical) also permits separability of
features in imagery, because the different geometric
properties of scatterers on and within the snow, firn,
and icepack may lead to either preservation, 90
change, or diffusion (depolarization) of polarized
incoming radar waves with respect to backscattered
ones (Figs. 4.11, 4.12).
Radar wave penetration into the snow, firn, and
ice pack, though unwanted for a number of applica-
tions, may also be exploited for the detection and
mapping of crevasses covered by snow (Figs. 4.13,
4.14). Radar interferometry (interferometric syn-
90 Glacier mapping and monitoring using multispectral data

Figure 4.11. Color composite of HH (sent H, received H) and HV channels (sent H, received V) of an ALOS
PALSAR winter scene over Kronebreen, Ny Ålesund, Svalbard (HH: yellow; HV: blue). Color differences in the
composite occur due to changes in the degree to which surface or near-surface backscattering mechanisms change
the polarization of the incoming radar signal. Dominant co-polarization (here: HH) points to corner reflection and
ice lenses or layers, while dominant cross-polarization (here: HV) points to volume scattering (e.g., in homogeneous
ice) (north to the top).

Figure 4.12. Color composite of a fully polarized ALOS PALSAR winter scene over Ny Ålesund, Svalbard
(Kronebreen to the lower right) (north approximately to the right).
 Spectral change detection and temporal data merging 91

Figure 4.13. Envisat ASAR backscatter over Ny Ålesund/Kronebreen, Svalbard. (Upper right) June 13, 2008;
(middle left) February 29, 2008; (middle right) September 26, 2008. February data were taken after a cold and
snowy period (lower panel), the June data after a period of slightly positive temperature with little precipitation, and
September data after a period of strong rainfall. Whereas glacier delineation is difficult on single dates due to
different ground conditions, glacier edges become better detectable in the multitemporal false-color composite of all
the scenes (upper left). (Lower panel) Blue bars: precipitation; red line: air temperature in Ny Ålesund (data from
met.no) (north to the top).

4.7 SPECTRAL CHANGE DETECTION

thetic aperture radar, or InSAR) for extraction of
DEMs is not covered in this chapter, but differential
radar interferometry (DInSAR) for measurement
of glacier flow is briefly discussed in Section 4.8.
Optical and microwave data for glacier remote
sensing are complementary, as information from
different regions of the spectrum may produce
new insights and understandings. These topics are
treated in Section 4.9.
AND TEMPORAL DATA MERGING
4.7.1 Overview
We term merging of either spectral or spatial
domain data of different acquisition times as multi-
temporal data merging. Two important multi-
temporal merging methods are: measurement of
elevation changes (Chapter 5) and terrain displace-
92 Glacier mapping and monitoring using multispectral data

Figure 4.14. ALOS PALSAR winter backscatter data over Kronebreen, Ny Ålesund, Svalbard. (Upper image)
PALSAR scene, cf. Fig. 4.11. (Lower left) Section of the same scene; (lower right) summer ASTER VNIR data over
the same section as the lower left panel. Even through snow cover, crevasse zones (visible in summer ASTER data)
become clearly detectable as high backscatter in radar images

ments (Section 4.8). Merging of these two categories
of multitemporal geometry data facilitates under-
standing and modeling of glacier dynamics. Vertical
changes and horizontal displacements are quantita-
tively combined with the kinematic boundary
condition at the surface (Kääb et al. 1998, Gud-
mundsson and Bauder 1999, Kääb and Funk
1999, Kääb 2001). Basic multitemporal merging
in the spectral domain consists in the overlay of
repeated image data, most often applied for digital
change detection (Singh 1989, Mouat et al. 1993,
Lillesand and Kieffer 2000). Change detection tech-
niques include
. postclassification comparison
. multitemporal classification
. multitemporal principal component analysis
. multitemporal false-color composites
. algebraic expressions
. change vector analysis
. change axis analysis.
Although mainly designed for application in the
spectral domain, some of the above strategies can
also be applied for other data types.
Common spectral (and also nonspectral) classifi-
cation procedures may be applied separately on
datasets of different times, and the classification
results compared thereafter (postclassification com-
parison). For instance, glacier-covered areas can be
detected by multispectral classification from a satel-
lite image of time 1, and again from a satellite image
of time 2. By simple algebraic expressions the areas
of glacier change can be extracted and quantified.
Instead of comparing results post classification,
terrain changes may also be detected by merging
multitemporal data within one classification proce-
dure by defining change classes and nonchange
classes (multitemporal classification).
The individual bands (or layers) of multitemporal
data may be combined into one new multilayer
dataset (Fujimura and Kiyasu 1999). Principal
component analysis of this new dataset may then
enable detection of terrain changes, as the principal
components are not correlated. Such an approach is
 Spectral change detection and temporal data merging 93

Figure 4.15. Deposits of the September 20, 2002 rock/ice avalanche at Karmadon in the North Ossetian Cau-
casus. Change detection is done by means of a multitemporal RGB composite (left panel). Red, ASTER band 3 of
July 22, 2001 (upper right panel); green and blue, ASTER band 3 of October 13, 2002 (lower right panel).
Avalanche track and deposits, as well as lakes dammed by deposits become detectable. Red-colored changes
on northern slopes are due to different shadow/illumination conditions between the acquisition dates. Figure can
also be viewed as Online Supplement 4.7.

Figure 4.16. Normalized difference index image (right panel) of two ASTER images over a glacier in Bhutan
(January 20 and November 20, 2001; left panel). Largest differences (black and white) occur at crevasses due to
their movement from left to right (black-and-white pattern) and at the retreating calving front of the glacier.
Differences in lateral moraines point to errors in orthorectification, presumably from DEM errors (SRTM) propagat-
ing into horizontal orthoprojection errors. Figure can also be viewed as Online Supplement 4.8.

of special use when change has to be extracted from
large datasets (Marshall et al. 1995).
Multitemporal false-color composites (FCCs)
represent a powerful tool for visualizing change
between two or three images (Figs. 4.13, 4.15). Such
FCCs might directly form the basis for mapping, or
serve as preparation and evaluation of other change
detection algorithms. The number of data acquisi-
tion dates included in a multitemporal FCC is
restricted to three unless PCA is used beforehand
to extract the most dominant changes from the
dataset. Multitemporal FCCs may also be used to
visualize and investigate the results of other change
detection techniques (see Section 4.7.2).
94 Glacier mapping and monitoring using multispectral data

Figure 4.17. (Left) Section of an RGB normalized difference index (NDI) of two Landsat scenes (September 22,
1992 and October 31, 2009). R: NDI of bands 5; G: NDI of bands 4; B: NDI of bands 2. (Right) Section of the 2009
scene.The different dates of the year (>1 month difference) lead to illumination differences that become visible as
apparent change (cf. Section 4.7.2). Glacier retreat shows up in red. The white ring around the lake in the middle
(Lake Sabai) reflects the reduction in lake area as a result of lake outburst on September 3, 1998. Figure can also be
viewed as Online Supplement 4.9.

Algebraic expressions such as subtraction, ratios,
or normalized difference indices between two or
more multitemporal datasets or derivatives of them
are often used for change detection (Figs. 4.16,
4.17). Detecting elevation differences from repeated
DEMs is a simple example of temporal data sub-
traction. Spectral differences between repeated
imagery indicate terrain changes, but are also
affected by different illumination and atmospheric
conditions. Under some circumstances, the ratios of
multitemporal imagery tend to normalize the effects
of illumination variations (Crippen 1988). Fig. 4.16
shows an example of detecting glacier movement
from repeated ASTER imagery from a multitem-
poral band ratio. The displacement of individual
terrain features such as crevasses (spatial changes)
was used to detect change over time. Spectral sig-
natures can also be used for change detection, as an
area becomes ice free due to glacier retreat.
The above change detection approaches may be
applied to raw image data, radiometrically cor-
rected data, or spectral–spatial transformed or pro-
cessed products (e.g., orthoimagery, filter products,
PCs). Application of change detection algorithms
can also be restricted to certain areas masked by
some previous classification (Lillesand and Kieffer,
2000).
For a given pixel of a multispectral image or
multilayer dataset, two or more (spectral) variables
can be plotted against each other for individual
acquisition dates. A change vector connects result-
ing points in the chosen variable feature space. The
magnitude and direction of this vector (or vector
cluster for a group of pixels) may be characteristic
of a certain type of change (e.g., plant succession in
glacier forefields, development of snowpack, or
glacier retreat (Lillesand and Kieffer 2000).
A two (or more) dimensional scatter plot of a
spectral band at time 1 versus the same band at
time 2 approximates the plot diagonal (with slope
of 1) in the case of no changes between the two
acquisition dates. For significantly changed pixels
the respective plot points lie apart from this diag-
onal, forming clusters representing discrete change
types. Change axis analysis defines a nonchange
axis (an above diagonal or a parallel or slightly
rotated axis) and perpendicular change axes.
Threshold coordinates in the change axis space
are then used to classify individual changed, or non-
changed, pixels (Lillesand and Kieffer 2000). The
method is also sensitive to changes occurring within
shadow.
For all change detection approaches, accurate
uniform geometry for multitemporal datasets is
crucial. Such techniques are, thus, usually applied
using imagery of the same sensor and same sensor

position (e.g., repeat tracks) or orthoimages. Loca-
tion differences between compared datasets might
influence change detection procedures heavily (Fig.
4.16). Similarly, illumination differences due to
variable solar geometry may create significant
apparent changes (Figs. 4.15, 4.17).
4.7.2 Image change evaluation by
subtraction of multispectral
anniversary pairs (ICESMAP)
Many chapters in this book show the results of
change detection and land cover characterization
using multispectral images. Image differencing is a
general processing technique developed to identify
uncorrelated zones within multitemporal image sets
(Singh 1989, Jensen 2005), where surface changes
have occurred during the time between image pair
acquisitions. This per-pixel image algebraic method
simply involves subtraction of one digital image
from another to produce a third difference (or
change) image. This can be expressed as
DXk;i;j ¼ DXk;i;j ðt2 Þ  DXk;i;j ðt1 Þ þ C
where DX is a pixel value for sensor band k, i and j
are image line and column coordinates respectively,
t2 refers to the later image, t1 refers to the earlier
image, and C is a constant for rescaling to unsigned
8-bit image values (although radiometric bit depths
may vary dependent upon sensor characteristics).
For multispectral imagery, the image spectral
bands, k, will vary by choice of the user, but for
our work with ASTER imagery k includes three
VNIR bands—band 3 (NIR), band 2 (red), band
1 (green)—composited into RGB color space. Ade-
quate change detection by image differencing can be
extremely challenging to do quantitatively if images
are acquired under different illumination or sensor
view conditions. Widely disparate shadowing and
complex scattering phase functions can be the most
difficult or impossible challenges to remedy (Fig.
4.17).
Here we provide background to one method,
Image Change Evaluation by Subtraction of Multi-
spectral Anniversary Pairs (ICESMAP), which was
developed for GLIMS by the University of Arizona
group (J.S. Kargel and G. Leonard) in order to
circumvent the need for complex radiometric and
geometric transformations of images over rugged
surfaces. For this book, ICESMAP has been
applied to glaciers in the Chugach Mountains,
Alaska (Chapter 13); Hoodoo Mountain, British
Columbia (Chapter 15); Mount Rainier, Cascades
Spectral change detection and temporal data merging 95
(Chapter 17); the Mount Everest, Himalaya area
(Chapter 24); and the Mount Cook area, New Zeal-
and (Chapter 29). This section is the first detailed
presentation of the methodology. However, the first
published application of the ICESMAP method-
ology was for an image pair acquired over Central
East Greenland (Kargel et al. 2012).
Alpine glaciers and adjacent terrains typically
undergo satellite-discernible surface changes due
to weather events, discrete geomorphological
events, and seasonal and interannual fluctuations.
Some of these changes integrate the effects of
climate change, ice motion, phase state, and mass
movements on glacier surfaces. Notable changes
seen in differenced anniversary image pairs may
reveal changes in glacier extent and texture/micro-
relief; surface patterns and area density of crevasses,
debris, meltwater ponds, and streams; snow distri-
bution, and grain size of snow, firn, and ice; melting
or freezing of wet snow; snow and rock avalanches;
and changes in vegetation types and density on
debris-covered glaciers, moraines, and beds of
drained glacier lakes. All of these material or
ð4:2Þ
topographic changes may potentially be resolved
spatially, spectrally, and/or radiometrically, which
is the core premise of image change detection: that
surface variations result in changes in radiance
intensities which are larger in magnitude with
respect to radiance changes resultant from other
sources unrelated to changes in the surface (Ingram
et al., 1981).
Other sources in this case refer to relative changes
in image pair radiometry due primarily to changes
with little relation to surface changes. These may
include variable view geometry and the scattering
phase function of the surface; different sensor gains,
shifting solar position in the sky related to time of
year and day, and differing atmospheric attenuation
conditions related to humidity and temperature,
optically thin condensation clouds, and dust (Jen-
sen, 1983). The minimization or removal of these
extraneous effects is required to identify and
quantify actual surface changes; the approach of
ICESMAP is simply to minimize these effects rather
than correct for them, noting that in practice it is
very difficult and often impossible to make an accu-
rate correction for all these variables. The problem
with corrective approaches is that usually not
enough is known, or uncertainties are too great,
to allow quantitative correction and isolation of
the effects of actual changes. The ICESMAP solu-
tion is simple control of nonsurface changes so that
there is little or nothing to correct. If the Sun is in
96 Glacier mapping and monitoring using multispectral data
the same position in the sky and the sensor view of
the surface remains the same, then there is no need
to correct differences in solar angle and differential
effects of shadowing and hill shading, differential
pathlength of photons through the atmosphere, dif-
ferential sensor gains, scattering phase functions,
and so on. Equalization of these factors is prefer-
able to having to correct for them. The limiting
factor, of course, is that images must be acquired
on or near anniversary dates (or, as we explain
below, otherwise have the Sun in the same place
in the sky) with the same sensor and with the same
view of the surface. The key is to select suitable
image pairs that commonly (but not always) exist
for a given area.
In some cases, single-band image differencing
may be sufficient to indicate where surface changes
have occurred (e.g., using either visible or NIR
bands to identify surface brightness and vegetation
changes, respectively). In this method, threshold
DN values within the difference image histogram
are selected by the user to highlight areas of sig-
nificant change (e.g., statistically selected ‘‘area
change’’ values, perhaps  2 or  2). Note that
areas of change in this example are represented
within the tails of the histogram; whereas values
clustering on or near the mean of the difference
image histogram represent areas of little to no
change. Whilst slightly more complex, the differ-
encing of multispectral image pairs can lead to a
richer portrayal of user-identifiable change features.
This is true since multispectral data can better
capture the unique compositional information con-
tained within the variable surface spectral responses
detected by the different channels of the sensor. The
resultant change image may subsequently contain
information related to changing surface material
compositions and their distributions. The resulting
difference image then requires postdifferencing clas-
sification or coding of surface changes by an expert
familiar with the terrain or target area. Compared
with single-band differencing analysis, it may be
more complicated to segment specific features of
change in multiband differencing, as it may become
necessary to identify what combination of image
bands are responsible for indicating a particular
surface change before additional thresholds or band
ratios are applied. Ideally, interpretations can be
rather intuitive, although if one image or both
images contain saturated bands (where the upper
range of recorded radiance is limited by gain sensi-
tivity), then the colors represented in the change
image may appear as strongly contrasting hues
but in fact can become difficult or impossible to
interpret. This potential problem underscores the
importance of applying images that have been
acquired with optimal GLIMS gain settings (i.e.,
gain settings set to maximize optimal image quality
for glaciers and features of interest for the particu-
lar area of study), or ignoring or carefully account-
ing for areas of the difference image where image
saturation is problematic.
The principal factors affecting image quality
needed for image differencing include atmospheric
conditions, cloud cover, sensor gain settings, and
Sun–target–sensor viewing geometry. For some
purposes, such as multitemporal mapping of the
calving front extent of a glacier, there is little pre-
processing to be done to an image series except to
assure proper image co-registration. In other cases,
such as the mapping of more subtle land cover units
or seeking to identify a wide range of changing land
cover/land use phenomena between images, it
becomes necessary to correct for various param-
eters of the changing observing geometry within
the image pair including sensor–solar azimuth angle
and solar zenith angle, scattering phase functions,
atmospheric absorptions, differing sensor gain set-
tings, and interinstrument calibrations. This can be
done, but is sometimes a complex process that may
ultimately prove futile if the images are acquired
under conditions that are too disparate from one
another. Subsequently, isolation of seasonal versus
interannual, versus long-term trends can be diffi-
cult. Therefore, for the purposes of assessment of
interannual and long-term glaciological trends,
change assessment is ideally done with images
acquired by the same imaging system in Sun-
synchronous orbits, using the same instrument
settings (e.g., gain) and viewing geometry, and
acquired on (or near) anniversary dates, and under
the same atmospheric conditions. For these con-
ditions, the solar–sensor azimuth and solar zenith
angle parameters are nearly identical; consequently,
the observing geometry and scene illumination are
the same. If atmospheric conditions also are similar
and ground features are unchanged, then the same
spectral reflectance image will be recorded, because
illumination, viewing, absorption, and scattering
functions are the same. Furthermore, if instrument
drift is known and corrected, then subtraction of
one image from the other will result in a black field
(zero difference). Normally, of course, this is not the
case as surfaces are changing, and we are interested
in these actual surface changes. In short, near-
anniversary date image pairs better insure that
 Spectral change detection and temporal data merging
spurious signals and artifacts, potentially resultant
from differing diurnal and seasonal Sun angle
effects, and seasonal vegetation phenological
differences are minimized.
For the current generation of orbiting Earth-
observing sensors, exact anniversary dates are
rarely acquired, but images acquired within a few
days of first image acquisition anniversary are fairly
common, particularly for more frequently imaged
areas of Earth. Importantly, image pairs should be
selected for the season and or annual date(s) that
may optimally capture the changing phenomena of
interest to the user. For satellite glacier studies, this
may include images acquired at the end of the re-
gional ablation season. Additionally, seasonality is
important for image date selections. Winter images
may be relatively darker, contain significant
shadowing, and be hindered by increased snow
cover. The former two problems may in part be
mitigated by acquiring image data from Sun-
synchronous sensors that collect scene data at local
midday times. In addition, there is wider tolerance
for nonexact anniversary date image pairs collected
near the solstices compared with the equinoxes; this
is because the Sun’s observed path through the sky
is at a minimum during the solstices, and maximum
during the equinoxes. As a result, an image pair
differing by one week collected near the solstice will
show far less shadowing differences than a similar
one-week pair acquired near an equinox. Another
approach may include selection of images acquired
over an equivalent number of days before and after
a solstice, so that the Sun again is at the same place
in the sky (though seasonal changes may become
significant). Shadows and surface scattering are not
identical but are similar, and with the same look
direction the total solar radiative pathlength is also
similar.
The success of image change assessment by this
method is dependent upon accurate image pair co-
registration. Serious misregistration can produce
ghosting and false indications of change, and even
misregistration by a few tenths of a pixel can pro-
duce false outlining of features, typically revealed
where material contacts are highlighted by a pair of
lines of differing colors. If registration is very pre-
cise, generally unchanging features (over time spans
of years to decades) such as hills, valleys, mountains
peaks, and other relief features, and various
material units that are visually prominent in source
images, virtually disappear as muted grayscale feat-
ures in the change image. As a quality minimum,
image co-registration should produce RMSEs of
97
less than 0.5 pixel, a precision possible with a
variety of commercial image-processing software
programs, or otherwise with user-generated
image-processing code. Georeferencing may also
be required if one or both images are not yet geo-
referenced and if it is important to assign geospatial
values to input and output products.
The following steps summarize a generalized
approach to applying ICESMAP. Different users
may find that they can modify or eliminate some
steps, depending on the image data type and quality
used, the software or code applied, and the desired
final products.
4.7.2.1 Generalized ICESMAP processing steps
Selection of VNIR multispectral image pair/series
should be carried out in the following way. Ideal
images are cloud and haze free, apply identical
sensor gains, and are acquired on the anniversary
date. However, images a few days off the anniver-
sary date usually provide useful results. Appendix
24.5 provides an error analysis for one example of
image differencing applied to an ASTER image pair
acquired near Mount Everest. ASTER Level 1B
image data are well suited for ICESMAP analysis,
with DN values in terms of scaled (8-bit) radiance.
The steps are:
1. Optional: apply atmospheric correction or dark
object subtraction from all bands.
2. Produce three-band image cubes (e.g., AST3–
AST2–AST1 ! RGB).
3. Subset images if necessary (i.e., reduce to area of
interest).
4. Co-register image pair/series (<0.5 pixel
RMSExy ).
5. Subtract earlier image from the later image.
6. Rescale to unsigned image format (full dynamic
range of original imagery, e.g., 8-bit, 16-bit).
Note: this step may occur automatically depend-
ing on the software applied or design of the
algorithm.
7. Render difference image into RGB space.
8. Post differencing, the user assigns change feat-
ures and defines what any anomalous change
features represent.
9. Optional: apply a threshold to individual or
multiple difference image band(s) in order to
segment and quantify desired change features
or apply a higher level customized classifier or
a clustering algorithm such as Fuzzy C Means.
98 Glacier mapping and monitoring using multispectral data
The change image from a pair of differenced three-
band images is rendered as an RGB image although
the DN values are rescaled so that the lowest nega-
tive value is set to zero, and the maximum change
value is rescaled to DN ¼ 255 (for 8-bit images). In
this case, unchanging features become a neutral
gray, and changed surfaces become brighter or
darker or colored. However, there are usually por-
tions of the difference image that show residual
features that cannot be completely removed, such
as small variations on the surface where shadows
have shifted slightly, where differential slopes have
slight illumination differences (especially when slop-
ing near grazing solar incidence). Changes in vege-
tation or soil or snow moisture can be of interest,
but differences in leaf moisture or slight differences
in soil moisture and other fairly small changes can
have large effects on a difference image. Typically
though, the change features above are rather minor,
against which more dramatic surface changes will
stand out boldly, particularly if the chosen image
pair is of high quality. Real changes will be high-
lighted in image brightness or color against a back-
ground of nonchanging surfaces that appear
relatively subdued in the change image.
Examples of difference image results and their
interpretations are provided in the chapters men-
tioned above. Image pairs used in this book’s
chapters on the Himalaya (Section 24.3.3) and the
Hoodoo Mountain area (Section 15.3.1) include
explanations of the methodology for those specific
image pairs, the interpretation of results, and the
sources of error and artifacts. Other chapters pro-
vide briefer explanations. In general, and not con-
sidering areas where an image or image pair is
saturated in one or more bands, areas where the
differenced image is blue represent a surface that
has become bluer (or less red), such as where a pond
has formed; where red, they have become more red
(or less blue), such as where a pond has drained;
and where green, band 2 has become brighter rela-
tive to bands 1 and 3. Image areas that appear black
represent areas that have become darker in all
bands; and white areas have become brighter in
all bands. One of the chief caveats to keep in mind
is that when one or both of the images are saturated
in one or more bands, the colors represented in the
change image are more difficult to interpret because
some information is lacking in those pixels. This
most commonly occurs within areas that are highly
reflective across all VNIR bands, such as within or
proximal to snowfields in accumulation zones.
Although such areas may appear in a multitude
of different hues, it is often still possible to identify
the transient snow line (e.g., ELA) especially if only
one of the images in the image pair is less saturated
or unsaturated. Table 4.1 shows 13 typical cases of
changing surface features within glaciated terrains,
from the perspective of how these features would
appear in the before image and subsequently in the
difference image.
4.8 ICE FLOW
The kinematic boundary condition at the surface
(Kääb and Funk 1999) or the mass conservation
relation (Cogley et al. 2011) express that a change
in terrain elevation over time at a certain location
and the vertical component of a three-dimensional
displacement of an individual particle at the surface
describe different kinematic quantities. A change in
elevation at a defined glacier location may be the
effect of mass balance, three-dimensional straining,
and/or mass advection, the latter of which is again a
function of glacier displacement and geometry.
Glacier displacements can be determined by dif-
ferent methods, though most of them only deliver
single components of the three-dimensional velocity
vector. Methods include:
. vertical differences between multitemporal DEMs
(see Chapter 5);
. qualitative analysis of movements, such as from
terrestrial indicators (e.g., lack of vegetation,
degree of lichen coverage and lichen size, surface
weathering), change detection techniques (see
above), or repeat image animation (flickering)
(entire topic not covered here) (Kääb et al.
2003b, Paul et al. 2007);
. digital matching of repeated optical imagery (see
below);
. differential radar interferometry (DInSAR) (see
below);
. repeated terrestrial and satellite geodesy (i.e., re-
peat survey of natural or artificial markers—not
covered here);
. analog and analytical photogrammetric methods
(not covered here);
. DEM matching (i.e., digital matching of repeat
DEM grids, similar to image matching—not ex-
plicitly covered here, but implicitly under image
matching);
. terrestrial real or synthetic aperture radar (not
covered here); and
Table 4.1. Examples of ICESMAP image difference characteristics.

Case Feature (change) Image ASTER (ordinal DN values) b Predifference image Difference image appearance
time a appearance (t2 image–t1 image)
Band 1 Band 2 Band 3

1 Glacier terminus clean ice t1 High Mod.–high Mod.–high White Dark gray–black
Glacier terminus (receded 2 pixels) t2 Low Low Low Gray–brown

2 Supraglacial pond absent t1 Low Mod.–high Mod. Gray–brown Blue

Supraglacial pond (appearance) t2 High Low Low Blue

3 Supraglacial pond t1 High Low Low Blue Red

Supraglacial pond (disappearance) t2 Low Mod.–high Mod Gray–brown

4 Glacier ice surface t1 Low–high Low–mod. Mod. White–light gray Mottled, mod.–dark gray
Glacier ice surface (flow 2 pixels) t2 Low–high Low–mod. Mod. White–light gray

5 Transient snowline t1 High Mod.–high Mod.–high White Gray–dark gray

Transient snowline (elevation change) t2 Mod. Mod. Mod. Light gray–white

6 Proglacial lake nonturbid t1 Low Low Low Dark blue Light blue–green
Proglacial lake (incrased turbidity) t2 Mod. Low–mod. Low Blue–green

7 Firn fine grained t1 High Mod.–high Mod.–high Light gray–light blue Mod.–dark blue
Firn (coarse grained) t2 High Mod.–high Mod. Mod. gray–mod. blue

8 Firn coarse grained t1 High Mod.–high Mod. Mod. gray–mod. blue Light yellow–orange/light red
Firn (fine grained) t2 High Mod.–high Mod.–high Light gray–light blue

9 Snow patch t1 High High High White Dark gray–black

Snow patch (disappearance) t2 Low Low–mod. Low Mod.–dark gray

10 Rock outcrop t1 Low Low–mod. Low Mod.–dark gray White

Rock outcrop (snow covered) t2 High Mod.–high Mod.–high White

11 Vegetation low LAI t1 Low Low Mod.–high Mod. red Mod. orange–dark red
Vegetation (high LAI) t2 Low Low High Bright red

12 Vegetation high LAI t1 Low Low High Bright red Light–mod. blue/green
Vegetation (low LAI) t2 Low Low Mod.–high Mod. red
Ice flow

13 Snow, ice, rock, vegetation, hills, etc. t1 Variable Variable Variable Variable Neutral gray, untextured
Same (unchanged) t2 Variable Variable Variable Variable
99

a
t1 ¼ before image; t2 ¼ after image. b For example, relative to 8-bit dynamic range values where 0 is lowest and 255 highest values, respectively. DN ¼ Digital Number; LAI ¼ Leaf Area
Index; mod. ¼ moderate.
100 Glacier mapping and monitoring using multispectral data
. mechanical methods such as strain wires (not
covered here).
The ICESMAP method (Section 4.7.2) maps areas
of significant ice displacement, but displacement
vectors are not determined.
4.8.1 Image choice and preprocessing for
image matching
A highly efficient method for measuring terrain dis-
placements is the comparison of repeated optical
imagery. If original imagery is used, the obtained
displacements have to be rectified using a corre-
sponding sensor model and orientation parameters
(Kääb et al. 1997, Kääb and Funk 1999, Kaufmann
and Ladstädter 2002). If orthoimages are used,
image comparison directly delivers the horizontal
components of the displacement vector. The
approach is, in principle, applicable to terrestrial,
aerial, and spaceborne imagery, and DEMs.
The choice of images to be used for matching has
to fulfill, or balance, the following conditions:
. the time difference between the images used has to
be large enough for displacements to exceed the
accuracy of the matches. Matching accuracy is
mainly, but not only (see Section 4.8.4), a function
of the image resolution and precision of the
matching algorithm.
. The time difference between the images has to be
small enough for surface transformations (melt,
shearing, etc.) to not inhibit the discovery of
corresponding features in multitemporal images.
Corresponding features may be destroyed in as
little as a few weeks for fast-flowing glaciers under
temperate climate conditions, or may last several
years for cold regions and/or slowly deforming
debris cover.
. The larger the time difference between the images
the better the signal-to-noise ratio in estimated ice
velocities.
. Images have to be co-registered at an accuracy
higher than the targeted measurement accuracy or
displacements. This co-registration can be part of
geometric preprocessing (e.g., orthorectification
with common tie or ground control points) or a
separate preprocessing step (i.e., co-registration
using stable ground points). Alternatively, stable
ground offsets can be measured as part of the
image-matching process, correction parameters
estimated from these, and displacements on
glaciers corrected accordingly. Note also remarks
on orthoprojection in Section 4.2.2, particularly
the error propagation resulting from the use of
DEMs to generate orthoimages.
Besides the geometric preprocessing of images to be
matched (mainly co-registration), radiometric pre-
processing might be useful; normalization between
the gray values of the image will generally improve
the results. Additionally, instead of a raster-based
image-matching process, points of interest with suf-
ficient gray value variations around them can be
detected and matchings performed at these loca-
tions only. Improved image-matching results are
reported if derivatives of the original images are
used, instead of the originals themselves (e.g., gradi-
ent images, also called orientation images, or prin-
cipal component; Frezzotti et al. 1998, Haug et al.
2010, Ahn and Howat 2011, Heid and Kääb 2012).
4.8.2 Image-matching techniques
Digital comparison between multitemporal
images—termed image matching, or intensity
matching/tracking, or texture matching/tracking—
may be accomplished by area-based matching tech-
niques (also called block matching), or feature-
based matching techniques. Block matching com-
pares complete gray value arrays (i.e., image sec-
tions) with each other, whereas feature-based
matching compares geometric forms such as edges
or polygons that are extracted from the imagery
and converted to vector features. For measuring
glacier movement, area-based methods seem prefer-
able due the fact that the reflectance variations to be
tracked are often smooth, at least for medium and
low-resolution data, and would thus not allow
extraction of distinct vector features. Algorithms
for area-based matching operate in the spatial
domain or in the frequency domain, or a combina-
tion of both.
Of the spatial domain techniques the most suit-
able for glacier flow are double cross-correlation
and least-squares matching (Gruen and Baltsavias
1987, Scambos et al. 1992, Kääb and Vollmer 2000,
Kaufmann and Ladstädter 2002, Kääb 2002,
Debella-Gilo and Kääb 2011), and of the frequency
domain techniques Fourier- and wavelet-based
algorithms are the most appropriate (Leprince et
al. 2007, Haug et al. 2010). All these area-based
methods rely on extracting an image section (refer-
ence template) from the reference image (e.g., time
1) and searching within a search window the most
similar search template in the search image (time 2).
 Ice flow 101

Figure 4.18. Surface velocity field for a section of Kronebreen, derived from ASTER imagery of June 26 and
August 6, 2001. Isolines indicate ice speed in meters per year. The surface velocities of Kongsvegen are too small to
be measured from repeated satellite imagery of 15 m resolution over a few weeks (underlying ASTER image of
August 6, 2001). Figure can also be viewed as Online Supplement 4.10.

For frequency domain methods, the search window
is equivalent to the search template.
A detailed list, explanation, and comparison of
the most commonly used matching methods in
glaciology is given by Heid and Kääb (2012).
The size of the search window has to be chosen
according to the expected maximum displacement,
such that the search template that corresponds with
the reference template can, in fact, be found in the
search window. The size of the reference and search
templates have to be chosen according to the tex-
tural characteristics of the ground surface. If the
reference template size is too small, the similarity
surface has no clear maximum; if the reference tem-
plate size is too large, computing time soars up, and
deformations within the template area degrade
matching accuracy or in extreme cases cause mis-
matches.
Some matching algorithms inherently output
subpixel precision such as least-squares or Fourier
transform–based ones. For other algorithms such
as cross-correlation, subpixel accuracy can be
achieved by either interpolating the images or
templates to higher resolution (oversampling), or
by peak fitting to the similarity surface at its max-
imum (Debella-Gilo and Kääb 2011). In cases
where the observed displacement greatly exceeds
the spatial resolution, subpixel precision might
not be necessary.
Digital motion measurements from repeated
optical imagery have been applied
. for ice sheets using satellite imagery (Scambos et
al. 1992, Whillans and Tseng 1995, Frezzotti et al.
1998, Haug et al. 2010, Ahn and Howat 2011),
. for arctic glaciers using satellite imagery (Lefau-
connier et al. 1994, Rolstad et al. 1997, Dowdes-
well and Benham 2003, Kääb et al. 2006; Fig.
4.18),
. for mountain glaciers using satellite, aerial. or
terrestrial imagery (Seko et al. 1998, Nakawo et
al. 1999, Evans 2000, Kääb 2001, 2002, 2005a, b,
Kääb et al. 2003a, Skvarca et al. 2003, Bolch et al.
2008, Scherler et al. 2008, Copland et al. 2009,
Quincey and Glasser 2009, Quincey et al. 2009a,
Heid and Kääb 2012).
The matching of orthoimages combined with
repeated DEMs provides horizontal displacements
and surface elevation changes. The vertical com-
ponent of the displacement can be estimated from
DEM gradients if surface parallel displacement is
assumed (Kääb and Funk 1999). The accuracy of
this approximation is determined by the representa-
tiveness of the DEM used compared with actual
terrain topography.
A combination of simultaneous multitemporal
and multiangle image matching where a surface
102 Glacier mapping and monitoring using multispectral data
particle is measured in all overlapping images of all
image acquisition dates is, in fact, able to deliver
three-dimensional surface particle displacements.
Such a procedure can be substantially simplified
by introducing approximated orthoimages (i.e.,
orthoimages computed from a coarse DEM; also
called pseudo-orthoimages) instead of the original
imagery (Kaufmann and Ladstädter 2002, Leprince
et al. 2007). The latter procedure and the above
combined orthoimage/DEM gradient approach
converge if DEMs with sufficiently high spatial
resolution are available. Repeat ASTER stereo
imagery (i.e., stereoscopic channels 3N and 3B)
open up the possibility of simultaneous matching
in both multitemporal and multiangular data for
direct measurement of three-dimensional displace-
ments.
4.8.3 Postprocessing and analysis
All image-matching results contain inaccurate
measurements or mismatches. If the probability
of such errors is insufficiently reduced by preproces-
sing steps such as point-of-interest operators (För-
stner 2000, Kaufmann and Ladstädter 2002,
Debella-Gilo and Kääb 2011), a number of post-
processing steps are available to detect and remove
erroneous measurements:
1. Minimum and maximum acceptable speeds (i.e.,
velocity magnitudes) can be set.
2. Most matching algorithms provide a measure of
matching quality, such as cross-correlation coef-
ficients or signal-to-noise ratios. A threshold can
be set to eliminate results below a certain quality
measure.
3. For single glaciers a directional slice (range of
direction angles) can be chosen for velocity vec-
tors to be accepted. Similarly, a maximum direc-
tional offset can be chosen between the original
displacement vectors and those from other
sources (e.g., the aspect from a DEM).
4. Simple filters can be used to identify potentially
erroneous measurements such as a running vec-
tor median or RMS windows.
5. Original measurements can be compared with a
low-pass filtered version of the displacement field
and an acceptance threshold set on the respective
difference of measurements (Heid and Kääb
2012). A similar procedure can be built into
the matching process itself to limit matching pos-
itions to those that are most probable (Evans
2000).
6. The original image 1 can be compared with a
simulated image 1S which is obtained by inter-
polating image 2 using reverse displacements
matched between 1 and 2. The correlation
between 1 and 1S is an indicator of matching
success, one that is relative because the overall
correlation level between 1 and 1S depends not
only on the displacement between times 1 and 2,
but also different imaging conditions, etc. The
method is thus particularly useful for algorithm
comparisons (Debella-Gilo and Kääb 2011).
7. Instead of applying image matching to a single
pair of raw or preprocessed images, a number of
versions of the input images can be matched
(PCs, directional filtered, etc.) and from the stack
of results the most probabe match can be
selected (Ahn and Howat 2011).
Glacier velocity fields from image matching can be
analyzed in various ways, such as by evaluating
transverse and longitudinal profiles, spatial gradi-
ents (i.e., ice deformation), temporal variations,
streamline and relative age calculations, and by
direct inclusion in numerical flow models, etc.
(Kääb 2005b).
4.8.4 Accuracy
The total error budget of displacement vectors
matched from repeat images entails:
. The precision of the matching algorithm, which
may be of the order of 0.1 to 0.2 pixels.
. The representativeness of a match for the match-
ing window area covered. Measured displace-
ments are often assigned to the center pixel of
the window, but can in theory originate from
anywhere inside the window, depending on visual
contrast and the algorithm used. If the window
contains deformation, this also introduces error
(Debella-Gilo and Kääb 2011, 2012).
. The geolocation accuracy of individual pixels,
which might be of the same order of magnitude
as the pixel size itself, and stems from instabilities
within the sensor (e.g., mechanical scanner; some-
times called co-registration accuracy) or instabil-
ities of the satellite (e.g., jitter) which are not
sufficiently captured by attitude measurements
(Nuth and Kääb 2011, Heid and Kääb 2012).
. Image co-registration or orthoprojection errors
(e.g., from GCP or DEM errors).
. The degree to which tracked intensity variations
actually reflect underlying glacier movement (e.g.,

Figure 4.19. Radar interferogram between an ERS 1/2 tandem pair of April 5 and 6, 1996. The perpendicular
baseline is 12 m, so that topographic phase contribution is small and the most visible of the color fringes is due to ice
movement. Kronebreen at the middle left shows coherence loss due to large speeds (cf. Fig. 4.18). The large fringes
at the middle right indicate slowly deforming sea ice.
surface transformations may change intensity var-
iations independent of movement).
. Mismatches in which an intensity feature in the
search image, which does not correspond to the
same feature in the reference image, shows a
higher similarity than actually correct corre-
sponding features (e.g., self-similar crevasse or
ogive patterns, snow dunes).
Obviously, where the images to be matched lack
intensity variations that are sufficiently large and
stable over time, such as is often the case for accu-
mulation areas, image matching gets unreliable or
fails totally. As a result, the method will typically be
suitable for ablation areas and sections of accumu-
lation areas with pronounced features such as
crevasses, rock avalanche deposits, etc. In addition,
some matching methods are inherently more robust
in low-contrast areas than others (Heid and Kääb
2012).
Ice flow 103
4.8.5 SAR offset tracking
and interferometry
Phase differences in an interferogram constructed
from two synthetic aperture radar (SAR) images
include (i) the phase difference due to acquisition
geometry (curved Earth phase, ellipsoid), (ii) the
phase difference due to actual topography above
the ellipsoid (topographic phase), (iii) the displace-
ment phase due to any terrain shifts between the
two acquisition times, (iv) atmospheric phase dis-
tortions, and (v) noise (Fig. 4.19). Removal of the
curved Earth phase from an interferogram, and
subtraction of the topographic phase from the total
phase difference gives a differential interferogram in
which only displacement, atmospheric effects, and
noise remain (differential SAR interferometry,
DInSAR). The topographic phase can either be
simulated from a DEM, or inferred from multiple
interferograms the difference between which can be
used to eliminate a constant displacement term
(Bamler and Hartl 1998).
104 Glacier mapping and monitoring using multispectral data
Figure 4.20. Surface displacements on Kronebreen derived from SAR offset tracking using Radarsat-2 fine beam
data of April 6 and 30, 2009. Speeds increase from zero (light-blue/cyan) over pink, yellow to blue (1.3 m/day).
Radar interferometry requires phase coherence
between the acquisition times of individual SAR
images. Such coherence over ice and snow is usually
lost within hours or days, a fact that limits the
applicability of satellite radar interferometry
usually to the one-day tandem phase of ERS 1
and ERS 2 during 1995 and 1996, and the occa-
sional 3-day repeat cycle orbits of ERS (e.g., 1991),
and a coordinated ERS/Envisat campaign. Longer
periods of phase coherence of up to one orbit cycle
of typically several weeks or even longer are usually
only found for ice bodies in particularly stable cold
and dry conditions (e.g., Canadian Arctic during
winter, Antarctica, etc.).
New opportunities for radar interferometry over
glaciers will be enabled by ESA’s Sentinel-1 (2013
or 2014 launch). The C-band SAR of these twin
spacecraft will have 6-day revisitation and thus
should permit expanded use of InSAR.
Radar interferometry provides only the line-of-
sight component of displacement. Other compo-
nents have to be projected (e.g., assuming move-
ment along the direction of steepest descent as
calculated from a DEM). A combination of inter-
ferograms from ascending and descending orbits is
thus able to provide two components of a three-
dimensional displacement vector, in particular
for high latitudes with large azimuth differences
between the two orbit directions (Kwok and Fahne-
stock 1996).
Numerous ice flow studies using radar inter-
ferometry exist, mainly in polar regions, but fewer
exisst under typical mountain topography, for
smaller glaciers, and in lower latitudes (Luckman
et al. 2007, Quincey et al. 2009b).
A totally different technique applicable to repeat
SAR data is offset tracking (in principle equivalent
to matching between optical images). Backscatter
intensity or radar phase texture is tracked between
multitemporal SAR images using image-matching
techniques as described above. Like all image-
matching techniques, SAR offset tracking provides
two-dimensional displacements in the image plane.
In case of phase coherence (see above) phase texture
can be tracked with high accuracy (radar speckle
tracking).
In case of coherence loss (i.e., where radar inter-
ferometry is no longer possible), there might still be
sufficient SAR intensity texture to be matched
between the repeat images. Where coherence is
completely lost it might be an indication of rapid
ice flow, and so this provides a method of mapping
active ice areas even where debris cover makes ice
mapping difficult. Thus, coherence loss mapping
can help distinguish between actively flowing
debris-covered ice and completely stagnant ice. A

multispectral image analog to coherence loss map-
ping is the ICESMAP method described above,
though the computation is different, and each
method can sense different phenomena besides ice
motion.
The main differences between SAR offset track-
ing and matching of repeat optical data consist (i) in
the potential of SAR images to contain backscatter
intensity texture even in cases where optical images
would not contain sufficient texture (e.g., snow
areas) and (ii) in the much higher noise level of
SAR data (radar speckle). This high noise level
requires matching algorithms that are particularly
robust against noise, and much larger matching
templates than those used for optical data have to
be employed. The latter requirement limits applica-
tion of the method to larger glaciers. New high-
resolution SAR sensors such as Radarsat-2 or
TerraSAR-X, however, allow glacier flow to be
studied for smaller glaciers too by reducing the
geographic area of the matching templates (not
the template size in number of pixels), and by often
showing more scene details to be potentially
matched over time (Fig. 4.20).
4.9 CHALLENGES, CONCLUSIONS,
AND PERSPECTIVES
Classification approaches based on multispectral
imagery are thoroughly developed and well estab-
lished. Significant progress can, however, be
expected from newly developed combinations of
input layers and combinations of classification
algorithms. For ice and snow applications, the
inclusion of thermal bands is little tested and should
be further investigated.
Paul (2002, 2004) found band ratios to be an
especially simple, robust, and fast method for
glacier mapping over large areas. Manual adjust-
ment of the necessary threshold seems to be prefer-
able to thresholds automatically optimized from
training areas (Rott and Markl 1989). For compli-
cated situations adaptive threshold variations over
one scene depending on variations of the ground or
illumination properties might be investigated.
There is still much potential in using spectral data
and its derivatives alone, but multidimensional
classification including, for instance, DEMs may
be most promising since they best reflect the nature
of high-mountain phenomena and processes. The
increasing availability of suitable DEMs supports
this trend. The manual application of expert
Challenges, conclusions, and perspectives 105
knowledge for delineation will be increasingly
reduced but never become redundant (Rott and
Markl 1989).
Hyperspectral analysis of the glacial, periglacial,
and paraglacial environment could provide new
advantages not only for classification, but also for
understanding numerous phenomena and pro-
cesses.
While promising results are already available for
geological applications, very little is known about
the hyperspectral response of alpine environments
within rugged topography (Keller et al. 1998,
Schläpfer et al. 2000, Gruber et al. 2003, Casey et
al. 2012). In contrast to multispectral applications,
hyperspectral methods have large potential to lead
much further beyond the human ability for visual
interpretation.
Microwave sensors respond especially sensitively
to snow and ice and their physical properties due to
varying liquid water content and varying penetra-
tion depths. However, much research is still needed
into the complex dielectric properties of snow and
ice under various conditions to facilitate linking
recorded microwave responses to the geophysical
characteristics of ground materials.
Optical and microwave data for glacier remote
sensing are in a number of aspects complementary,
so that their combination might offer new insights
and approaches. In principle, the combination of
sensor data from both the optical visible infrared
region and the microwave/radar regions could
assist glacier analysis in the following ways:
. DEM generation and elevation change detection:
a major shortcoming of optical stereo DEMs is
their dependence on visual contrast. InSAR DEM
generation is unaffected by this but relies instead
on the sensitivity of measuringphase coherence.
Optical DEMs can serve as initial solutions for an
InSAR DEM by solving the modulo 2 ambiguity
during phase unwrapping, in particular over low-
coherence areas. Such a DEM gets its overall
vertical scale from an auxiliary DEM and its
details from radar interferometry. Similarly, the
absolute vertical scale of an InSAR DEM that is
usually uncertain due to baseline uncertainty can
be refined by fitting an InSAR DEM to an
auxiliary (e.g., optical) DEM or altimetry data
(Moholdt and Kääb 2012)
. Quantification of glacier movement: optimal con-
ditions for differential radar interferometry, radar
offset tracking, and repeat optical image matching
differ significantly from each other, and may vary
106 Glacier mapping and monitoring using multispectral data
even over one scene at one point in time. Radar
interferometry requires phase coherence which is
limited to small displacements and deformations
(among other factors). Radar offset tracking does
not suffer from such problems, but is still more
sensitive to ground changes than repeat optical
data and requires much larger matching windows.
On the other hand, sufficient contrast in the back-
scatter features necessary for SAR offset tracking
is usually found over much larger areas than the
visual contrast necessary for optical matching.
. Spectral analysis of glaciers: spectral fusion of
optical and SAR data might be especially prom-
ising for spectral analysis of glaciers (Rott and
Strobl 1992, Rott 1994, Hall et al. 1995a, 2009,
König et al. 2001, Warren and Brandt 2008).
In sum, despite finding optical multispectral
imaging to be powerful for glacier remote sensing
and radar/microwave techniques having their own
advantages, combining the two approaches to a
constellation undergoing coordinated targeting
would bring about many opportunities for funda-
mental advances in glacier remote sensing.
4.10 ACKNOWLEDGMENTS
This chapter was supported by the ESA Climate
Change Initiative project Glaciers\_cci
(4000101778/10/I-AM), ESA GlobGlacier (21088/
07/I-EC), the EU’s Seventh Framework Pro-
gramme (FP/2007–2013)/ERC Grant Agreement
No. 320816, Norwegian Space Centre/ESA PRO-
DEX (4000106033) and The Research Council of
Norway through the CORRIA (185906/V30) and
RASTAR projects (208013/F50).
4.11 REFERENCES
Adhikary, S., Nakawo, M., Seko, K., and Shakya, B.
(2000) Role of supraglacial ponds in the ablation pro-
cess of a debris-covered glacier in the Nepal Hima-
layas. In: Nakawo, M., Raymond, C., and Fountain,
A. (Eds.), Debris-covered Glaciers, International Asso-
ciation of Hydrological Sciences, Rennes, France,
pp. 43–52.
Ahn, Y., and Howat, I. (2011) Efficient automated glacier
surface velocity measurement from repeat images using
multi-image/multichip and null exclusion feature
tracking. IEEE Transactions on Geoscience and Remote
Sensing, 49(8), 2838–2846.
Albert, T.H. (2002) Evaluation of remote sensing tech-
niques for ice-area classification applied to the tropical
Quelccaya Ice Cap, Peru. Polar Geography, 26(3), 210–
226.
Andreassen, L., Paul, F., Kääb, A., and Hausberg, J.
(2008) Landsat-derived glacier inventory for Jotunhei-
men, Norway, and deduced glacier since the 1930s. The
Cryosphere, 2, 131–145.
Aniya, M., Sato, H., Naruse, R., Skvarca, P., and
Casassa, G. (1996) The use of satellite and airborne
imagery to inventory outlet glaciers of the Southern
Patagonia Icefield, South America. Photogrammetric
Engineering & Remote Sensing, 62, 1361–1369.
Atwood, D., Meyer, F., and Arendt, A. (2010) Using L-
band SAR coherence to delineate glacier extent. Cana-
dian Journal of Remote Sensing, 36, S186–S195.
Bamler, R. and Hartl, P. (1998) Synthetic aperture
radar interferometry. Inverse Problems, 14, R1–R54.
Bhambri, R., Bolch, T., Chaujar, R.K., and Kulsh-
reshtha, S.C. (2011) Glacier changes in the Garhwal
Himalaya, India, from 1968 to 2006 based on remote
sensing. Journal of Glaciology, 57(203), 543–556.
Binaghi, E., Madella, A., Madella, P., and Rampini, A.
(1993) Integration of remote sensing images in a GIS
for the study of alpine glaciers. In: Winkler, P. (Ed.),
Remote Sensing for Monitoring the Changing Environ-
ment of Europe, Balkema, Rotterdam, The Nether-
lands, pp. 173–178.
Binaghi, E., Madella, P., Montesano, M., and Rampini,
A. (1997) Fuzzy contextual classification of multi-
source remote sensing images. IEEE Transactions on
Geoscience and Remote Sensing, 35(2), 326–339.
Bishop, M., Shroder, J., and Hickman, B. (1999) High
resolution satellite imagery and neuronal networks for
information extraction in a complex mountain en-
vironment. Geocarto International, 14, 17–26.
Bishop, M.P., Bonk, R., Kamp, U., and Shroder, J.
(2001) Terrain analysis and data modeling for alpine
glacier mapping. Polar Geography, 25(3), 182–201.
Bishop, M., Olsenholler, J., Shroder, J., Barry, R., Raup,
B., Bush, A., Copland, L., Dwyer, J., Fountain, A.,
Haeberli, W. et al. (2004). Global Land Ice Measure-
ments from Space (GLIMS): Remote sensing and GIS
investigations of the Earth’s cryosphere. Geocarto
International, 19(2).
Bolch, T., and Kamp, U. (2006) Glacier mapping in high
mountains using DEMs, Landsat and ASTER data.
Grazer Scriften der Geographie und Raumforschung,
41, 37–48.
Bolch, T., Buchroithner, M., Kunert, A., and Kamp, U.
(2007) Landsat-based inventory of glaciers in western
Canada, 1985–2005. EARSeL eProceedings, 27, 403–
410.
Bolch, T., Buchroithner, M.F., Peters, J., Baessler, M.,
and Bajracharya, S. (2008) Identification of glacier
motion and potentially dangerous glacial lakes in the
Mt. Everest region/Nepal using spaceborne imagery.

Natural Hazards and Earth System Science, 8(6), 1329–
1340.
Bolch, T., Menounos, B., and Wheate, R. (2010a)
Landsat-based inventory of glaciers in western
Canada, 1985–2005. Remote Sensing of Environment,
114(2010), 127–137.
Bolch, T., Yao, T., Kang, S., Buchroithner, M., Scherer,
D., Maussion, F., Huintjes, E., and Schneider, C.
(2010b) A glacier inventory for the western Nyainqen-
tanglha Range and the Nam Co Basin, Tibet, and
glacier changes 1976–2009. The Cryosphere, 4(3),
419–433.
Bolch, T., Pieczonka1, T., and Benn, D. (2011) Multi-
decadal mass loss of glaciers in the Everest area (Nepal
Himalaya) derived from stereo imagery. The Cryo-
sphere, 5(2), 349–358.
Bourdelles, B. and Fily, M. (1993). Snow grain size deter-
mination from Landsat imagery. Annals of Glaciology,
17, 87–92.
Bronge, L., and Bronge, C. (1999) Ice and snow-type
classification in the Vestfold Hills, East Antarctica,
using Landsat-TM data and ground radiometer meas-
urements. International Journal of Remote Sensing, 20,
225–240.
Brown, D., Lusch, D., and Duda, K. (1998) Supervised
classification of types of glaciated landscapes using
digital elevation data. Geomorphology, 21, 233–250.
Casey, K., Kääb, A., and Benn, D. (2012) Geochemical
characterization of supraglacial debris via in situ and
optical remote sensing methods: A case study in
Khumbu Himalaya, Nepal. The Cryosphere, 6, 85–100.
Clark, R. (1999) Spectroscopy of rocks and minerals, and
principles of spectroscopy. In A. Rencz (Ed.), Remote
Sensing for the Earth Sciences: Manual of Remote
Sensing, Third Edition, John Wiley & Sons, New
York,, 358 pp.
Cogley, J., Hock, R., Rasmussen, L., Arendt, A., Bauder,
A., Braithwaite, R., Jansson, P., Kaser, G., Möller, M.,
Nicholson, L. et al. (2011) Glossary of Glacier Mass
Balance and Related Terms (Technical Report 86, IHP-
VII Technical Documents in Hydrology, IACS Con-
tribution No. 2), UNESCO–IHP.
Copland, L., Pope, S., Bishop, M., Shroder, J., Clendon,
P., Bush, A., Kamp, U., Seong, Y., and Owen, L.
(2009) Glacier velocities across the central Karakoram.
Annals of Glaciology, 50(52), 41–49.
Crippen, R. (1988) The dangers of underestimating the
importance of data adjustments in band ratioing. Inter-
national Journal of Remote Sensing, 9(4), 767–776.
Crippen, R.E. (1989) A simple spatial filtering routine for
the cosmetic removal of scan-line noise from Landsat
TM p-tape imagery. Photogrammetric Engineering and
Remote Sensing, 55(3), 327–331.
Debella-Gilo, M., and Kääb, A. (2011) Sub-pixel pre-
cision image matching for measuring surface displace-
ments on mass movements using normalized cross-
References 107
correlation. Remote Sensing of Environment, 115(1),
130–142.
Debella-Gilo, M., and Kääb, A. (2012) Measurement of
surface displacement and deformation of mass move-
ments using least squares matching of repeat high reso-
lution satellite and aerial images. Remote Sensing, 4(1),
43–67.
Della Ventura, A., Rampini, A., Rabagliati, R., and Ser-
andrei Barbero, R. (1983) Glacier monitoring by satel-
lite. Il Nuovo Cimento, C-1(6), 211–221.
Dowdeswell, J., and Benham, T. (2003) A surge of Per-
seibreen, Svalbard, examined using aerial photography
and ASTER high-resolution satellite imagery. Polar
Research, 22(2), 373–383.
Dozier, J. (1989). Spectral signature of alpine snow cover
from the Landsat Thematic Mapper. Remote Sensing
of Environment, 28, 9–22.
Engeset, R., and Ødegard, R.S. (1999) Comparison of
annual changes in winter ERS-1 SAR images and
glacier mass balance of Slakbreen, Svalbard. Inter-
national Journal of Remote Sensing, 20(2), 259–271.
Evans, A.N. (2000) Glacier surface motion computation
from digital image sequences. IEEE Transactions on
Geoscience and Remote Sensing, 38(2), 1064–1072.
Förstner, W. (2000) Image preprocessing for feature
extraction in digital intensity, color and range images.
In: A. Dermanis, A. Gruen, and F. Sans (Eds.), Geo-
matic Methods for the Analysis of Data in the Earth
Sciences (Lecture Notes in Earth Sciences), Springer-
Verlag, Berlin, pp. 165–189.
Frezzotti, M., Capra, A., and Vittuari, L. (1998) Compar-
ison between glacier ice velocities inferred from GPS
and sequential satellite images. Annals of Glaciology,
27, 54–60.
Fujii, Y. (1977) Field experiment on glacier ablation
under a layer of debris cover. Journal of the Japanese
Society of Snow and Ice, 39, 20–21.
Fujimura, S., and Kiyasu, S. (1999) Comparison between
glacier ice velocities inferred from GPS and sequential
satellite images. Paper presented at IEEE Geoscience
and Remote Sensing Symposium, IGARS-99, pp. 1192–
1194.
Gillespie, A., Kahle, A., and Walker, R. (1986) Color
enhancement of highly correlated images, I: Decorrela-
tion and HSI contrast stretches. Remote Sensing of
Environment, 20, 209–235.
Gratton, D., Howarth, P., and Marceau, D. (1990) Com-
bining DEM parameters with Landsat MSS and TM
imagery in a GIS for mountain glacier characteriza-
tion. IEEE Transactions on Geoscience and Remote
Sensing, 28(4), 766–769.
Gratton, D., Howarth, P., and Marceau, D. (1993) Using
Landsat-5 Thematic Mapper and digital elevation data
to determine the net radiation field of a mountain
glacier. Remote Sensing of Environment, 43, 315–331.
Gruber, S., Schläpfer, D., and Hoelzle, M. (2003) Snow-
free ground albedo derived from DAIS 7915 hyper-
108 Glacier mapping and monitoring using multispectral data
spectral imagery for energy balance modeling in high-
alpine topography. Paper presented at Eighth Inter-
national Conference on Permafrost, Extended
Abstracts, pp. 47–48.
Gruen, A., and Baltsavias, E. (1987) High-precision
image matching for digital terrain modelling. Photo-
grammetria, 42, 97–112.
Gudmundsson, G., and Bauder, A. (1999) Towards an
indirect determination of the mass-balance distribution
of glaciers using the kinematic boundary condition.
Geografiska Annaler, 81A(4), 575–583.
Haeberli, W., and Epifani, F. (1986) Mapping the dis-
tribution of buried glacier ice: An example from Lago
Delle Locce, Monte Rosa, Italian Alps. Annals of
Glaciology, 8, 78–81.
Hall, D., Chang, A., and Siddalingaiah, H. (1988) Reflec-
tances of glaciers as calculated using Landsat 5 The-
matic Mapper data. Remote Sensing of Environment,
25(3), 311–321.
Hall, D., Chang, A., Foster, J., Benson, C., and Kovalick,
W. (1989) Comparison of in situ and Landsat derived
reflectances of Alaskan glaciers. Remote Sensing of En-
vironment, 28, 23–31.
Hall, D., Bindschadler, R., Foster, J., Chang, A., and
Siddalingaiah, H. (1990) Comparison of in situ and
satellite derived reflectances of Forbindels Glacier,
Greenland. International Journal of Remote Sensing,
11, 493–504.
Hall, D., Williams, R., and Bayr, K. (1992) Glacier reces-
sion in Iceland and Austria. EOS Trans. Am. Geophys.
Union, 73(12), 129, 135, 141.
Hall, D., Williams, R., and Sigurdsson, O. (1995a)
Glaciological observations on Bruarjökull, Iceland,
using synthetic aperture radar and Thematic Mapper
satellite data. Annals of Glaciology, 21, 271–276.
Hall, D.K., Riggs, G., and Salomonson, V. (1995b)
Development of methods for mapping global snow
cover using moderate resolution imaging spectro-
radiometer data. Remote Sensing of Environment,
54(2), 127–140.
Hall, D., Williams, R., Casey, K., DiGirolamo, N., and
Wan, Z. (2006) Satellite-derived, melt-season surface
temperature of the Greenland ice sheet (2000–2005)
and its relationship to mass balance. Geophysical
Research Letters, 33(11).
Hall, D., Nghiem, S., Schaaf, C., DiGirolamo, N., and
Neumann, G. (2009) Evaluation of surface and near-
surface melt characteristics on the Greenland ice sheet
using MODIS and QuikSCAT data. Journal of Geo-
physical Research—Earth Surface, 114.
Haug, T., Kääb, A., and Skvarca, P. (2010) Monitoring
ice shelf velocities from repeat MODIS and Landsat
data: A method study on the Larsen C ice shelf, Ant-
arctic Peninsula, and 10 other ice shelves around Ant-
arctica. The Cryosphere, 4(2):161–178.
Heid, T., and Kääb, A. (2012) Evaluation of existing
image matching methods for deriving glacier surface
displacements globally from optical satellite imagery.
Remote Sensing of Environment, 118, 339–355.
Hoelzle, M.,Mittaz, C., Etzelmöller, B., and Haeberli, W.
(2001) Surface energy fluxes and distribution models of
permafrost in European mountain areas: An overview
of current developments. Permafrost and Periglacial
Processes, 12(1), 53–68.
Holben, B., and Justice, C. (1981) An examination of
spectral band rationing to reduce the topographic
effect on remotely sensed data. International Journal
of Remote Sensing, 2, 115–133.
Huggel, C., Kääb, A., Haeberli, W., Teysseire, P., and
Paul, F. (2002) Satellite and aerial imagery for analys-
ing high mountain lake hazards. Canadian Geo-
technical Journal, 39(2), 316–330.
Hugli, H., and Frei, W. (1983) Understanding anisotropic
reflectance in mountainous terrain. Photogrammetric
Engineering & Remote Sensing, 49(5), 671–683.
Ingram, K., Knapp, E., and Robinson, J. (1981) Change
Detection Technique Development for Improved
Urbanized Area Delineation (Technical Report Tech-
nical memorandum CSCITM- 81/6087), Computer
Sciences Corporation, Falls Church, VA.
Jensen, J. (1983) Urban/suburban Land Use Analysis:
Manual of Remote Sensing 2, American Society of
Photogrammetry, Falls Church, VA.
Jensen, J. (2005) Introductory Digital Image Processing: A
Remote Sensing Perspective, Second Edition, Pearson
Prentice Hall, Upper Saddle River, NJ.
Kääb, A. (2001) Photogrammetric reconstruction of
glacier mass balance using a kinematic iceflow model:
A 20-year time-series on Grubengletscher, Swiss Alps.
Annals of Glaciology, 31, 45–52.
Kääb, A. (2002) Monitoring high-mountain terrain
deformation from repeated air- and spaceborne optical
data: Examples using digital aerial imagery and
ASTER data. Photogrammetry and Remote Sensing,
57, 39–52.
Kääb, A. (2005a) Combination of SRTM3 and repeat
ASTER data for deriving alpine glacier flow velocities
in the Bhutan Himalaya. Remote Sensing of Environ-
ment, 94(1), 463–474.
Kääb, A. (2005b) Remote Sensing of Mountain Glaciers
and Permafrost Creep, Vol. 48, Schriftenreihe Phy-
sische Geographie, Glaziologie und Geomorphodyna-
mik, University of Zürich, Zürich.
Kääb, A. and Funk, M. (1999) Modelling mass balance
using photogrammetric and geophysical data: A pilot
study at Griesgletscher, Swiss Alps. Journal of Glaciol-
ogy, 45(151), 575–583.
Kääb, A. and Vollmer, M. (2000) Surface geometry,
thickness changes and flow fields on creeping mountain
permafrost: Automatic extraction by digital image
analysis. Permafrost and Periglacial Processes, 11(4),
315–326.
Kääb, A., Haeberli, W., and Gudmundsson, G. (1997)
Analysing the creep of mountain permafrost using high

precision aerial photogrammetry: 25 years of monitor-
ing Gruben rock glacier, Swiss Alps. Permafrost and
Periglacial Processes, 8(4), 409–426.
Kääb, A., Gudmundsson, G., and Hoelzle, M. (1998)
Surface deformation of creeping mountain permafrost:
Photogrammetric investigations on rock glacier Mur-
tel. Paper presented at Proceedings of the Seventh Inter-
national Permafrost Conference, Yellowknife, Canada,
pp. 531–537.
Kääb, A., Paul, F., Maisch, M., Hoelzle, M., and Hae-
berli, W. (2002) The new remote-sensing-derived Swiss
glacier inventory, II: First results. Annals of Glaciology,
34, 362–366.
Kääb, A., Huggel, C., Paul, F., Wessels, R., Raup, B.,
Kieffer, H., and Kargel, J. (2003a) Glacier monitoring
from ASTER imagery: Accuracy and applications.
EARSeL eProceedings (LISSIG Workshop, Berne,
March 11–13, 2002), 2(1), 43–53.
Kääb, A., Isakowski, Y., Paul, F., Neumann, A., and
Winter, R. (2003b) Glaziale und periglaziale Prozesse:
Von der statischen zur dynamischen Visualisierung.
Kartographische Nachrichten, 535), 206–212.
Kääb, A., Lefauconnier, B., and Melvold, K. (2006) Flow
field of Kronebreen, Svalbard, using repeated Land-
sat7 and ASTER data. Annals of Glaciology, 42, 7–13.
Kääb, A., Berthier, E., Nuth, C., Gardelle, J., and
Arnaud, Y. (2012) Contrasting patterns of early
twenty-first-century glacier mass change in the Hima-
layas. Nature, 488, 495–498.
Kargel, J.S., Ahlstrøm, A.P., Alley, R.B., Bamber, J.L.,
Benham, T.J., Box, J.E., Chen, C., Christoffersen, P.,
Citterio, M., Cogley, J.G. et al. (2012) Greenland’s
shrinking ice cover: ‘‘Fast times’’ but not that fast
(Brief communication). The Cryosphere, 6(3), 533–537.
Kaufmann, V., and Ladstädter, R. (2002) Spatio-
temporal analysis of the dynamic behaviour of the
Hochebenkar rock glaciers (Oetztal Alps, Austria) by
means of digital photogrammetric methods. Grazer
Schriften der Geographie und Raumforschung, 37,
119–140.
Kayastha, R., Takeuchi, Y., Nakawo, M., and Ageta, Y.
(2000) Role of supraglacial ponds in the ablation
process of a debris-covered glacier in the Nepal Hima-
layas. In: M. Nakawo, C. Raymond, and A. Fountain
(Eds.), Debris-covered Glaciers, International Associa-
tion of Hydrological Sciences, Rennes, France, pp. 71–
81.
Keller, P., Keller, I., and Itten, K. (1998) Combined
hyperspectral data analysis of two Alpine lakes using
CASI and DAIS 7915 imagery. Paper presented at
EARSeL Workshop, University of Zurich, Zurich,
Switzerland.
Kelley, R., Engeset, R., Kennett, M., Barrett, E., and
Theakstone, W. (1997) Characteristic snow and ice
properties of a Norwegian ice cap determined from
complex ERS SAR. Paper presented at Proceedings
References 109
of the Third ERS Symposium, ESA, Noordwijk, The
Netherlands.
Klein, A.G., Seltzer, G.O., and Isacks, B.L. (1999) Mod-
ern and last local glacial maximum snowlines in the
Central Andes of Peru, Bolivia, and Northern Chile.
Quaternary Science Reviews, 18, 63–84.
Knap, W., Brock, B., Oerlemans, J., and Willis, I. (1999)
Comparison of Landsat-TM derived and ground-
based albedos of Haut Glacier d’Arolla, Switzerland.
International Journal of Remote Sensing, 20(17), 3293–
3310.
Koelemeijer, R., Oerlemanns, J., and Tjemkes, S. (1993)
Surface reflectance of Hintereisferner, Austria, from
Landsat 5 TM imagery. Annals of Glaciology, 17, 17–
22.
König, M., Winther, J.-G., and Isaksson, E. (2001)
Measuring snow and ice properties from satellite.
Reviews of Geophysics, 39(1), 1–27.
Kwok, R., and Fahnestock, M. (1996) Ice sheet motion
and topography from radar interferometry. IEEE
Transactions on Geoscience and Remote Sensing,
34(1), 189–200.
Lefauconnier, B., Hagen, J., and Rudant, J. (1994) Flow
speed and calving rate of Kronebreen Glacier, Sval-
bard, using SPOT images. Polar Research, 13(1), 59–
65.
Leprince, S., Ayoub, F., Klinger, Y., and Avouac, J.
(2007) Co-registration of optically sensed images and
correlation (COSI-Corr): An operational methodology
for ground deformation measurements. Paper pre-
sented at IEEE Geoscience and Remote Sensing Sym-
posium, IGARS-2007, 1943–1946.
Liang, J. (2004) Quantitative Remote Sensing of Land
Surfaces, John Wiley & Sons, New York.
Lillesand, T., and Kieffer, R. (2000) Remote Sensing and
Image Interpretation, Fourth Edition, John Wiley &
Sons, New York.
Luckman, A., Quincey, D., and Bevan, S. (2007) The
potential of satellite radar interferometry and feature
tracking for monitoring flow rates of Himalayan
glaciers. Remote Sensing of Environment, 111(23),
172–181.
Markham, B., and Barker, J. (1985) Spectral character-
ization of the Landsat Thematic Mapper sensors.
International Journal of Remote Sensing, 6(5), 697–716.
Marshall, G., Rees, W., and Dowdeswell, J. (1995) The
discrimination of glacier facies using multitemporal
ERS-1 SAR data. In: J. Askne (Ed.), Sensors and En-
vironmental Applications of Remote Sensing, Balkema,
Rotterdam, The Netherlands, pp. 263–269.
Mattson, L., Gardner, J., and Young, G. (1993) Ablation
on debris covered glaciers: An example from the
Rakhiot Glacier, Punjab, Himalaya. Snow and Glacier
Hydrology, IAHS, 218, 289–296.
Mihalcea, C., Mayer, C., Diolaiuti, G., D’Agata, C.,
Smiraglia, C., Lambrecht, A., Vuillermoz, E., and Tar-
tari, G. (2008) Spatial distribution of debris thickness
110 Glacier mapping and monitoring using multispectral data
and melting from remote-sensing and meteorological
data, at debris-covered Baltoro Glacier, Karakoram,
Pakistan. Annals of Glaciology, 48, 49–57.
Mittaz, C., Hoelzle, M., and Haeberli, W. (2000) First
results and interpretation of energy-flux measurements
of Alpine permafrost. Annals of Glaciology, 31, 275–
280.
Moholdt, G., and Kääb, A. (2012) A new DEM of the
Austfonna Ice Cap by combining differential SAR
interferometry with ICESat laser altimetry. Polar
Research, 31(18460).
Mouat, D., Mahin, G., and Lancaster, J. (1993) Remote
sensing techniques in the analysis of change detection.
Geocarto International, 2, 39–50.
Nagler, T. and Rott, H. (2000) Retrieval of wet snow by
means of multitemporal SAR data. IEEE Transactions
on Geoscience and Remote Sensing, 38(2), 754–765.
Nakawo, M. and Young, G. (1982) Estimate of glacier
ablation under a debris layer from surface-temperature
and meteorological variables. Journal of Glaciology,
28(98), 29–34.
Nakawo, M., Yabuki, H., and Sakai, A. (1999) Charac-
teristics of Khumbu Glacier, Nepal Himalaya: Recent
change in the debris-covered area. Annals of Glaciol-
ogy, 28, 118–122.
Narama, C., Kääb, A., Duishonakunov, M., and
Abdrakhmatov, K. (2010) Spatial variability of recent
glacier area changes in the Tien Shan mountains, Cen-
tral Asia, using Corona (1970), Landsat (2000), and
ALOS (2007) satellite data. Global and Planetary
Change, 71(12), 42–54.
Nicholson, L., and Benn, D. (2006) Calculating ice melt
beneath a debris layer using meteorological data.
Journal of Glaciology, 52(178), 463–470.
Nuth, C., and Kääb, A. (2011) Co-registration and bias
corrections of satellite elevation data sets for quantify-
ing glacier thickness change. The Cryosphere, 5(1),
271–290.
Østrem, G. (1959) Ice melting under a thin layer of
moraine, and the existence of ice cores in moraine
ridges. Geografiska Annaler, 41(4), 228–230.
Painter, T. (2011) Comment on Singh et al., ‘‘Hyperspec-
tral analysis of snow reflectance to understand the
effects of contamination and grain size.’’ Journal of
Glaciology, 57(201), 183–185.
Painter, T., Duval, B., Thomas, W., Mendez, M.,
Heintzelman, S., and Dozier, J. (2001) Detection and
quantification of snow algae with an airborne imaging
spectrometer. Applied and Environmental Microbiol-
ogy, 67(11), 5267–5272.
Painter, T.H., Dozier, J., Roberts, D.A., Davis, R.E., and
Green, R.O. (2003) Retrieval of subpixel snow-covered
area and grain size from imaging spectrometer data.
Remote Sensing of Environment, 85(1), 64–77.
Paul, F. (2001) Evaluation of different methods for glacier
mapping using Landsat TM. EARSeL eProceedings, 1,
239–245.
Paul, F. (2002) Changes in glacier area in Tyrol, Austria,
between 1969 and 1992 derived from Landsat 5
Thematic Mapper and Austrian glacier inventory data.
International Journal of Remote Sensing, 23(4), 787–
799.
Paul, F. (2004) The new Swiss glacier inventory 2000:
Application of remote sensing and GIS. PhD thesis,
University of Zurich.
Paul, F., and Kääb, A. (2005) Perspectives on the produc-
tion of a glacier inventory from multispectral satellite
data in the Canadian Arctic: Cumberland Peninsula,
Baffin Island. Annals of Glaciology, 42, 59–66.
Paul, F., Kääb, A., Maisch, M., Kellenberger, T., and
Haeberli, W. (2002) The new remote-sensing-derived
Swiss glacier inventory, I: Methods. Annals of Glaciol-
ogy, 34, 355–361.
Paul, F., Huggel, C., and Kääb, A. (2004) Mapping of
debris-covered glaciers using multispectral and DEM
classification techniques. Remote Sensing of Environ-
ment, 89, 510–518.
Paul, F., Kääb, A., and Haeberli, W. (2007) Recent
glacier changes in the Alps observed by satellite:
Consequences for future monitoring strategies. Global
and Planetary Change, 56, 111–122, doi: 10.1016/j.
gloplacha.2006.07.007.
Paul, F., Barry, R., Cogley, J., Frey, H., Haeberli, W.,
Ohmura, A., Ommanney, C., Raup, B., Rivera, A.,
and Zemp, M. (2009) Recommendations for the com-
pilation of glacier inventory data from digital sources.
Annals of Glaciology, 50(53), 119–126.
Pietroniro, A., and Leconte, R. (2000) A review of
Canadian remote sensing applications in hydrology,
1995–1999. Hydrological Processes, 14, 1641–1666.
Quincey, D. and Glasser, N. (2009) Morphological and
ice-dynamical changes on the Tasman Glacier, New
Zealand, 1990–2007. Global and Planetary Change,
68(3), 185–197.
Quincey, D., Copland, L., Mayer, C., Bishop, M.,
Luckman, A., and Belo, M. (2009a) Ice velocity and
climate variations for Baltoro Glacier, Pakistan.
Journal of Glaciology, 55(194), 1061–1071.
Quincey, D., Luckman, A., and Benn, D. (2009b) Quan-
tification of Everest region glacier velocities between
1992 and 2002, using satellite radar interferometry and
feature tracking. Journal of Glaciology, 55(192), 596–
606.
Qunzhu, Z., Meisheng, C., Xuezhi, F., Fengxian, L.,
Xianzhang, C., and Wenkun, S. (1983) A study of
spectral reflection characteristics for snow, ice and
water in the north of China. In: B. Goodison (Ed.),
Hydrological Applications of Remote Sensing and
Remote Data Transmission, International Association
of Hydrological Sciences, Rennes, France, pp. 451–
462.
Rana, B., Nakawo, M., Fukushima, Y., and Ageta, Y.
(1997) Application of a conceptual precipitation-runoff
model (HYCYMODEL) in a debris-covered glacier-

ized basin in the Langtang Valley, Nepal Himalaya.
Annals of Glaciology, 25, 226–231.
Richards, J.A. (1993) Remote Sensing Digital Image Anal-
ysis, Second Edition, Springer-Verlag, Berlin.
Rolstad, C., Amlien, J., Hagen, J., and Lundén, B. (1997)
Visible and near-infrared digital images for determina-
tion of ice velocities and surface elevation during a
surge on Osbornebreen, a tidewater glacier in Sval-
bard. Annals of Glaciology, 24, 255–261.
Rott, H. (1976) Analyse der Schneeflocken auf Gletschern
der Tiroler Zentralalpen aus Landsat-Bildern. Zeit-
schrift für Gletscherkunde und Glazialgeologie, 12(1),
1–18.
Rott, H. (1994) Thematic studies in Alpine areas by
means of polarimetric SAR and optical imagery.
Advances in Space Research, 14(3), 217–226.
Rott, H., and Markl, G. (1989) Improved snow and
glacier monitoring by the Landsat Thematic Mapper.
Workshop on Landsat Themaic Mapper Applications
(Technical Report SP-1102), ESA, Noordwijk, The
Netherlands.
Rott, H., and Strobl, D. (1992) Synergism of SAR and
Landsat TM imagery for thematic mapping in complex
terrain. Advances in Space Research, 12(7), 425–431.
Rowan, L., and Mars, J. (2003) Lithologic mapping in the
Mountain Pass, California area using Advanced
Spaceborne Thermal Emission and Reflection Radio-
meter (ASTER) data. Remote Sensing of Environment,
84(3), 350–366.
Sabins, F. (1999) Remote sensing for mineral exploration.
Ore Geology Reviews, 14(34), 157–183.
Sakai, A., Takeuchi, N., Fujita, K., and Nakawo, M.
(2000) Role of supraglacial ponds in the ablation pro-
cess of a debris-covered glacier in the Nepal Hima-
layas. In: M. Nakawo, C. Raymond, and A.
Fountain (Eds.), Debris-covered Glaciers, International
Association of Hydrological Sciences, Rennes, France,
pp. 119–130.
Sandmeier, S., and Itten, K. (1997) A physically-based
model to correct atmospheric and illumination effects
in optical satellite data of rugged terrain. IEEE Trans-
actions on Geoscience and Remote Sensing, 35(3), 708–
717.
Scambos, T.A., Dutkiewicz, M.J., Wilson, J.C., and
Bindschadler, R.A. (1992) Application of image
cross-correlation to the measurement of glacier veloc-
ity using satellite image data. Remote Sensing of the
Environment, 42(3), 177–186.
Scherler, D., Leprince, S., and Strecker, M.R. (2008)
Glacier-surface velocities in alpine terrain from optical
satellite imagery: Accuracy improvement and quality
assessment. Remote Sensing of Environment, 112(10),
3806–3819.
Scherler, D., Bookhagen, B., and Strecker, M.R. (2011)
Spatially variable response of Himalayan glaciers to
climate change affected by debris cover. Nature
Geoscience, 4, 156–159.
References 111
Schiefer, E., Menounos, B., and Wheate, R. (2008) An
inventory and morphometric analysis of British
Columbia glaciers, Canada. Journal of Glaciology,
54(186), 551–560.
Schläpfer, D., Bojinski, S., Schaepman, M., and Richter,
R. (2000) Combination of geometric and atmospheric
correction for AVIRIS data in rugged terrain. Paper
presented at AVIRIS Workshop. In: R.O. Green (Ed.),
Summaries of the Ninth Annual JPL Airborne Earth
Science Workshop, Pasadena, California (JPL Publ.
00-18). Jet Propulsion Laboratory. Available at
ftp://popo.jpl.nasa.gov/pub/docs/workshops/00_docs/
Schlapfer_web.pdf
Schowengerdt, R.A. (2007) Remote Sensing: Models and
Methods for Image Processing, Third Edition,
Academic Press, San Diego, CA (ISBN 0-12-628981-
6).
Seko, K., Yakubi, H., Nakawo, M., Sakai, A., Kadota,
T., and Yamada, Y. (1998) Changing surface features
of Khumbu Glacier, Nepal Himalayas revealed by
SPOT images. Bulletin of Glaciological Research, 16,
33–41.
Shukla, A., Arora, M., and Gupta, R. (2010). Synergistic
approach for mapping debris-covered glaciers using
optical-thermal remote sensing data with inputs from
geomorphometric parameters. Remote Sensing of En-
vironment, 114, 1378–1387.
Sidjak, R., and Wheate, R. (1999) Glacier mapping of the
Illecillewaet Icefield, British Columbia, Canada, using
Landsat TM and digital elevation data. International
Journal of Remote Sensing, 20(2), 273–284.
Singh, A. (1989) Digital change detection techniques
using remotely sensed-data. International Journal of
Remote Sensing, 10(6), 989–1003.
Skvarca, P., Raup, B., and Angelis, H.D. (2003). Recent
behaviour of Glaciar Upsala, a fast flowing calving
glacier in Lago Argentino, southern Patagonia. Annals
of Glaciology, 36, 184–188.
Strozzi, T., Paul, F., and Kääb, A. (2010) Glacier map-
ping with ALOS PALSAR data within the ESA Glob-
Glacier project. Paper presented at Proceedings of the
ESA Living Planet Symposium.
Suzuki, R., Fujita, K., and Ageta, Y. (2007) Spatial dis-
tribution of thermal properties on debris-covered
glaciers in the Himalayas derived from ASTER data.
Bulletin of Glaciological Research, 24, 13–22.
Takeuchi, N. (2009) Temporal and spatial variations in
spectral reflectance and characteristics of surface dust
on Gulkana Glacier, Alaska Range. Journal of Glaciol-
ogy, 55(192), 701–709.
Taschner, S., and Ranzi, R. (2002) Comparing the oppor-
tunities of Landsat-TM and ASTER data for monitor-
ing a debris covered glacier in the Italian Alps within
the GLIMS project. Paper presented at Proceedings
IGARSS 2002, Toronto, ISBN 0-7803-7537-
8(4):1044–1046.
112 Glacier mapping and monitoring using multispectral data
Vermote, E., Tanré, D., Deuze, J., Herman, M., and
Morcette, J. (1997) Second simulation of the satellite
signal in the solar spectrum: An overview. IEEE Trans-
actions on Geoscience and Remote Sensing, 35(3), 675–
686.
Volesky, J., Stern, R., and Johnson, P. (2003) Geological
control of massive sulfide mineralization in the Neo-
proterozoic Wadi Bidah shear zone, southwestern
Saudi Arabia, inferences from orbital remote sensing
and field studies. Precambrian Research, 123(2), 235–
247.
Warren, S.G. (1982) Optical properties of snow. Reviews
of Geophysics and Space Physics, 20(1), 67–89.
Warren, S., and Brandt, R. (2008) Optical constants of ice
from the ultraviolet to the microwave: A revised com-
pilation. Journal of Geophysical Research—Atmo-
spheres, 113(D14).
Wessels, R.L., Kargel, J.S., and Kieffer, H.H. (2002)
ASTER measurement of supraglacial lakes in the
Mount Everest region of the Himalaya. Annals of
Glaciology, 34, 399–408.
Whalley, W. (1979) Relationship of glacier ice and rock
glacier at Grubengletscher, Kanton Wallis, Switzer-
land. Geografiska Annaler, Series A, Physical
Geography, 61(12), 49–61.
Whillans, I., and Tseng, Y.-H. (1995) Automatic tracking
of crevasses on satellite images. Cold Regions Science
and Technology, 23, 201–214.
Williams, R.S., Jr., Hall, D.K., and Benson, C.S. (1991)
Analysis of glacier facies using satellite techniques.
Journal of Glaciology, 37(125), 120–128.
Williams, R., Hall, D., Sigurdsson, O., and Chien, J.
(1997) Comparison of satellite-derived with ground-
based measurements of the fluctuations of the margins
of Vatnajökull, Iceland, 1973–92. Annals of Glaciology,
24, 72–80.
Winther, J. (1993). Landsat TM derived and in situ sum-
mer reflectance of glaciers in Svalbard. Polar Research,
12, 37–55.
Wiscombe, W.J. and Warren, S.G. (1980) A model for the
spectral albedo of snow, i: Pure snow. Journal of the
Atmospheric Sciences, 37, 2712–2733.
Zahnen, N., Jung-Rothenhäusler, F., Oerter, H.,
Wilhelms, F., and Miller, H. (2003) Correlation
between Antarctic dry snow properties and backscat-
tering characteristics in RADARSAT SAR imagery.
EARSeL eProceedings, 2, 140–148.