International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
  
2. AUTOMATIC TIE POINT SELECTION AND IMAGE 
MATCHING 
2.1 Selection of Image Patches 
In order to find corresponding image tie points for registration, 
one needs to select and match image features with high 
information content. Our approach is to automatically search 
for image patches with good edges and corners. There are many 
filters (Bergholm, 1987, Iverson, 1995, Rothwell 1995 and 
Smith 1997) to detect edges and corners. We chose the Canny 
filter (Canny, 1986) to locate the edges in the image and then 
Harris (Harris, 1988) corner detector to detect corners from the 
detected edges. In this way, we reduce unnecessary edges, keep 
the corners with most feature information, and also increase the 
processing speed for image matching. 
There are a few processing steps in the detailed Canny detector 
(Canny, 1986): Gaussian smoothing, gradient computation, 
non-maximal suppression and hysteresis. We used a 11x11 
Gaussian window with a standard deviation o of 2.0 for the 
Guassian smoothing. A 3x3 Sobel filter is then used to derive 
the edge strength, i.e. the gradient of image intensity. In 
performing hysteresis, pixels with edge strength greater than 
98.5 percent of all strength or any connected pixel with strength 
greater than 50 percent are marked as edge pixels. In Harris 
(Harris, 1988) corner detection, a 5x5 Gaussian window with a 
standard deviation o of 1.0 is used as the gradient weight, and 
0.04 is chosen as the constant value in the corner/edge response 
function. 
The edges and corners detection are only applied to the high 
resolution PAN image. Image patches that contain one or more 
corner points are used as candidates to approximately select the 
corresponding image patches in the XS image based on the a- 
priori imaging parameters provided in the image header files. 
2.2 Hierarchical local matching 
After corner detection, a hierarchical local matching approach 
is implemented to match the image patches. The PAN and XS 
images are scaled by various factors to create a few hierarchical 
layers, with the highest resolution layer having a pixel size 
equal to that of the PAN image. The matching of two patches at 
various layers is computed by maximising the normalized cross 
correlation: 
  
  
N-1N-1 we = 
» 2 (A5 4x; -B) 
C(4, B)=——L (1) 
N-1 N-À 5 MN i 
NEON 
i-0 j-0 i-0 j-0 
where  C(4, B) = the normalized cross correlation between 
the scaled PAN and XS image patches 
N = window size of image patch 
Aj; and B; = individual pixel values in the PAN and 
XS patches, respectively 
A and B- average pixel values in the PAN and XS 
patches, respectively. 
940 
[mage matching starts with the lowest resolution layer. The 
panchromatic patch that contains one or more detected corners 
is used as the reference. By shifting the multispectral patch 
around its initial position, the location where the normalised 
cross correlation of the two image patches is highest within a 
range of shifting is selected as the matched position. For each 
higher resolution layer, the matched position derived from the 
previous lower resolution layer is used as the initial position 
and matching is performed to improve the accuracy of the 
matched position. A threshold of 0.85 for the cross correlation 
coefficient of the highest resolution layer has been set to reject 
those image patches with poor matching. 
The centre coordinates of the matched image patches are used 
as the respective coordinates of tie point pairs. 
3. SENSOR MODEL AND REFINEMENT 
3.1 Direct Sensor Model 
SPOTS image is delivered with header file or metadata file, 
which contains the sensor model imaging parameters such as 
satellite state vectors, attitudes, pixel look directions etc. Details 
of the SPOTS sensor model, including the use of the imaging 
parameters is described in SPOT Satellite Geometry Handbook, 
2002. 
The basic model described in the SPOT Satellite Geometry 
Handbook is the direct sensor model which relates the pixel 
coordinates (samp, line) to geographical coordinates (/at, lon) 
as follows: 
lon = F(samp,line,h) Q) 
lat = G(samp,line,h) 
where F and G are sets of various functions, collectively 
known as the direct sensor model. The relation of F and G to 
the various imaging parameters is briefly described below: 
For every pixel (samp, line) in an image, the samp coordinate is 
related to pixel look direction within the sensor, V^ while the 
line coordinate is related to time 4, The attitude (yaw, pitch 
roll) of the sensor at each line is given in the header file. The 
satellite sensor position and velocity can be interpolated at 
from set of time-tagged satellite state ectors. From the pixel 
look direction, attitude, position Rand velocity, the 
intersection of look vector with the topographic surface with 
altitude / above the earth ellipsoid gives the geographic (lon, 
lat) coordinates. 
3.2 Inverse Model 
For orthorectification and model refinement by least squares 
adjustments, it is desirable to relate the pixels coordinates 
(line, samp) in terms of geographical coordinates (lon, lath). 
samp - f (lon,lat,h) 3) 
line = g(lon,lat,h) 
Int 
me 
He 
he 
ba 
Fr 
ab 
co 
me 
an 
wi 
hi 
co 
3.4 
In 
ge 
sin 
an 
un 
int 
he 
chi 
DI 
im 
By 
arc 
pit 
wh