You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

International cooperation and technology transfer
Mussio, Luigi

Gianfranco FORLANI
Department of Civil Engineering - University of Parma
Viale delle Scienze, 43100 Parma
E-mail: forlani@parmal.eng.unipr.it
Department of Surveying (DIIAR) - Politecnico of Milan
P.zza L. da Vinci 32, 20133 Milano
E-mail: livio@mail.polimi.it
Commission VI - Working Group 3
KEY WORDS: Digital Elevation Models, Orthophotos, Accuracy
This paper reports on the application of automatic generation of DEMs in rock quarries, to figure out the excavation
volumes by repeated surveys. Accuracies of some percentage of the volume’s estimate are acceptable, the ultimate goal
being planning and managing the exploitation of the quarries. The aim of the investigation is therefore to check whether
automatic DEM generation satisfy this accuracy level and what kind of restrictions should be put on flight planning,
image resolution, scanning accuracy, additional terrain information, matching techniques to improve the performance of
the DPW in this environment. A stereoscopic model from a 1997 flight over a quarry near Brescia has been processed
with two different digital photogrammetric stations and the results compared to those obtained with an analytical plotter
as well as by computing DEM errors from the left and right orthophotos of the pair.
The automatic generation of digital elevation models
by digital photogrammetric techniques is now several
years old. Virtually all phogrammetric workstations
include such modules and DEM’s are routinely
generated in production environment. Tests on the
accuracy of the technique as well as workshops and
meetings on the subject, though, witness that the
current level of sophistication of the correlation
techniques cannot match the results of analytical
plotters in every condition.
Several parameters are relevant trying to asses the
behaviour of the algorithms: image resolution, image
texture, terrain roughness and slope, vegetation or other
terrain superstructures (buildings, vehicles, etc.). As far
as image resolution is concerned, empirical tests
suggest that this parameter is almost independent of
image scale (Baltsavias and Kaeser, 1998) (which is
obviously kept at the smallest value for the specific
task) and that it may be chosen in the range from 20 to
30 micrometers with just a slight quality decay as
resolution decreases. Using larger resolution values,
the DEM quality becomes too poor.
Overall DEM accuracy, defined as the RMS of the
discrepancies between terrain heights and its
interpolated values from the DEM, may be
characterized by three factors (Ackermann, 1996):
• the accuracy in elevation o z = m b h/B G p which
depends on image scale, object to baselenght ratio
and measurement accuracy in image space;
• DTM grid size: interpolation errors varies with
terrain roughness: the denser the grid the smaller
the error, for a given terrain;
• detection and filtering of non-terrain features
(buildings, trees, etc.).
The accuracy of correlation techniques is normally
evaluated in percentage of the pixel size. Therefore,
increasing scanning resolution may seem the
straightforward way to improve it. In fact,
notwithstanding storage and data handling problems,
still severe if pixel size goes below 20 pm, empirical
studies suggest that the current generation of
photogrammetric scanners is affected by significant
noise when scanning images at the maximum available
optical resolution (below 10 pm). Perhaps more
important than resolution is image texture:
homogeneous regions or repetitive patterns may lead to
wrong identification of conjugate pairs, leading to
inaccurate results even if the terrain is not too rough.
Steep slopes and terrain breaklines, on the other end,
lead to strong perspective differences and occlusions
that stereo correlation algorithms cannot cope with. As
image scale decreases, their effect is less relevant; the
same apply also to non-terrain features.
At large scales, filtering out trees, buildings and so on
is automatically performed by some systems, based on
appropriate smoothing constraints on the curvature of
the interpolating surface or other techniques. This
works fine as far as these objects are isolated and there
are no terrain breaklines in the vicinity. Forest areas
and urban areas are still hard to deal with in large scale.