In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
519 
in order to avoid misclassifications. In spectrally heterogeneous 
to very heterogeneous images, the classification had to be 
restricted to sufficiently homogeneous areas, since the 
segmentation of training trees led only in those areas to a 
reliable classification. As a result, in heterogeneous images, an 
increased number of training trees was necessary for a reliable 
classification, causing a much bigger effort. 
4. RESULTS 
4.1 Image data homogenization 
As mentioned above, the heterogeneity of the image data within 
their spectral bands caused a major problem, with the infra-red 
band showing the least differences between the images. This 
means, that even within one flight strip adjacent images showed 
decisive spectral differences (figures 1 and 2). 
Figure 2: Detail of two adjacent images (true color illustration) 
within one flight strip (cf. to figure 1): left side: original image 
prior to pre-processing; right side: enhanced and homogenized 
image data. 
Figure 1: Example for the spectral heterogeneity of the used 
image data: Adjacent ortho-rectified images within one flight 
strip of the research area (illustration in true color without 
histogram stretch). 
The quality control of the original image data histograms 
proved, that the reflection values for all spectral bands only 
covered about 1/3 of the possible histogram width. For this 
comparison, the histogram of digital image data from a flight 
campaign in 2006, equally using a matrix camera system, has 
been selected. The histograms of these images showed an 
almost ideal distribution over the whole possible reflection 
spectrum (figure 3). The shortening of the pixel-value area 
within the LVG image data from 2008 complicated the spectral 
delimitation and visual perceptibility of a tree species 
decisively. 
As another severe quality limitation of the image data in use, 
spectral differences between each single spectral band of the 
Vexcel images have been identified. The reflection values 
within the spectral bands red, green and blue varied up to 15 % 
in one flight strip. This strong spectral differentiation between 
the spectral bands originally averted the transfer of the 
classification algorithms from one image to its neighbors. 
These effects necessitated a homogenization and a histogram 
stretching of each spectral band as an essential prerequisite to 
the subsequent classification. Figure 2, 4 and 5 illustrate the 
extent of the spectral differences between the single images and 
the various spectral bands within the single image as well as the 
efficiency of the developed methodology for the image data 
improvement. 
<* o • 11* f*~afGo— 
C.e*M<} \ ■r*»\ 
C**s* 
«* 0 a i a O ® ~3 “* 
Ow*i PK**o*n Hwfc*** ;p**û**} 
6*>FuM!*9rt C'a»«! 
Figure 3: Histogram comparison of blue band. Left original, 
right after enhancement 
Figure 4: Two adjacent images after the pre-processing (True 
color illustration after pre-processing with broadened value 
range) 
Figure 5: Two adjacent images after the pre-processing (infra 
red illustration)