ROC. 
oblems 
rnative 
metric- 
and the 
ructure 
lata. In 
'rstood, 
evel of 
iminary 
error is 
zes and 
ground 
ıt10Ns 
|-map 
uit of 
map 
yman, 
|-map 
ssing, 
olems 
is the 
Since 
after 
uman 
dures 
1ative 
i or 
| that 
n is 
lapiro 
recise 
ge is 
‚such 
give 
per is 
data, 
tems, 
el of 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004 
automation. Cadastral parcels surrounded by wall features are 
of particular interest in the paper, and research is being 
focused on the analysis of the geometric characteristics of walls 
and land parcels. It should be noticed that prior knowledge 
about the error model in the procedures of image-and-map 
matching has not been fully understood. An error model is vital 
to give reasonable thresholds for optimising the automatic 
procedures of the proposed GSM technique. 
In Section 2, an error model is proposed not only to implement 
the algorithm in order to achieve high level of automation, but 
also to provide a sound theoretic basis for error evaluation of 
the proposed algorithms. The algorithm of the proposed GSM 
technique is validated using a sub-scene of standard Quickbird 
image and the corresponding cadastral map given by the 
relevant authority, as described in Section 3. An analysis is 
done using manual image-and-map registration in Section 3 to 
be compared with the automatic approach in Section 4. 
Automatic techniques have been developed and tested to match 
image features and the corresponding vector data, as shown in 
Section 4. Brief discussion and conclusions are covered in 
Section 5. 
2. THEORY 
2.1 Basis of Image-and-Map Registration 
The basic requirements of image-and-map registration include 
well-defined co-ordinate systems and identifiable features for 
image and map space. On the one hand, spatial information 
systems provide specific layers of vector data (polygons) over 
areas of interest in co-ordinate system of selected map 
projection. The landscape of areas of interest can change 
following local development or construction works, however, 
most of the cadastral parcels remain unchanged. On the other 
hand, the cadastral parcels characterized by boundary lines, 
such as walls, provide features found in high-resolution optical 
images taken by space-borne advanced sensors, as shown in 
Fig.l. 
  
  
   
  
  
  
(CQuickbird Image Copyright 2002, Digital Globe) 
Figure 1. (Left) A patch (165 lines by 182 pixels) of Quickbird 
standard image at Taoyuan, Taiwan. (Right) A segment ofa 
cadastral map provided by the Government of Taoyuan County, 
Taiwan. The circled node shows a wall feature. 
Since that walls are observable features appearing in high 
resolution satellite images or on air photos and are 
approximate boundary lines of cadastral parcels, therefore, it is 
straightforward to utilize geometric structure defined by 
lines/polygons to register map and image, leading to the 
proposal of a geometric-structure-matching technique in 
this paper. 
2.2 Basis of Geometric-Structure-Matching 
‘Structure’ can mean relational structure or semantic 
structure as the terms used by the pattern recognition and 
computer vision community (Shapiro and Haralick, 1981; 
Wang, 1998), however, it is referred to as the geometric 
structure in this paper. Roughly speaking, the so-called 
geometric structure can be given by the co-ordinates of 
nodes of each polygon, or endpoints of each line. A 
cadastral parcel given by governmental spatial information 
systems is defined by numerous nodes with known ground 
co-ordinates, which give an exact geometric structure that 
can be employed to guide the search of wall features in 
image space, provided that wall features are detectable and 
applicable. It is observed that wall features in satellite 
images show an U-shape intensity profile normal to the 
bearings of walls as shown in Fig.2. The bottom of U-shape 
curve corresponds to the shadow of a wall illuminated by 
the sun. 
380 
375 
370 
: 365 
:$ 360 
: 0355 
350 
345 
340 
  
Figure 2. An intensity profile normal to the bearing of a 
wall. 
In order to detect wall features in high-resolution satellite 
images, the height of wall and the sun azimuth/angle data 
provided by image header data have to be introduced to 
form a wall model in image space. The width of the shadow 
of walls of height A under solar illumination of solar 
altitude © is derived as s=hcot® on the ground. Given 
that h=2m and 0 =72¢ ¢ the shadow of the wall exhibits a 
dark line of width 0.65m, or 1~2 pixels, in a vertical 
satellite image. In case of oblique photography with tilt 
angle ©, the wall itself can be observed and be projected 
onto image space, showing a bright line of width w given by 
w=hcotQ. Thus, a dark-and-bright line pair exhibits a 
wall observed in an oblique high-resolution satellite image, 
giving a U-shape model of an intensity profile. In automatic 
approach, each polygon has to be employed to search across 
the boundary lines, pixel by pixel, to find and record 
candidate locations regarding to the U-shape model of 
intensity profile. The extent to be searched is determined 
by an error model, to be established in next section. 
For constructing a correct U-shape model of each polygon 
along the wall feature, the length between the contiguous 
nodes in the polygon is used as the first weighting factor to 
reduce the effects resulted from obscured features, 
assuming that a long wall feature always keeps the same 
radiometric characteristics along the boundary line. Since