Page 231 - Digital Analysis of Remotely Sensed Imagery

Image Geometric Rectification 195

The accuracy of determining the position and orientation of the
sensor using direct georeferencing lies typically between 10 and
20 cm (RMS), and 15 and 30 arc seconds (RMS), respectively (Litho-
poulos et al., 1999). This level of accuracy has been confirmed by Kinn
(2002). Evaluated against 24 GCPs, a ground positioning accuracy of
around 1 m was achieved in both easting and northing (Mostafa and
Schwarz, 2000). The accuracy for height is lower, at a range from 1.5 to
about 3 m. Horizontal coordinates and height have a standard devia-
tion of 0.9 and 1.8 m, respectively. The ground positioning accuracy
was further improved to 0.5 m in planimetry (1s) and 1.6 m in height
from stereopairs at an average image scale of 1:12,000 (Mostafa and
Schwarz, 2001). The highest ground positioning accuracy is achieved
at 0.2 m of planimetric accuracy and 0.3 m in height using GPS/INS-
aided block triangulation of both nadir and oblique images. This
accuracy level is sufficiently accurate for mappings at scales <1:5000.

5.8.2 Comparison with Polynomial Model
As shown in Eqs. (5.13) and (5.14), polynomial-based image transfor-
mation handles only the horizontal position of pixels. It is unable to
deal with the third dimension of pixels (i.e., height). Therefore, any
positional shift caused by topographic relief on the ground cannot be
removed via the application of the two transformation models. The
polynomial method is suitable for georeferencing satellite images
obtained at a very small scale. If ground control is available, this
method is preferable in transforming images from the local coordi-
nate system to the global system. In this way it is feasible not only to
coregister multiple images but also to correct geometric distortions
inherent in the input image. Very easy to implement, polynomial
image rectification is advantageous over direct georeferencing in that
it does not require information on satellite orbit and sensor calibra-
tion (El-Manadili and Novak, 1996). However, it has the following
three disadvantages:
• First, it requires maintaining a large number of well-distributed
GCPs.
• Second, it lacks physical interpretation of the model beyond
the second order.
• Third, it is unable to handle the positional shift caused by
topographic relief displacement. Of the two sources of
geometric distortions in remote sensing images, those caused
by orbital parameters exert a global impact on all pixels in an
image. With the polynomial model, such distortions can be
effectively dealt with. However, those caused by variation in
topography are impossible to address.

The influence of topographic relief is local and random in nature.
This influence can be effectively tackled by a stochastic approach of