ET EEE 
" re sc ace ss 
Eh htt a A LES 
the solutions, So they must be estimated in 
proper way. 
The solution of exterior orientation elements 
can be obtained with very high accuracy, if 
the above mentioned techniques are applied. It 
is shown from experiments that root mean 
squares (RMS) of control points is 0.6 pixel 
and RMS of check points is less than ] pixel. 
2. DATA ORGANIZATION FOR IMAGE MATCHING 
After the solution of the exterior orientation 
elements and its variation rate of the left 
and right images, the terrain elevation can be 
calculated using the conjugate points 
coordinates after antomatic image matching. In 
order to generate the DEM data in regular grid 
base, the interpolation processing must be 
implemented to convert the elevation data of 
irregular grid to that of regular grid, 
Considering the geometric characteristic of the 
SPOT data. We designed a new strategy for data 
organization before image matching procedure. 
We take one image (left or right) as a master 
image. The geometric rectification between the 
master image and the topographic map is 
performed by means of the polynomial 
transformation. The rectification result is 
illustrated in Tab.1, in which one can be seen 
that the planimetrie errors are limited within 
1 pixel. Approximately, if we create the 
regular grid on the master image, the grid can 
be taken as a orthophoto-grid. Based on this 
grid , the multilevel image matching and 
intersection are performed for each point of 
the grid and then the elevation data can be 
obtained, that is to say , the DEM has been 
generated. As can be seen, we have accquired 
DEM data in regular grid directly. The 
advantges of the above strategy are that the 
interpolation procedure can be avoided and the 
accuracy loss from the interpolation is also 
eliminanted. 
3.  MULTILEVEL IMAGE MATCHING 
In Accordance with the geometric character- 
istics of SPOT image, the strategy of three 
level image matching is adopted as follows: 
the first level, image to image registration, 
the second level, image matching with accuracy 
of 1 pixel to generatse the coarse parallax 
grid, and the third level, high precision 
image matching with modified Least squares 
matching algorithm which can reach sub-pixel 
accuracy. 
4 
876 
3,1 Level 1 image matching 
In most image matching strategies for SPOT 
stereopairs, the "epipolar’ lines images are 
needed to be generated for restricting the 
correlation searching range. However. it is 
impossible to define exactly the epipolar 
geometry for SPOT , since a pair of SPOT 
images does not have corresponding pairs of 
straight epipolar lines. Dowman (Dowman, 1987) 
has put forward a method to settle this problem, 
but his method must need to know DEM data. Zhou 
(zhou, 1988) has proposed another way in which 
epipolar lines of SPOT stereopairs are arranged 
approximately using the ploynomial fitting 
technique, The main advantage of the Zhou’s 
method is that it need not any orientation and 
terrain elevation information, 
In this paper, the ‘epipolar lines” images 
strategy which can be taken as a 
preprocessing procedure for matching has not 
been adopted. We take the approach of image to 
image registration as the first level 
procedure in multilevel matching. The approach 
is implemented by resampling the right or left 
image ( the other is considered to be a master 
image) , according to the ploynomial 
registration relation which is created by 
an amount of well-distributed corresponding 
control points selected by semiautomatic or 
automatic manner. It is shown from the test 
results that the ranges of parallax variation 
in all directions are restrricted by the 
registration approach , which is obviously 
different from the "'epipolar" processing 
approach since the latter confines only the 
parallax variation in verticle direction. So 
we called this step of registration processing 
as level 1 image matching 
Tab. 1 illustrates the experiments resuts for 
two test areas. As can be seen that only a few 
control points are used and the parallax 
variation is well restricted , especially. in 
vertical direction, the maximum parallax is 
less than 3 pixels which is profitable to the 
successive matching processing. it should be 
mentioned that the block by block image 
registration processing should be implemented 
in the area where the relief is distinct. 
3.2 Level 2 image matching 
The purpose of this level is that the parallax 
grids with accuracy of 1 pixel ‘should be 
generated as fast as possible. firstly, the 
image segmentation is carried out in the 
master (left or right) image. In the SPOT image, 
there are some areas, which we called matching 
dead-areas, such as clouds, snow, water and