ET 
2, DIRECTION DEPENDANT RADIANCE BEHAVIOUR IN AIRBORNE 
MULTISPECTRAL SCANNER DATA 
Scanner data from 4000 m flight altitude were investigated which were collected during 
  
the "German Airborne Sensing Programme (FMP)" over thé test site Hartheim near Freiburg, . 
on 1. 4.76 (scene nr. 02/76/002/092/03). The following channels were selected: 
channel 1 = 0.41 um, channel 3 7» 0.52 um, channel 6 = 0.64 pm, channel 9 = 0.82 um. 
Ground truth for the central part of the scene were kindley supplied by Prof. Hildebrandt, 
Freiburg. 
For the data evaluation about 900 training fields were marked in this scene for bare soil, 
winterwheat and grass, based on ground truth information and near the edges of the scene 
on the colour of training fields. Then means and covariance matrices for each training 
field were calculated, sorted and plotted against scan angle (figure 1). Furthermore for 
each class in each channel first, second and fourth order polynomials were determined for 
means by least squares adjustment with weights proportional to training field size. 
Figure 1 shows the result of these calculations. 
Because the flight was conducted around noon (13.06 o'clock, sun elevation = 42°, 
sun azimuth = 212.5°) from east to west (flight direction = 264°), the left part of the 
image ( - 50°) is brighter, the right side (+ 50°) is darker. The examples in figure 1 
show typical results in the near infrared (channel 9) and red (channel 6) wavelength 
region. The direction dependant radiance behaviour is different for each object class and 
wavelength. Second order polynomials can sufficiently approximate the direction dependance 
for all analyzed objects. Although fourth order polynomials better approximate the real 
distribution of data, they are feasible only with many data points (e.g. bare soil or 
winterwheat). If there are few training areas (e.g. grass) however local peculiarities can 
influence this curve stronger than a second order approximation. On the other hand second 
order polynomials also can be falsified by a gap in training area distribution (e.g.grass, 
left side). 
In addition for the same training areas the variances and covariances have been plotted 
against scan angle (figure 1). 
By study ingthese graphs aweak but not significant trend seems to exist. The higher variances 
in the left part of the image could be caused by the higher mean values in that region 
and simulate in this way a not really existing trend. 
186 
a 
channel 9 
11 
MEAN 
œ 
channel 9 
VARIANCE 
channels 6/9 
COVARIANCE 
There 
range 
range 
matri: 
this s 
subdi: 
point: 
angle