a i as
en EE wR
—————————À— € €——
- 1490 -
Match-points may be set artificially. Following an idea of M. Evans (2),
we have positioned small mirrors (45x45 cm2) which reflect radiation from
the sun into the sensor of the satellite as it overflies the test site. The
experiment shows that the "mirror-pixels" are clearly visible if the sun-
reflectors are positioned properly. Each mirror precisely defines its
pixel.
Careful examination shows that the distance between corresponding mirror
pixels will vary. The sum of the misregistration is about 1.5 pixels. It
appears that mirrors are quite good at marking an area but the positions of
the mirror pixels in different Landsat frames are not sufficiently precise
for pixel-wise overlay.
III Ceonetrical.transformat ions
A more reliable method of finding image displacements consists in matching
the frames by cross-correlation calculus. The frames arefirst of all aligned
visually by conspicuous image features. Final register of images is achieved
iteratively by refining their mutual positions in correspondance with the
position of the maximum of the cross-correlation function. This simple pro-
cedure is based on purely translational transformation.
For an easy visual assessment of the result of correlation the positioned
pictures together with an image of their pixel-wise differences are shown in
Fig. Ic.
In a perfect overlay the difference-picture should be of uniformly grey.
There are two reasons why this will in general not be the case:
l. the spectral variations of the elements of our landscape are relatively
large due both to the seasonal changes and to different atmospherical
conditions.
2. purely translational transformation is inadequate: other geometrical
transformations have to be implemented.
As Fig. lc demonstrates steep contrast gradients like lakeshores or rivers
are especially sensitive to a poor fit. In order to compensate the residuals
further computations are necessary.
A series of at least 6 subframes, 32 by 32 pixels in size, is chosen within
the correlation area used for translational alignment. Correlation calculus
on the small subframes or windows again yields a set of local displacement
vectors. The following transformations were considered:
1. rotation
2. affine transformation
3. interpolation with a polynomial of 2nd degree.
Fig. 1 shows the influence of the transformations in different pictures. It
appears that the most important improvements were achieved by rotation giving
best fit for extended linear features like lakeshores and rivers. Affine
transformation and non-linear interpolation did not improve the results
significantly beyond mere rotation, as judged by the value of the corre-
lation coefficient.
per THER.
UU EM Re en: dee 7 | rn
