IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 4, APRIL 2001 885
Spatio-Temporal Landscape Analysis in Mountainous
Terrain by Means of Small Format Photography: A
Janette Aschenwald, Karin Leichter, Erich Tasser, and Ulrike Tappeiner
Abstract—A method is presented in order to georectify
high-oblique terrestrial images of high mountainous terrain taken
by means of small-format camera. Using the distinct topographic
situation of the study area located in a small alpine valley, an
automatic camera has been mounted on the opposing hill slope
providingdaily photographs of the area ofinterest. To usethe data
in a geographic information system (GIS) a specific georectifica-
tion method (JUKE method) was developed employing a digital
elevation model (DEM), several reference and ground control
points (GCPs), and the focal length of the lens. The method was
useful for single-stage image rectification but was also applied
successfully on pre-analyzed time-series images (“summary”
images). Hence, the procedure enables the analysis of detailed,
dynamic landscape ecological processes. Accuracy was mainly a
measure of the quality of the GCPs, which were difficult to define
in this remote area. The relatively economic data capturing and
transforming procedure makes the method interesting for various
applied disciplines in order to gain detailed spatio-temporal
data. This method might be of considerable benefit, particularly
in mountainous terrain where it is often difficult to capture
continuous spatio-temporal information.
Index Terms—Digital elevation model (DEM), georectification,
landscape ecology, mountain areas, oblique terrestrial images,
small format photography, time-series.
NE IMPORTANT aspect concerning landscape analysis
is the investigation of temporal changes of spatial phe-
nomena in mountain ecosystems. In the course of the past years,
this topic has received considerable attention in the literature as
landscape changes can be observed frequently and often lead to
serious alterations of ecosystem processes , .
However, accurate temporal investigations are rare, though
the analysis of detailed time series can lead to important find-
ings about landscape ecological processes. Integrated in a GIS,
the temporal component enables the analysis and simulation of
substantial dynamic processes , .
Manuscript received December 16, 1999; revised October 16, 2000. This
work was supported in part by the fourth framework of the EU, Project
ECOMONT (ENV4-CT95-0179), and by the INTERREG-II Project INTE-
J. Aschenwald is with the Department of Economic Policy and Economic
Theory, University of Innsbruck, Innsbruck, Austria (e-mail: janette.aschen-
K. Leichter and E. Tasser are with the European Academy Bolzano, Bolzano,
Italy (e-mail: email@example.com; firstname.lastname@example.org).
U. Tappeiner is with the Institute of Botany, University of Innsbruck, Inns-
bruck, Austria, and also with the European Academy Bolzano, Bolzano, Italy
Publisher Item Identifier S 0196-2892(01)02153-2.
But the possibilities of capturing time series with high tem-
poral resolution are very limited, particularly in mountain re-
gions. Conventional mapping systems sufferfrom variousdraw-
backs. Mapping in the field is extremely time consuming and
thus not feasible for short, regular time intervals. Aerial cam-
paigns are very costly and hence are usually restricted to only a
few selected dates. Satellite imagery is still not provided at suf-
ficient spatial and temporal resolutions needed for research in
small mountainous areas.
One economic alternative would be the use of a small format
camera that provides terrestrial photographs. High oblique im-
ages are usually captured in such cases due to the camera posi-
tion in relation to the study site. This consequently results in a
problem of georectification which needs to be considered seri-
ously in areas of mountainous terrain , .
Due to the complex image geometry, high oblique pho-
tographs are normally used for documentation purposes only.
In contrast to aerial photographs, where the optical axis needs
to be near-vertical, the axis is shifted severely (in fact, it can
be close to horizontal). The georectification of such images
for integration in a GIS is rarely pursued. Consequently, it
is not possible to apply any of the standard procedures for
georectification offered by various GIS and remote sensing
packages. With the help of terrestrial photogrammetry ,
 specific procedures need to be developed. Hence, there do
exist a few special programs for the geometric correction of
oblique (aerial and terrestrial) photographs, most of which are
monoplotting procedures. To name a few, the program-system
BUSCH-DHM was applied by Waldhäusl
et al.  in order
to map the quality of vegetation on ski slopes from terrestrial
photographs. Aschenbrenner  analyzed glacier surfaces
derived from terrestrial and aerial images using the package
SCOP-MONOPLOT. Also, Kirnbauer et al.  used digital
monoplotting for examining snow cover patterns from oblique
nonmetric aerial photographs. Warner et al.  developed
a monoscopic measurement system for vertical and oblique
All these techniques mainly support the transformation of
selected image information, namely, identified boundary lines,
and hence do not allow the georectification of the entire image
contentina raster data format.Thisstudyhowever,isfocusedon
developing a method that enables the production of orthophotos
from high oblique terrestrial imagery. Moreover, the analysis of
a time series of images is of interest and therefore, the goal of
the developed method is to support the requirements of such an
analysis as well. As snow distribution is a spatio-temporal phe-
0196–2892/01$10.00 © 2001 IEEE
886 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 4, APRIL 2001
nomenon of crucial importance for high mountain ecosystems
, , snow pattern analysis is chosen in order to test the
developed model with a detailed time series.
The objective of this study is to develop a simple and eco-
nomic method to:
• capture data series of spatial information in mountainous
terrain with any temporal resolution desired by means of
a small-format camera;
• georectify terrestrial, high-oblique photographs for input
to a GIS;
• apply the developed georectification technique on both
single-stage and pre-analyzed time series (“summary”)
TUDY AREA AND MATERIAL
A. Study Area
The study area is situated in the upper Passeier Valley, South
Tyrol, Italy, in the central eastern Alps. It is a narrow, v-shaped
valley with fairly steep slopes. The area of ecological interest
is located on the orographic right, south exposed hill slope,
comprising the alpine meadows of the farmers of the village
Walten and the adjoining forest below. It extends over approx-
imately 3 km
and ranges from 1200 m to 2350 m in altitude.
For details, see .
B. Automatic Camera and Position
As a capturing system, a 35 mm single lens reflex camera,
namely a Pentax A-3 with Takumar 28–80 mm objective, was
used with Kodak Elite Chrome 100 color slide films. To be
able to take photographs automatically, an electronic timer was
added to the camera.
The topographic situation of the study area, as described
above, provided an optimal solution for the positioning of the
camera. On the opposing hill slope of the study site the camera,
covered by an open metal box, was mounted within a forest
clearing at an altitude of 1521 m. From this position, the central
and most important section of the study site was clearly visible.
The focal length was adjusted to 35 mm to cover the entire
area of interest. The horizontal distance to the area of interest
ranged between 1250 m and 2350 m. Only approximately 3%
of the photographed area could not be seen due to the position
of the camera in relation to the study site. This was checked
by a calculated visibility map. To avoid confusion within the
georectification procedure, the corresponding areas needed to
be excluded from the calculations.
The lensaperture and the exposure time were set to automatic
to allow adjustment to varying weather conditions. By means of
the timer one photograph was taken every day at noon (12.00 h
CET). As a result, a very detailed time series of more than four
years has been established up to now. The study itself, however,
focuses on the period between November 1996 and September
C. Preparation of Photographs
For further processing and integration with GIS, the color
slides were transformed into digital form. They were scanned
with a resolution of 1100 dpi (1536
1024 pixels) by a com-
Fig. 1. Projection of the three-dimensional (3-D) DEM onto the
two-dimensional (2-D) image plane.
: ground coordinates of
: image coordinates of the projected DEM.
mercial company and saved in Kodak Photo CD format (.pcd).
This resulted in a file size of approximately 4.5 MB for each
image. The geometric resolution varied somewhat within the
images (0.5–2.5 m) due to scale variations. The nominal pixel
size in the center of the image covered an area of about 1
For practical reasons the images needed to be pre-processed.
The extent of the images was reduced somewhat on all sides to
a final image size of 1349
D. Digital Elevation Model
Working with single images in mountainous terrain required
the use of a high-quality and high-resolution digital elevation
model (DEM) . The DEM was generated from 5 m con-
tour lines and the incorporation of auxiliary line and point data
(e.g., ridges, edges, river network, peaks, and depressions) by
applying a sophisticated linear interpolation technique, namely,
the Stochastic Contour Interpolation Method (SCIM) . As
a result, a DEM with a spatial resolution of 5 m and a sub-
pixel accuracy was available. The DEM was georeferenced to
the Italian grid system (Gauß–Boaga projection, International
Spheroid oriented to Rome M. Mario).
A. JUKE Method
differs significantly from the standard monoplotting procedure.
In monoplotting, the image rays from the camera of the identi-
fiedboundary lines areintersected pointby point with thedigital
elevationmodel to find the
-coordinatesof thelines. In the
JUKE method, however, the DEM is projected onto the image
plane (Fig. 1), which enables the relation of the ground coor-
dinate system (
) with the image coordinate system ( ).
This is an efficient and fast technique to obtain the information
pixel by pixel.
Monoplotting has essential disadvantages.
• displacement of objects due to misinterpretation when
identifying boundary lines in the photographs, in partic-
ular when the distance between the object and the camera
position is large;
• displacement of objects due to small intersection angles
between the image rays and the DEM, in particular for
• additional inaccuracies caused by linear interpolation be-
tween the intersected points.
ASCHENWALD et al.: SPATIO-TEMPORAL LANDSCAPE ANALYSIS 887
Fig. 2. (a) Digital photo. (b) Projected DEM. (c) Both data superimposed.
Applyingthe JUKE method, these inaccuraciestypical forthe
monoplotting technology do not occur. The intrinsic character-
istic of the JUKE method, namely to project the DEM onto the
image plane, guarantees a definite relation of the points(despite
nonlinear effects). Once the transformation parameters (inner
and outer orientation) are determined, all points are projected
by only a few calculation steps. Therefore, it is also a computa-
tionally efficient method. Furthermore with the help of ground
control points (GCPs) the accuracy of the JUKE method (rela-
tive to the accuracyof the DEM and the GCPs) can be specified.
The necessary mathematical equations and derivations are to
be found in specific photogrammetry text books , .
B. Input Information
For the calculations, the coordinates of various reference
points and GCPs were necessary as input for the JUKE method.
The exact ground position of the camera was measured by
means of differential GPS surveying (spatial accuracy:
As a reference receiver, a Trimble 4000 (L1) was available, and
as mobile receiver, a Trimble Pro XL was used.
The ground location of the perspective center of the images
was needed to identify the direction of the optical axis of the
lens system (outer orientation). Normally, the optical axis
should have been horizontal, but a horizontal orientation could
not be guaranteed in practical operation because of some
technical hitches in mounting the camera. The respective center
point of the images was established by connecting the opposing
corners of the images on the screen with very fine lines (inner
orientation). The corresponding ground coordinates were
determined by defining their position on an already available
orthophoto of the study area (scale 1 : 5.000, date of flight
August 23, 1997) . With this information, the location and
orientation of the respective image plane with regard to the
DEM were obtained.
Some of the essential mathematical computations could only
be carried out with the help of several GCPs. Features that were
clearly identifiable in the image were selected, and their image
coordinates were ascertained. Their ground position was either
located with the highest possible precision on the available
orthophoto or measured by means of GPS. Image as well as
ground coordinates were obtained for 15 GCPs.
For each pixel corner of the DEM, a unique identification
number (ID) as well as the ground coordinates (
calculated and written to an ASCII-file. Knowing the exact po-
sition of the camera, those pixels that were not visible from the
camera location were calculated and marked on the DEM. All
pixel corners covered by this masked area were then deleted
from the ASCII-file.
By means of the JUKE method and the aforementioned input
information, the necessary parameters were calculated. Conse-
quently, all points of the DEM prepared in the ASCII-file were
reflected, shifted, projected, rotated, and multiplied with the
enlargement scaling factor to get the corresponding points on
the image plane. The output ASCII-file now contained three
columns: the identification number and the
- and -coordi-
nates of the pixel corners of the DEM expressed in the coor-
dinate system of the image.
The resulting point information of the new ASCII-file (pro-
jected DEM) was then converted into an ArcInfo (Workstation
7.2) raster file (grid) with the same extent and pixel size as the
prepared image. It was now possible to superimpose the two
grids (projected DEM and image) spatially and combine their
information (Fig. 2). The digital number of each pixel of the
image, which represented the relevant radiometric information,
was appended to the pixel information (the ID number) of the
projected DEM. These two columns were then written into a
third ASCII-file. By means of the ID number, the ground coor-
dinates (i.e., the
, and coordinates of each pixel corner)
were still known. Consequently, the radiometric information of
the image could be depicted in the national map projection and
ground coordinate system. To provide an adjusted value for the
entire pixel, the values of the four surrounding pixel corners
were averaged. The image was now available for overlay with
other spatial data and for incorporation into landscape models.
Forthis study,thegeorectificationwas conducted with single-
stage images as well as with pre-analyzed “summary” images
(amount of days covered with snow) as the analysis of time se-
ries was focused on.
D. Quality of the JUKE Method
The mean Euclidean distance wasused to evaluatethe quality
of the JUKE method. This measure is defined as the mean de-
viation of the calculated points from their known position, in
contrast to the root mean square error (RMSE), which measures
the root of the mean squared deviation.
The set of GCPs was divided randomly into two disjunct sets,
one of which wasthe training set and the other one of which was
the validation set. The training set was used exclusively to com-
pute the orientation of the optical axis. The points of the vali-
dation set were employed to calculate the mean Euclidean dis-
tance. Because of the random choice of these two sets, the mea-
sure could also be biased in both directions and thus, the mean
Euclidean distance on the validation set was not a consistent
estimator for the model performance. Therefore, the bootstrap
888 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 4, APRIL 2001
Fig. 3. Summary of the model structure. Gray input data; surrounded by broken line
procedure; and black statistical evaluation.
method was applied to achieve statistically correct statements
about the performance of the JUKE method and its standard de-
viation . The set of reference points was not only divided
once into two sets but
times. The procedure of calculating
the parameters of the JUKE method with the help of the training
points and subsequently the corresponding performance on the
validation set (mean Euclidean distance) was repeated
The average over all these performances gave a consistent esti-
mator of the quality of the method and a consistent estimator of
the standard error could also be achieved.
E. Practical Realization
The model was programmed in Visual Basic 6.0 and can
be downloaded from the homepage with the internet address
requirement is a 486 computer with Windows NT or 9
The program requires the following inputs:
• camera position;
• center point of the photo;
• focal length;
• text file of the data (DEM);
• text file of reference and GCPs;
• number of bootstrap replications (optional);
• ratio of the number of training points to the number of
validationpoints (at themoment,this ratio isfixed to 2: 1).
The ground coordinates of the center point of the images,
which was the intersection point of the image diagonals, was
very difficult to measure. On the ground, this intersection point
was located within the forest and it was thus impossible to re-
late it to a specific image object. Consequently, the ground co-
ordinates of the image center point could not be ascertained
with confidence. To obtain still quite good results, a raster with
a cell size of 1 m was put around the supposed center in the
-, -, and -directions in the DEM, each raster point repre-
senting a possible center. With each possible center the model
was calculated 100 times and the one with the smallest mean
Euclidean distance on the validation set was selected as the best
center. This procedure guaranteed a good performance even if
the center was not clearly identifiable.
As a summary, Fig. 3 demonstrates graphically the structure
of the entire procedure.
F. Method for Time Series
The focus of this project was not only to analyze single-stage
images but a time series of photographs. Therefore, a useful
procedure for automation had to be developed.
In this section, two methods of time-series analysis are de-
scribed. On the one hand each single image could be georec-
tified separately by means of the above introduced procedure.
The thematic analysis would then be conducted afterwards.
On the other hand the images of a specific time period could
first be added up pixelwise. Consequently, only the resulting
summary image needed to be georectified. The latter procedure
offered the great advantage of identifying and measuring the
reference and control points just once for the reference image.
Hence, a considerable amount of time could be saved this way.
Due to temperature changes leading to subtle movements of
the metal container, which was used for camera suspension, a
variation of the image extent for the various photographs was
expected. Therefore, it was not possible to add up all single im-
ages without making arrangements. Applying the second anal-
ysis procedure, relative image registration was a prerequisite
as image analysis was carried out before the georectification
procedure. To provide the same properties for all images of a
continuous month series, a relative registration was carried out,
adapting the extent and resolution of each image to one specific
reference image, i.e., a warping was performed. By means of
four control points identifiable in all images, the co-registration
was carried out. It was now possible to superimpose all these
Images which could not be used due to bad weather (fog,
clouds, dust) were taken off the time series and substituted with
images from the day before or after which ever fitted best. Con-
ASCHENWALD et al.: SPATIO-TEMPORAL LANDSCAPE ANALYSIS 889
Fig. 4. Locations of 15 GCPs G1–G15 and nine test points T1–T9. The broken line characterizes the used image extent.
sequently, all zones without significance for this study were cut
offand masked out,respectively.Hence, the skywas masked out
defining a threshold using a clear summer photograph. A clear
winter photograph was used for masking out the forest. The ra-
diometric information of thepixel in eachimage was then trans-
ferred into binary data by defining a threshold, “0” standing for
Consequently, all images of a specific snow period were super-
imposed and the digital numbers (0 or 1) were added up pixel-
wise, resulting in an image representing the cumulative score of
days with snow cover at each cell. Finally, the model was calcu-
lated for the reference image only and was then superimposed
with the “summary” image.
This technique would represent a comfortable automation
procedure for time series analysis, but first, the quality of this
procedure needed to be checked. Thus, nine test points were
selected, all of which were not covered with snow during the
whole period. The pixel representation of these selected objects
in the summary image and their appearance in the single images
Two main questions are addressed in this section.
• How suitable was the model for georectification of single-
• Was the described method successful and useful for time-
A. Georectification: Single Image
A single image was prepared and the DEM was calculated
correspondingly by means of the developed JUKE method. Fi-
nally, both data sets were superimposed to obtain the informa-
tion of interest (Fig. 2). Fifteen GCPs were available to com-
pute the parameters ofthe model. Great attentionwas paid to the
fact that these points were distributed approximately uniformly
throughout the area (Fig. 4). Theywere randomly separated into
ten points for the training set and five points for the validation
set. The more points included in the training set, the better the
accuracy of the model will be.
As an example, an image taken January 11, 1997 was geo-
rectified. For a quantitative judgement, the mean Euclidean dis-
tance on the validation set was calculated to be 2.78 pixels.
According to the different resolutions of the image, caused by
oblique view and topography, this result represented an average
deviation of approximately 1.4 m in the lower, 2.8 m in the
middle, and 7 m in the upper part of the image. With regard
to the final resolution, which was given by the resolution of the
DEM(5m),the mean Euclidean distance correspondedto a high
This remarkable result was also supported by repeating the
calculation 100 times. The mean performance amounted to 2.32
pixels with a standard error of 0.39pixels. These results empha-
sized the quality of the JUKE method.
Furthermore, it was checked if the applied technique of
finding the best center by putting a raster around the supposed
center could improve the performance. The raster width was
-, -, and -direction, and the lateral length of the
entire raster was 100 m. The mean performance could be
improved only slightly. But this fact was not surprising because
of the goodness of the previous results. With the optimized
center, the mean performance amounted to 2.14 pixels with a
standard error of 0.41 pixels.
Furthermore, the quality of the model was checked by ap-
plying the methodto eightsingle images from different filmsse-
lected randomly. The image coordinates ofall GCPswere ascer-
tained for each image separately and a raster as described above
was put aroundthe supposed center point of each photo. The se-
lected photos showedclearly the variation of the center (Table I)
andconsequentlyof the optical axisand the GCPs.Ameandevi-
ation of 1.6 pixels in
-direction and approximately 1.5 pixels
890 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 4, APRIL 2001
EST CENTER COORDINATES ( ), MEAN PERFORMANCE ( ), AND STANDARD DEVIATION (S. D.) OF JUKE METHOD FOR
DIFFERENT SINGLE PHOTOGRAPHS
Fig. 5. The “summary” image representing the cumulative score of days with snow cover at each cell (a) before and (b) after the georectification procedure.
in -direction were measured after georectifying the eight se-
lected images separately. As the performance results demon-
strated, this fact was not a problem for single image georecti-
fication because of the flexibility of the model that calculated
the optical axis and the optimal center for each. But it was nec-
essary to identify the image coordinates of the GCPs for each
photograph separately. Therefore, time-series analysis by geo-
rectifyingeachsingle imageof a specificperiod wasshowntobe
possible, but of course this would be a time consuming method.
The slight instability of camera adjustment could be a disad-
vantage for time-series analysis using the second method. But
the following section demonstrates theusefulness of the second,
automated, and less time intensive procedure.
B. Georectification: Time-Series
For all images of a continuous seven month series, a rela-
tive registration was carried out, adapting the extent and reso-
lution of each image to the specific reference image dated from
January 11, 1997. This was done with the help of four control
points identifiable in all images. Consequently, 204 images of
the snow period 1997/98 were superimposed, and the digital
numbers (0 or 1) were added up pixelwise, resulting in an image
representing the cumulative score of days with snow cover at
each cell [Fig. 5(a)].
Finally, the JUKE method was calculated for the reference
image (January 11, 1997) only. The grid of the projected
DEM was then superimposed with the summary image and
consequently retransformed into the national coordinate system
Forcheckingthe quality of the co-registrationprocedure,nine
smallobjects (e.g., trees,rocks) that werenot coveredwith snow
during the whole period were selected to serve as test points
(Fig. 6). The pixel values representing these objects in the sum-
mary image were compared to the desired pixel values, which
ASCHENWALD et al.: SPATIO-TEMPORAL LANDSCAPE ANALYSIS 891
Fig. 6. (a) Position of test point T . (b) Magnified illustration of the snow situation around test point T . The pixel values in the summary image (actual values)
differ from the desired pixel values. Instead of the ideal value of zero at the location of T
, the actual values were 105 and 88, i.e., a minimum of 88 images
out of a total of 163 images were positioned slightly incorrectly. (c) Actual and desired snow distribution around test point T
represented diagrammatically. (d)
Superimposed information of the desired and the actual snow distribution around test point T
. Test point T and the mean of all nine test points represented
were deduced from the single images. As an example, consid-
ering only one row of pixels, test point T
[Fig. 6(a)] should be
depicted ideally by two pixels side by side, each with the value
0 (desired values), the adjoining pixels should have values of
163 [Fig. 6(b)].Actually, however,test pointT
with values of 105 and 88 in the summary image (actual values).
The values of the adjoining pixels increased from 124 up to 163
[Fig. 6(b)]. This showed that a minimum of 88 single images
out of a total of 163 images were not positioned exactly. This
resulted in the objects being somewhat blurred in the summary
image with a measured positioning error of four pixels to the
right (east) and of two pixels to the left (west) [Fig. 6(c)].
This procedure was repeated for all nine test points. On av-
erage, the objects were expanded incorrectly two pixels (1–5
m) to the left side and five pixels (2.5–12.5 m) to the right side
[Fig. 6(d)]. This gave evidence that thesummary image wasdis-
torted somewhat to the right side of the photograph. It was thus
inevitable to introduce some simplifications, but the developed
procedure was still robust enough to yield useful results.
As we have shown, the JUKE method is useful for single
image georectification. One sensitive parameter of the model is
the identification of the ground coordinates of the center point
of the images. Another is the determination of the coordinates
of GCPs. By means of an optimization routine, the respective
center can be slightly optimized.
Table I demonstrates that the performance of the model is dif-
ferent depending on the image. Since the center point is opti-
mized for each image and the optical axis is calculated every
time, the range of the performance is not due to variations of
the center point. Rather, it is caused by the difficult identifica-
tion of the GCPs.
Hence, accuracy is mainly a measure of the quality of the
GCPs. In the investigated area, ground control is somewhat dif-
ficult to define, as it is situated in a rather remote region ,
. At lower altitudes, there are several buildings that can be
used but at higher altitudes, reliable GCPs are rare. There are
only a few elements, such as huts and path intersections, that
can be identified on the photographs. Moreover, image scale be-
comes smaller toward higher altitudes and thus small errors in
determining the image coordinates of the GCPs result in more
serious positioning errors. It waspossible to measure the ground
coordinates of a few GCPs by means of a differential GPS, thus
providing high quality GCPs. Others were defined by means of
an available orthophoto of the area. The difficulty of defining
reliable GCPs could be facilitated considerably by placing large
signals at the respective position in the terrain in advance .
The ground coordinates could then be measured with high ac-
curacy. The GCPs would be clearly identifiable in all digital
photos, and their image coordinates could be ascertained with
highest possible confidence.With such preparations, the perfor-
mance of the JUKE method would improve notably.
The accuracy of the summary image is additionally strongly
influenced by the co-registration procedure. As some posi-
tioning errors are already introduced during this process, the
accuracy of the resulting georectified image is somewhat lower
than the mean accuracy of several images which are georecti-
fied separately. Employing precise, predefined GCPs also for
the co-registration procedure would improve the accuracy of
the summary image as well.
Another critical point is the rather small number of GCPs.
Only 15 GCPs are available, which have to be split into training
and validation sets to be able to obtain consistent estimators
892 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 39, NO. 4, APRIL 2001
for the performance of the model. Generally, the more training
points available the higher the possibility to fit the model per-
fectly. Fifteen reference points are a minimum requirement but
nevertheless the results are promising.
Because of the accuracy of our model, we did not consider
any calculations for correcting aberrations (for example barrel
or pincushion distortion) which are caused by the optical lens
system. This might be due to the fact that only the inner part
of the images was analyzed. As an advantage, the study area
is situated in the middle of the images where these nonlinear
aberrations are negligible.
However, considering the final resolution of 5 m, the mean
deviation in the middle of the images of about 3 m is within
one pixel and thus is satisfying. This is in particular true for the
fieldof application of this study since snowcoverboundaries are
desired to an accuracy of a few meters. Applying monoplotting
procedures, Aschenbrenner  achieved a positional accuracy
of 5 to 10 m, Kirnbauer et al.  estimated inaccuracies in the
order of 5 m.
For each application concerning time-series analysis, one
needs to weigh up if the georectification procedure has to be
carried out separately for each image thus being time-con-
suming but providing higher accuracy or if a somewhat less
accurate pre-analyzed image is sufficient as input in the georec-
tification procedure and provides a considerable saving of time.
Up to now, the model performance is computed in units of
pixels. This has not been an imperative problem but a compu-
tational effort which, after we have shown the practicability of
our model, will be solved in a future project.
The spatial resolution of the final product mainly depends on
three parameters: a) the scale and resolution of the original pho-
tograph, b) the scanning resolution, and c) the pixel size of the
DEM. In high oblique photographs, scale variation is severe and
needs to be considered. It is not only a function of relief to-
pography (ground elevation) as it is for true vertical aerial pho-
tographs,butit is alsoaffectedby object distance aswellasposi-
tion and angular orientation of the camera . Hence, scale is
different in all directions. Scale is largest at the bottom of the
images, becoming increasingly smaller toward the top of the
image. This phenomenon is closely related to pixel resolution
of the scanned images. Due to scale variation each pixel covers
a different ground area. Pixels are represented at about the fol-
lowing resolutions: bottom 0.5 m, center 1 m, upper part 2.5 m.
Thechosen image resolution of1536
mise between high detail recognition and storage space. How-
ever, detail recognition is still sufficient for the purpose of this
study. The DEM is represented with a pixelsize of 5 m, i.e., with
anotablycoarser spatial resolution thanthe scanned images,and
consequently, this is the determining factor for the final resolu-
tion. Depending on the respective image scale, every second to
fifth image pixel information only is taken when combining the
projected DEM with the image information. At the top of the
image, where scale is the smallest, the projected points of the
DEM are distributed much closer than throughout the bottom
part. Perspective views influence the spreading of the projected
DEM points. A better final resolution can thus be reached by
improving the spatial resolution of the DEM, and adjusting it to
the image resolution determined by the other two factors men-
tioned above.However,as a higher input resolution would result
in a considerably larger data file and consequently would lead
to a notably increased calculation time, this possibility must be
Working with high oblique images implies dealing with non-
visible areas. For the JUKE method, these zones do not in-
duce any problems as long as the corresponding pixel corners
are excluded from the data file before calculation. However, as
these areas are lost for analysis it is advisable to choose the
study area and the camera location in a way which avoids large
hidden zones. Hence, study areas with highly variable exposi-
tion should not be favored. In this study only about 3% of the
investigated area was not visible. Although no snow cover in-
formation is available for these hidden zones, it is possible to
predict reliable snow distribution values also for these pixels by
means of a multiple regression analysis .
To apply the presented method, no experience in photogram-
metry is required. Beside having available the JUKE method
(software), it is necessary to have access to a raster-GIS in order
to combine the model-output file with the digital photograph
and for visualization. All data transfer is done by the convenient
ASCII format. Data preparation can thus be performed by sev-
eral, commonly available programs. Hence, the method is fea-
sible for anybody with general computing knowledge and some
basic training in GIS and digital image processing.
The developed JUKE method proved successful for single-
stage as well as for time series analysis. Although it is not pos-
sible to reach photogrammetric precision, the georectified data
can be used in a GIS-environment for investigations at scales of
about 1:5000 to 1:10 000. Single-stage images were georecti-
fiedwithhighlysatisfactory accuracy. Including an optimization
procedurefor definingthe imagecenter yielded evena slight im-
provement. For time series analysis, one has to accept a some-
whatloweraccuracy,as additional positioning errors arealready
introduced by the image co-registration procedure. The higher
the quality and the precision of the input data, i.e., the DEM
and the GCPs, the higher the accuracy of the final product. As
high quality DEMs and precisely measured GCPs become more
readily available, these limitations are likely to be made up in
Particularly in high mountainous terrain, where it can be dif-
ficult and expensive to acquire field data, aerial photographs or
satellite images, this method can be of considerable interest and
benefit to applied field studies in order to gain spatio-temporal
data. The relatively economic data capturing and transforming
procedure enables a successful application of the method within
The authors would like to thank the input of the anonymous
ASCHENWALD et al.: SPATIO-TEMPORAL LANDSCAPE ANALYSIS 893
 A.Cernusca, U. Tappeiner,and N. Bayfield, Land Use Changes in Euro-
pean Mountain Ecosystems. ECOMONT: Concepts—Results. Berlin,
Germany: Blackwell, 1999, p. 368.
 D. Pan, G. Domon, S. de Blois, and A. Bouchard, “Temporal
(1958–1993) and spatial patterns of land use changes in Haut-Saint-Lau-
rent (Quebec, Canada) and their relation to landscape physical at-
tributes,” Landscape Ecol., vol. 14, pp. 35–52, 1999.
 M. J. Wooster and D. A. Rothery, “Time-series analysis of effusive vol-
canic activity using the ERS along track scanning radiometer: The 1995
eruption of Fernandina Volcano, Galapagos Islands,” Remote Sens. En-
viron., vol. 62, no. 1, pp. 109–117, 1997.
 P.R. Paul, V. S. Kumar,C. L. V. R.Rao, K. Seidel, and H. Haefner, “Op-
erational remote sensing for snowcover, runoff and avalanche applica-
tions in the Himalayas,” in Proc. Operationalization of Remote Sensing
in Developing Countries, vol. 4, Enschede, The Netherlands, 1993, pp.
 P. Meyer, “A parametric approach for the geocoding of Airborne vis-
ible/infrared imaging spectrometer (AVIRIS) data in rugged terrain,”
Remote Sens. Environ., vol. 49, pp. 118–130, 1994.
 P. V. Bolstad, “Geometric errors in natural resource GIS data: Tilt
and terrain effects in aerial photographs,” Forest Science, vol. 38, pp.
367–380, April 1992.
 G. Hildebrandt, Fernerkundung und Luftbildmessung für
Forstwirtschaft, Vegetationskartierung und Landschaft-
sökologie. Heidelberg, Germany: Herbert Wichmann Verlag, 1996,
 C. C. Slama, Manual of Photogrammetry, 4th ed. Washington, DC:
Amer. Soc. Photogramm., 1980.
 P. Waldhäusl, C. Enthofer, and H. Kager, “Bildauswertung durch
Verschneiden des Strahlenbündels mit dem digitalen Höhenmodell,”
ÖzfVuPh, vol. 3, pp. 155–166, 1986.
 J. Aschenbrenner, “Orthophoto und monoplotting in der gletscherkar-
tographie,” Salzburger Geograph. Arbeiten, vol. 21, pp. 54–77, 1992.
 R. Kirnbauer, G. Blöschl, P. Waldhäusl, and F. Hochstöger, “An anal-
ysis of snow cover patterns as derivedfrom oblique aerial photographs,”
in Proc. Vienna Symp.—Snow, Hydrology and Forests in High Alpine
Areas, vol. 205, Vienna, Austria, 1991, pp. 91–99.
 W. S. Warner and W. Carson, “Mapping monoscopically: An innovative
system for vertical and oblique aerial photographs,” ITC J., vol. 3, pp.
 J. C. Coughlan and S. W. Running, “Regional ecosystem simulation:
A general model for simulating snow accumulation and melt in moun-
tainous terrain,” Landscape Ecol., vol. 12, pp. 119–136, 1997.
 T. R. Allen and S. J. Walsh, “Characterizing multitemporal Alpine
snowmelt patterns for ecological inferences,” Photogramm. Eng.
Remote Sensing, vol. 59, pp. 1521–1529, Oct. 1993.
 U. Tappeiner, E. Tasser, V. Mayr, and B. Ostendorf, “Passeier Valley
composite landscape, Italy,” in Land Use Changes in European
Mountain Ecosystems. ECOMONT: Concepts—Results, A. Cernusca,
U. Tappeiner, and N. Bayfield, Eds. Berlin, Germany: Blackwell,
1999, pp. 46–59.
 K. I. Itten and P. Meyer, “Geometric and radiometric correction of
TM data of mountainous forested areas,” IEEE Trans. Geosci. Remote
Sensing, vol. 31, pp. 764–770, July 1993.
 B. Ostendorf, V. Mayr, and U. Tappeiner, “GIS—Contents and goals,”
in Land Use Changes in European Mountain Ecosystems. ECOMONT:
Concepts—Results, A. Cernusca, U. Tappeiner, and N. Bayfield,
Eds. Berlin, Germany: Blackwell, 1999, pp. 180–187.
 E. Tasser, U. Tappeiner,andA.Cernusca,“SüdtirolsAlmen im Wandel,”
Athesia, p. 267, 2001, to be published.
 J. Johnston and J. Dinardo, Econometric Methods. Singapore: Mc-
 M. F. Buchroithner and R. Fischer, “High-quality 1: 25000 and
1: 50000 satellite image maps for high-mountainous terrain,” in 25th
Int. Symp. Remote Sensing and Global Environmental Change, Graz,
 A. Almer, J. Raggam, and D. Strobl, “High-precision geocoding of re-
mote sensing data of high-relief terrain,” in Symp. High-Mountain Re-
mote Sensing Cartography, Schladming, Austria, 1990.
 U. Tappeiner, G. Tappeiner, J. Aschenwald, E. Tasser, and B. Ostendorf,
“Modelingof Snow-CoverDurationin an Alpine Region,”in Ecol. Mod-
elling, to be published.
JanetteAschenwald received the M.Sc. degree inphysics and the Ph.D. degree
in economics from theUniversity of Innsbruck, Innsbruck, Austria, in 1993 and
She is currently an Assistant Professor with the Department of Statistics, So-
cial and Economic Faculty, University of Innsbruck. Her research interests are
concentrated on statistical procedures and the fields of application of artificial
Karin Leichter received the M.Sc. degree in geography from the University of
Innsbruck, Innsbruck, Austria, in 1996, and the M.Sc. in environmental remote
sensing from the University of Aberdeen, Aberdeen, U.K., in 1997.
She is a Researcher with the European Academy Bolzano, Bolzano, Italy,
where she is responsible for GIS work. Her research interests are focused on
GIS and remote sensing applications in mountain areas in relation to ecological
and environmental aspects.
Erich Tasser received the M.Sc. and Ph.D. in terrestrial ecology from the Uni-
versity of Innsbruck, Innsbruck, Austria, in 1994 and 2000, respectively.
He is currently a Researcher with the European Academy Bolzano, Bolzano,
Italy. His main research interests are in the field of mountain agriculture, sus-
tainable development in mountain areas, and landscape planning. He focuses
methodically on GIS and statistical procedures.
Ulrike Tappeiner received the Ph.D. degree in biology from the University of
Innsbruck, Innsbruck, Austria, in 1985.
She is qualified as a University Lecturer (habilitation) in the field of ecology
with the University of Innsbruck and is the Head of the research area “Alpine
Environment” with the European Academy Bolzano, Bolzano, Italy. She has a
great deal of experience in ecological research of mountain environments at a
European scale. Her main interests are inthe field of applied research with focus
on ecological effects of land-use changes in mountain ecosystems, ecology of
ski runs, and environmental impact assessment and environmental accounting.