ant
r
id
of
ray
n of
on
ith
Comparison of the actual z-values (fig. 11, 12, 13) -as they result due to
differing sensor characteristics -with the permissible solution interval
defined in (4.4.5) shows that the condition z £7 is approximately satis-
fied for an area size of 100 x 1080 pixels only,but by no means for the
other area sizes.
An all-positive eigenvector is given only if the secondary diagonal elements
of the Q-matrix do not disappear simultaneously [2]. The values of the Q..
(i # j) decrease as the test areas grow small and the possibility of HJ
nonexistent all-positive eigenvectors grows accordingly. The variation of
the z-values can consequently be interpreted as a superposition of the
sensor-speci fic behavior and the stability of the eigenvalue problem [34.
5. Conclusion, Prospects
The extensive investigation of the North Sea showed a significant depen-
dence on position for the mean gray values and the largest eigenvalues. The
sensor-dependence is particularly evident in the substantial variations
observed in the all-positive eigenvector. This leads to the conclusion that
at least for the land cover type ''open waters'' there is no typical error
ellipsoid for the object. On the contrary, the determining parameters such
as ellipsoid center, length and direction of the semimajor axis, are sub-
ject to very large variations in some cases, a crucial factor being the
size of the investigated area: Based on the three different area sizes used
for analysis, the optimum as to eigenvalues was found to be at 100 x 100
pixels, even if the remaining root mean square errors of orientation can
amount to 709. This fact will certainly make classification considerably
more difficult. In addition, the training areas used in general practice
cover substantially less than 100 x 100 pixels which is likely to aggravate
the problems pointed out in the above.
As the analysis also shows, a relatively certain classification is only
possible within small image submatrices (100 x 100 pixels). On account of
their different properties the results fail to be transferable to adjacent
areas.
0f all influences affecting the accuracy of classification, position-depen-
dent systematic gray value variations (see: mean values) as well as some
random portions (see: ellipsoid dimensions) can be accounted for, resp.
minimized, through reduction of the area of analysis, yet the ellipsoid
function remains a position function. Improvement of sensor-dependent
ellipsoid orientation, however, seems feasible only through employing either
one single sensor or several sensors of identical calibration within a sen-
sor array. The latter technical requirement for sensor array scanners on
the one hand will probably be difficult to realize, on the other it con-
stitutes an indispensable precondition for improving the discrimination of
multispectral characteristics and consequently for object recognition.
177