le
LY
y
le
LS
computed from the same adjustment, the normalised residual is r-distri-
buted (Pope, 1976). However we could because of the high redundance in
this application use the normal distribution independently of how the
standard error is computed.
The computation of the weight coefficient matrix of the residuals is
extensive, even if only the needed diagonal is computed. However, the
diagonal terms of this matrix will be very close to unity because of
the high redundance, and need never to be computed. The, from computa-
tional aspect, complicated data-snooping turns into a simple test of
the size of the difference between the images in each pixel. The disad-
vantage with data snooping is that the design matrix has to be updated
between the iterations.
IDEAL WINDOW SIZE
The influence of the window size on the precision and the reliability
is experimentally investigated in this project. The investigated window
sizes are 12X12, 16X16, 20X20, 30X30, 40X40 and 50X50 pixels with a
multiplicative radiometric parameter and with and without affine para-
meters. In order to find an ideal window size a connection between the
window size and the precision of the measured parallaxes is necessary.
The following variance component model was used
2 2 2 2
g 3500 Leon NRO n
xy m n h
where
n = the number of observations (pixels)
o? ; o* : o = method, noise and height dependent components
We could find an extreme point where
n, z fd / m )
The optimal window size thus being
1/2
m = n
0 0
RESULTS AND DISCUSSION
Accuracy of the Manual Measurements
The results have been influenced by errors from the manual measure-
ments. It should be noticed that is not possible to ignore the influen-
ce from the manual measurements in the discussion.
Wilcoxon's Signed Rank Test;
In some parts of the discussion it is of interrest to see whether two
methods have significant differences in precision. The Wilcoxon's
signed rank test for pairwise observations has been used in this inves-
tigation. It has the advantage of beeing distribution free. The null
hypothesis that two groups have the same standard deviation was tested
against the two-ended hypothesis that they had different standard
deviations at the 5% level.
01579.