1 2004
ided
good
‚the
ithm
ning
itch)
and
on to
ated
the
ven
ocal
r to
that
area
1 0.3
dest
nely
rain.
flat
the
cant.
TOrs
1 js
ling
und
led)
na
ures
vey
due
such
nes.
tof
dary
ling
s up
the
lon.
tral
a of
ying
rof
SD
It. is
vide
ead
ince
ctor
the
500
n of
) or
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004
|
|
|
| 1/1,200 (rural areas) for the old version. The new version of
| cadastral maps in Taiwan is reliable and is employed in this
| paper. It is estimated that an error, resulted from the precision
of digitizers and the process in digitization, up to 0.3 mm is
possible, which leads to an error of 15-30cm as a result.
Obviously, the error of digitisation of cadastral maps is still
less than one pixel in Quickbird imagery. To sum up, the error
model based upon equation (2) and the above inference gives
an error of 6 pixels due to various causes. The error model also
implicitly indicates that using wall features for the registration
of Quickbird images and vector data is liable to result in errors
up to 6 pixels or 4.2m.
3.1 Test Image and Cadastral Map
À sub-scene of a panchromatic standard (rectified) Quickbird
image taken on 26^ May, in 2002, in Taoyuan County, Taiwan,
covering an area of 2.5x1.6 km”, as shown in Fig.3, where
undulation of terrain surface is under 5m in the test area, is to
be registered with a cadastral map conveying several parcels
3. TEST DATA
with wall features, as shown in Fig.4.
Figure 3. A sub-scene of Quickbird standard image over
Taoyuan, Taiwan. Size:2.5x1.6 km’.
(Courtesy by Taoyuan County Government, Taiwan, ROC)
Figure 4. Four polygons extracted from a cadastral map, to be
registered with the satellite image.
The geometric consistency of the standard image can be
validated using a two-dimensional conformal (four-parameter)
transformation, which describes the geometric conformation
between the two sets of co-ordinates in terms of their
geometric relationship, such as a scaling factor, a rotational
angle and two translations, for GPS-measured ground points
and manually measured co-ordinates on the satellite image.
The header data of a satellite image gives a scale of the
image pixel to actual ground distance, and the co-ordinates
of the image points are recorded and transformed into a
local co-ordinate system. Since the extent of the test arcas
is relatively small, Datum shifts are of no concern in
validating the test images. Table 1 gives the results of
geometric inconsistency check on the two sets of measured
co-ordinates. The geometric inconsistency or the precision
of rectification of the test image, in terms of 6 GPS-
measured points between two sets of co-ordinates derived
by using four parameters transformation, shows that the
root of mean squared error (RMSE) of the 6 points is
approximately 2 pixels (6; 7 2 pixels), as demonstrated in
section 2.3. This is obviously owing to a relatively flat
terrain surface in the test areas.
5 GES sama RMSE| Mean | Max. | Min. |Extent
points
Easting 1.6 0 +2.1 | -2.3 | 44
Northing 1.2 0 *tLil-k44 2.5
Error Vector Length| 2.0 (Unit: pixels)
Table 1. The evaluation of the geometric consistency of
the sub-scene Quickbird standard image.
3.2 Manual Image-and-Map Registration
Automation is aimed at replacing the role of human
operators in some sorts of process or systems. Thus, human
knowledge about the principles of the photogrammetric
processes is essential for the evaluation of the desired
automatic process. In order to validate the proposed error
model, 22 ‘control features’ (nodes of polygons of cadastral
parcels) are selected manually from the test image and
vector map. Among those points (‘control features’), 8
points are used as GCPs in a projective transformation
between the test image and the vector map, and the other
14 points as check points. The best estimations of the
parameters of each transformation function are derived
using a least squares adjustment procedure, and these
parameters formulate a transformation function allowing
the other 14 points to be transformed and compared. Figure
5 shows the residual error vectors of 8 GCPs using a
projective transformation between the co-ordinates
manually measured in image and on a vector map,
respectively. The results summarized as in Table 2 show
the accuracies of 14 check points using a projective
transformation between the co-ordinates measured in the
test image and on the vector map. Figure 6 demonstrates
the distribution of the resultant errors of the check points,
showig that there is no systematic error. The RMS errors
shown in Table 2 give magnitudes of the resultant errors
derived from the process of manual image-and-map
registration and are comparable with the theoretical error
predicted in Section 2.
