tion 
itive 
98). 
nner 
ISOT, 
10us 
ith a 
olar 
t an 
the 
aths, 
er of 
jm, 
the 
"reat 
d in 
The 
' are 
scan 
d by 
airs" 
Poli, Daniela 
  
Figure 6. Difference between odd Figure 7. Differences between nadir image (left) and forward 
and even columns. image (right). Northern Adria. 
The ATSR2 GBT data were reduced to 8-bit and linearly stretched between the minimum and maximum value, cutting 
0.5% on both sides of the histogram (excluding the pixels assigned with an error code). To reduce the image quality 
problems (Fig. 6), a 3x3 median filter has been used before the generation of the image pyramid. As no a priori values of 
the cloud heights are given to the matching algorithm, a hierarchical matching procedure with 3 pyramid levels is applied 
so that the maximum possible parallax at the highest level is only 1-2 pixels. Every pyramid level is enhanced and 
radiometrically equalized with a Wallis filter. 44919 points are selected with an interest operator in the first pyramid level 
because it is likely that the same points are well detectable also in the third level and the original image. 
The matching was done with the unconstrained mode of the Multi-Photo Geometrically Constrained Matching Software 
package developed at our institute (Baltsavias, 1991), which is based on Least-Squares-Matching (Griin, 1985). Two 
independent results were processed through all pyramid levels, one working on the grey-level image, one working on the 
derivative image. For both sets, a 3x3 mean filter was included in the matching of the original image to smooth closed 
regions introduced by the previously applied median filter. 
Due to the different looking angle and the not equal pixel size of the two images, matching can be very difficult, 
especially with thin differently looking clouds (Fig. 7). It is also necessary to work with different matching parameters as 
every version can be good for some areas but unsuccessful in other regions. Fig. 7 shows the Northern Adria region 
where the second method (derivative image) could solve the problem of the strong illumination difference between nadir 
and forward image (see also Fig. 8). 
  
Figure 8. Left: unsuccessful matching on grey-level image, right: successful matching on derivative image. 
The solutions of the two sets are first independently quality-controlled with absolute tests on the x- and y-disparities and 
relative tests on the correlation coefficient, sigma 0 and the change from the approximations in x- and y-direction. 
Afterwards, the two sets are combined in the following way: points which just exist in one version after the quality 
control are taken without further examination. For the other points, the x- and y-disparities, sigma 0 and the correlation 
coefficient is compared to decide which point version is taken. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 1167