habbar bleibt. Die Methode wird an mehreren Beispielen er-
probt.
1. Introduetion
With multispectral scanning of the earth's surface from sat-
ellites or airplanes, mean radiation intensities of areal
elements are measured in several fixed wave-lengths ranges
and digitized. Considering the fact of equal radiation fea-
tures for equal topographic objects, one can conclude from
these spectra to the topographic objects producing them,
provided that size and amount of the selected frequency win-
dows are sufficient. This procedure of multispectral classi-
fication leads possibly together with other recognition al-
gorithms by means of digital image processing systems to the
automatic output of land use maps. The separability of ob-
jects, however, is considerably determined, as already im-
plied above, by the non-linear dependency of the multispec-
tral data between themselves. For this reason, it is neces-
sary to reduce them to the significantly different feature
vectors (data vectors of the spectral data sets), thus also
accelerating their processing on the computer. The computa-
tion of correlation coefficients (and procedures derived
thereof [1]) between the feature vectors does not yield the
desired information on spectral differentiations in detail,
because these statistic methods suppress differences for
those objects the portion of which is only small in the to-
tal data quantity, but the importance of which can be essen-
tial for eartographie purposes (e.g. roads, hydrography).
For this reason, a method is applied covering each pixel and
representing it in correct position, if its value exceeds a
threshold (Student-t-value) determined by least squares ad-
justment. Thus, from a noisy structure of these pixels a
linear dependency can be concluded. From a systematic struc-
ture (linear or areal) follows that the pixels do not depend
on the object and its position.