ibul 2004 
ters: two 
> rotation 
factor — s 
projection 
form, as 
ing object 
projection 
imaging 
| constant 
projection 
along the 
€ scanner 
ction for 
additional 
projection 
allel one, 
h scene if 
| model, it 
o epipolar 
zed scene 
for DEM 
its can be 
presented 
esampling 
gan et al., 
Equations 
sing some 
) 0, K, Áx, 
eters (4, — 
t and right 
meters for 
] parallel 
y-parallax 
x-parallax 
ortional to 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B1. Istanbul 2004 
e Use the parallel projection parameters for the original and 
normalized scenes to transform the grey values from the 
original scenes into the resampled scenes according to 
epipolar geometry. 
3. DEM GENERATION 
So far, the original scenes have been resampled according to 
epipolar geometry. The resampled scenes exhibit two main 
properties. First, conjugate points are located along the same 
rows. Second, the x-parallax between conjugate points is 
proportional to the height of the corresponding object point. The 
following subsections briefly discuss the utilization of the 
normalized imagery for DEM generation. The generation 
process involves four steps: primitive extraction, primitive 
matching, space intersection, and interpolation. 
3.1 Primitive Extraction 
At this stage, a decision has to be made regarding the primitives 
to be matched in the normalized scenes. Possible matching 
primitives include distinct points, linear features, and/or 
homogeneous regions. The choice of the matching primitives is 
crucial for ensuring the utmost reliability of the outcome from 
the DEM generation process. In this research, point features are 
chosen. This choice is motivated by the fact that the parallel 
projection model, which will be used for space intersection, can 
only deal with point primitives. 
Forstner interest operator (Forstner, 1986) is used to extract 
distinct points from the imagery. The operator identifies points 
with unique grey value distribution at their vicinity (e.g., corner 
points and blob centres), thus reducing possible matching 
ambiguities (refer to Figures 5 and 6 for a sample of the 
extracted points). The next section discusses the matching 
procedure of these points. 
3.2 Primitive Matching 
The outcome from the interest operator is a list of distinct points 
in the left and right scenes. This section deals with the 
identification of conjugate points,. which is known as the 
matching/correspondence problem. The solution to this problem 
can be realized through defining the location and the size of the 
search space as well as establishing matching criteria for 
evaluating the degree of similarity between conjugate points. 
3.2.1 Centre of Search Space 
The search space outlines the area where conjugate points to the 
selected matching: primitives are expected. For normalized 
scenes, conjugate points should lie along the same rows. The 
relative location between conjugate points along the epipolar 
lines as described by the x-parallax depends on the height of the 
corresponding object point. An approximate value for the x- 
parallax can be derived by knowing the average height 
associated with the area under consideration. 
3.2.2 Size of the Search Space 
For normalized imagery, the size of the search space in the right 
scene across the epipolar lines should be exactly equal to the 
size of the template centred at the matching primitive in the left 
scene. In other words, the search is done in one direction along 
the epipolar lines as represented by the rows of the normalized 
scenes. However, since there is no guarantee that the y-parallax 
between conjugate points in the normalized scenes is exactly 
zero, the size of the search space across the rows is selected to 
be slightly larger than the size of the template. 
395 
On the other hand, the size of the search space along the rows 
depends on the height variation within the imaged area. Rugged 
terrain would require a wider search space compared to tha 
associated with relatively flat terrain. 
o 
c 
t 
3.2.3 Matching Criteria 
The matching criteria deal with establishing a quantitative 
measure that describes the degree of similarity between a 
template in the left scene and a matching window, of the same 
size, within the search space in the right scene. Either 
correlation coefficient or least squares matching could be used 
to derive such a similarity measure. Since the involved imagery 
is captured by the same platform within a short time period, one 
should not expect significant variation in the appearance of 
conjugate features in the acquired scenes. Therefore, we used 
the correlation coefficient for deriving the similarity measure 
between selected features in the left scene and hypothesized 
matching candidates in the right scene. The centre of the 
matching window that yields the highest correlation coefficient, 
which is larger than a predefined threshold (e.g., 0.7), is chosen 
as the initial match of the selected point in the left scene. 
To eliminate possible mismatches, initially matched points 
undergo a consistency check using neighbouring matches. More 
specifically, the x-parallax (P,) value of a matched point is 
compared to those associated with the surrounding matched 
points. For a given point, the mean and standard deviation of the 
x-parallax values of the neighbouring points (i.e., # and o, 
respectively) are computed. Then, if the x-parallax for that point 
is significantly different from those associated with 
  
Px-u 
js - ul » Threshold ) 
T 
highlighted as a mismatch and is eliminated. 
neighbouring points (e.g., the point is 
3.3 Space Intersection 
Following the matching process, conjugate points undergo an 
intersection procedure to derive the ground coordinates of the 
corresponding object points. The parallel projection formulas 
(Equations 2) are used for such computation. For a conjugate 
pair in the left and right scenes, one can formulate four 
equations with three unknowns (X, Y, Z - the ground coordinates 
of the corresponding object point). These coordinates can be 
derived through a least-squares adjustment procedure. 
3.4 Interpolation 
So far, the ground coordinates of matched interest points, which 
passed the consistency check, are derived through space 
intersection. These points are irregularly distributed and are not 
dense enough to represent the object space. Therefore, they 
need to be interpolated. In this research, Kriging is used to 
interpolate the resulting object space points into regular grid. 
The Kriging methodology derives an estimate of the elevation at 
a given point as a weighted average of the heights at 
neighbouring points. However, the weights are stochastically 
derived based on the statistical properties of the surface as 
described by the elevations at the matched interest points 
(Allard, 1998). 
So far, we outlined a comprehensive methodology for DEM 
generation from high-resolution satellite scenes using an 
approximate model.. The reliability of the matching process is 
enhanced by utilizing scenes that are resampled according to 
epipolar geometry. The following sections deal with 
experimental results from real datasets to evaluate the validity 
and the performance of the developed methodology.