International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 Inter 
=) €® object 
= ! candidate roofs 
= candidate roofs [relative boundary non objects) 
=} candidate roofs (relative boundary vegegation) 
= @ stone like 
© gravel 
© stone 
@ slate 
@ stone like. no class 
& brick 
roof vegetation 
copper 
zinc 
rejected candidates (relative boundary vegetation) 
rejected candidates [relative boundary non objects) 
@ vegetation 
(©) non objects 
  
  
Figure 10: Class hierarchy Figu 
e Some materials show a significant different spectrum 
than the others, e.g. zinc and copper. 
e Some spectra of different materials are quite similar, 
e.g. stone plates and gravel. 
e Spectra of same material differ significantly due to the 
surface orientation in relation to the sun angle/ illumi- 
nation, e.g. brick or slate. 
Therefore, the main tasks are (1) to find specific charac- 
teristics of the spectra and select channels from the hyper- 
spectral data for the classification, and (2) to find quan- 
tities derived from the available channels, which reduce 
the influence of illumination. Furthermore, those materials 
showing a significant spectrum should be classified first, 
thus leading to a hierarchy in classification. The hierarchy Fig. 
we used is depicted in Fig. 10. First, we classify objects Om fh 
and non objects using the height information from laser spect 
scanning (first and last pulse). In a second step we derive a paris 
set of candidate roofs to be classified, by removing vegeta- 30.21 
tion areas from the objects applying an NDVI (channel 25 slate: 
and 15 of the HyMap-data) and smaller segments based on spect 
their size and their neighbourhood relations to segments of mate 
the classes non object and vegetation. Thus, this classifi- exam 
cation procedure may in principal also be applied, if only ing n 
a nDSM from first pulse data or derived from other sensor point 
data is available. The roof segments are now classified ac- 
cording to their material. For this purpose, we first have to 
define membership functions for each class and feature to 5 R 
be used, starting with those material classes with the most 
significant spectral differences to other materials. Zinc has . In fi 
high reflection values in the first channels and show some appre 
characteristic slopes, but these features seem to be differ- camp 
ent for new and and older zinc roofs. Therefore, the fuzzy some 
or(max) is used to compute the membership function value build, 
from the values of each feature. The spectrum of copper 
has a significant decrease from channel 8 to 20. Brick Fig. | 
shows an increase in the spectrum from the first channels Based 
to the last, which seems in our case to be independent from ficatit 
the age of the material. Slate, stone plates and gravel are ing m 
quite similar with respect to their spectra, but show differ- fist 3 
ences in channel combinations, although not as significant clude 
as decreases or increases of the spectra of the other mate- result 
rials above. Therefore, we tried different approaches for Foal s 
the computation of the class membership values based on theda: 
Figure 9: Segmentation (roof planes) and(min), or(max), and mean(arith.) and introduced also grave 
a class stone like, if no class of gravel, stone, or slate 1s (max) 
assigned. sificat 
1014