tion of
jally
)r is im-
à Maximum
; should
lihood
>-Versa,
with the
111
jerung
n der
8. Ph.D.
terns on
nterna-
ermany ,
, Acade-
Graw-Hi ll.
tion
as, TTC
ive Ap-
DIGITAL IMAGE PROCESSING OF REMOTE SENSING IMAGERY:
A COMPARATIVE STUDY ON DIFFERENT OBJECTIVE FUNCTIONS
IN CORRELATION PROCESS
MANFRED EHLERS
Institut fiir Photogrammetrie
Universitat Hannover
D - 3000 Hannover 1
ABSTRACT
In this paper five different objective functions for correlation process have
been compared in an empirical test on identical images with added random noise
in varying magnitudes. The accuracy of the correlation function is computed on
20 correlation points for each image. One result is that phase correlation
turns out to be relative undistortable against low signal-to-noise-ratio. On
the other hand a careful selection of correlation points according to the vari-
ances is of great influence in accuracy and sharpness of the objective function.
INTRODUCTION
In recent years digital correlation turns out to be a useful tool to detect
identical points in different images for rectification of remote sensing image-
ry (Gópfert 1980). Even in images of natural landscape like tidal lands digital
correlation methods have been applied successfully (Ehlers and Dennert-Moller
1981).
Tidal lands of the German North Sea coast are the test sites of the 'Sonder- -
forschungsbereich' (SFB) 149, a special research organization for ‘Surveying
and Remote Sensing at Coasts and Oceans'. One great disadvantage for digital
correlation in images of this terrain is the very low signal-to-noise-ratio
(SNR) which leads to inaccurate parallaxes (Ehlers and Wrobel 1980).
So preprocessing of the images, i.e. twodimensional digital filtering is one
method to increase correlation accuracy (Ehlers 1982, Fórstner 1982).
On the other hand the linkage between accuracy and objective function in corre-
lation process has not been investigated systematically although different
functions for this purpose have been presented (Ehlers and Wrobel 1980,
Gópfert 1980, Kuglin et al. 1979, Makarovic 1980).
In the following we present the results of a test series to answer this ques-
tion. The test computations have been executed on the image processing system
MOBI-DIVAH of the SFB 149 (MOBI, DIVAH-Handbuch ).
OBJECTIVE FUNCTIONS
The correlation concept of the SFB 149 is based on a modular structure. 50 it
allows an easy exchange of different objective functions for the correlation
process. Five of them have been tested:
a) The 'normal' product moment correlation coefficient,
b) the correlation intensity coefficient that has been derived out of coherent-
optical considerations (Gópfert 1980). The image signals are mapped on the
complex plane and the intensity of the complex correlation function is com-
puted. The coefficient is weighted by a parameter Py which depends on the
71