International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
3.4. Automatic Image Matching 
Automatic image matching is still a wide field of research, due 
to the need to use it with images which can vary a lot in content, 
radiometry and geometry. A variety of automated image 
matching methods exist. Some of them have been implemented 
and used at the Institute of Digital Image Processing (Joanneum 
Research, 1999). In these methods, a reference dataset, i.e. a 
pixel matrix or a vector of features derived from the image, is 
moved over a search window defined in the other image, and 
the respective correlation coefficient (or other similarity or 
difference criterion) between the corresponding pixel matrices 
or feature vectors is evaluated. 
The main methods being used presently are the so-called 
product moment correlation method and the feature vector 
matching method. In the first method, the corresponding point 
is defined to be at the maximum normalised correlation 
coefficient within the search window. A description can be 
found in Ehlers (1983) or Haberacker (1989). For the second 
method, a set of feature images is created for both images using 
various filters or other operators. For each pixel of the reference 
image, its feature vector is compared to the feature vectors in 
the search window of the other image. Then, the position of the 
minimum feature vector Euclidean distance defines the 
corresponding point. A description of this method can be found 
in Paar and Polzleitner (1992). For either method, 
backmatching by inverting the image matching procedure can 
be applied in order to detect and reduce matching failures. 
3.5. Software Implementation 
Methods for geocoding have being developed at the Institute of 
Digital Image Processing since the middle of the eighties. First 
multisensor applications of developed methods are described in 
Buchroithner (1989). The know-how and methods have been 
also used in the ERS data geocoding system GEOS, which has 
been developed at the German PAF (Schreier, 1993). 
During the last years, the various developed methods, including 
image geocoding, have been combined in the commercial 
Remote Sensing Software Package Graz (RSG, see Joanneum 
Research, 1999). This software package is designed for general 
geometric processing and applications of remote sensing 
images. The RSG software package is fully integrated into the 
ERDAS/Imagine system. The main software functions are: 
• Geometric modelling of single as well as multiple images 
acquired by various sensors. 
• Integration/Optimisation of given orbit and imaging 
parameters of these sensors. 
• High precision geocoding including a DEM. 
• Stereoscopic extraction of 3-D information by automated 
stereo matching. 
• Generation and optimisation of digital elevation models 
from image matching results. 
• SAR interferometry: generation of coherence, fringe and 
phase images from single look complex SAR data. 
• DEM generation from interferometric phase images. 
4. MULTISENSOR IMAGE COREGISTRATION AND 
GEOCODING EXAMPLES 
4.1. Polynomial Rectification 
Polynomial rectification will again be only briefly mentioned, 
since it is to a high degree inadequate, particularly for 
geocoding of multisensor image data. As an example, a 
multispectral SPOT image, acquired at an off-nadir angle of ca. 
22 degrees over the “Oetztal”-area in the Austrian Alps, was 
geocoded by the polynomial approach. Terrain elevations in this 
area are in the range between 1500 and 3500 m. 
Table 3. Matching accuracy of geocoded Landsat TM and 
polynomial rectified SPOT image (pixel size 20 m). 
East (pel) 
North (pel) 
Distance (pel) 
Mean 
3.15 
0.80 
7.61 
Std. Dev. 
8.37 
2.21 
5.23 
The output product was compared to a geocoded Landsat TM 
image being generated by parametric geocoding. The geometric 
differences between these products are clearly visible and were 
quantified by applying an automatic image matching to these 
images, the results being summarised in Table 3. As it can be 
seen, the average displacement amounts to 7.61 pixels of 20 
meters, i.e. 152 m. The standard deviation is around 100 m and 
the maximum displacements (not given in the table) are in the 
range of 500 m. 
4.2. Geocoding 
Illustrative examples for geocoding applications are shown 
below. Figure 9 shows a set of multisensor image data of the 
high-alpine “OetztaT’-area including optical and SAR images. 
For the geocoded Landsat TM and SPOT XS images, a 
correspondence check was performed using automatic image 
matching. Statistics of the measured displacements are 
summarised in Table 4, while Figure 10 shows the distance 
(displacement) values in a grey value coded presentation. Both 
Table 4 and Figure 10 show an obvious systematic 
characteristic of the discrepancies, indicated by the mean values 
and the striped features, respectively. 
Table 4. Matching accuracy of geocoded Landsat TM and 
SPOT ortho-image. 
East (pel) 
North (pel) 
Distance (pel) 
Mean 
-0.32 
1.25 
1.60 
Std. Dev. 
0.90 
0.74 
0.65 
A further experiment was related to the fine-registration 
between geocoded Landsat TM and SPOT XS images using the 
coregistration procedure mentioned in section 3.3. The 
matching accuracy of the fine-registered images was checked 
again by image matching, leading to the statistical differences 
summarised in Table 5. It can be seen, that the mean 
discrepancy between the images could be reduced from 1.6 
output pixels (Table 4) to 0.49 output pixels through fine