les
o LOT DIRECTION DEPENDANT CLASSIFICATION OF AIRBORNE MULTISPECTRAL
nis par SCANNER DATA
oxception
Lure do BERTHOLD PFEIFFER -
civées, E
:indements Institut für Photogrammetrie und Topographie
"tive. Universität Karlsruhe
ENGLERSTRASSE 7
7500 KARLSRUHE GERMANY
ABSTRACT
un tra- Tae
'océdé. The classification of airborne multispectral scanner data over the entire scan angle yields
onse :
misclassifications mainly near the edges, due to non-representative statistical values. The
tent des analysis of data show object dependant brightness- and hue-shifts with scan angle, which
: d vie : : . aat
Se garcer can sufficiently be approximated by second orderpolynomials. This results indicate , that
ion des the direction dependant radiance of objects must be considered in the classification of
es airborne data.
Guvont The maximum likelihood algorithm requires statistical values (means, covariance matrices)
eS canaux representative for the actual scan position. To reach this goal, training fields regularly
distributed over the strip are used to determine first or second order polynomials for means
and covariance matrices for each class and wavelength. Using these polynomials values
for every scan position are calculated and applied during classification .Direction dependant
classification yields a marked improvement with homogeneous classification over the entire
strip, without errors near the edges.
I. INTRODUCTION
The direction dependant radiance behaviour of natural objects cause a considerable bright-
ness variation with view angle in airborne multispectral scanner deta. The evaluation of
such data, e.g. the multispectral classification leads to misclassifications near the edges.
Therefore in most cases only the central part of the image is classified. To evalute the
entire strip, possibilities for direction dependant classification should be developed.
A modification of the maximum likelihood classification is suggested, which varies the
statistical values with scan angle. This method is independant of the data source and could
also be applied to radar data.
185
sum m
——M RR t SERERE aa