852 
find out thermal-land-cover-types was attempted. The 
data used for clustering were the temperature-value- 
frequency-histcgrams (30 one-degree temperature vari 
ables) of the micropixels being united in the macro 
pixels of 25x25, 50x50 and 100x100 pixels size. As 
input for the clustering algorithm only the number n 
of the desired clusters has to be specified and in an 
iterative process all the macropixels were classified 
to the nearest of the n cluster centers. As distance 
measure the Euclidian distance was used. The best 
interpretable results were obtained by choosing six 
cluster centers. The cluster-macropixels were plotted 
in the same manner as the temperature-macropixels 
and Fig.3 to Fig.6 show the results. 
In the 100 x 100 macropixelfield only the dense 
woodland (cluster 6), large fields with maize (clus 
ter 5) and harvested fields or asphalted areas (clus 
ter 4) can be identified quite well. The rest are more 
or less mixed areas and not clearly interpretable. 
In the 50 x 50 macropixelfield (Fig. 4) the clusters 
can be interpreted in the following way: 
Cluster 1: fields cultivated with maize, single-fami 
ly-houses with large vegetated areas around. 
Cluster 2: single-family-houses , small regular struc 
ture, quarters with modern house-blocks and large 
vegetated areas in between. 
Cluster 3: green fields or meadows. 
Cluster 4: bare soil areas, harvested fields, industri 
al areas, railway and other traffic areas, motorway. 
Cluster 5: mixed areas with high rate of vegetation. 
Cluster 6: wood, large parks, water. 
In Fig. 6 the 25 x 25 macropixelfield possesses the 
following clusters: 
Cluster 1: centers of large green fields, meadows, 
mixed fields. 
Cluster 2: single-family-houses with regular structure. 
Cluster 3: green fields, edges of fields, meadows, 
clearings. 
Cluster 4: inner city without vegetation, harvested 
fields, motorway, house-blocks with warm roofs, 
railway areas. 
Cluster 5: harvested fields, industrial areas, rail 
way areas, sports-grounds. 
Cluster 6: wood, house-blocks with large shadows, 
partly gardens and tree groups. 
When we compare Fig.4 and 5 where only five clusters 
were determined we can see the mixture in cluster 
ordination (see table 6). 
6 CONCLUSION 
It is not possible to classify thermal images into 
very special land use classes like we can do with 
multispectral data because most of the land use clasj 
which can be separated very well in visible and near 
infrared light have overlapping sinature intervalls 
in one dimensional thermal infrared data. This can be 
seen very clearly from a table where different land 
cover classes and their thermal signature are listed 
(see Seger & Mandl 1985, p.73). 
So only thermal more or less homogenious classes or 
urban cover types can be separated. This separation 
can be done for instance by image enhancement methods 
namely generalization in the radiometric and/or spa 
tial domain or unsupervised classification using 
clustering of simple texture parameters like the 
temperature-value-histogram parameters in our study. 
The extracted classes should only be characterized 
in a qualitative way and a correlation with a ground 
truth map showing conventinal land use classes will 
only bring good results when the classes have very 
rough definition. 
We have to choose a certain level of generalization 
in all domains (spatial, radiometric and temporal) 
and choose corresponding methods to answer the questio: 
of the special work. 
REFERENCES 
Lee, D.O. 1984. Urban climates. Progress in Physical 
Geography 8:1-31. 
Nübler, W. 1979. Konfiguration und Genese der Wärme 
insel der Stadt Freiburg. Freiburger Geographische 
Hefte, Heft 16. Freiburg i. Br. 
Oke, T.R. & F.G.Hanneil 1970. The form of the urban 
heat island in Hamilton, Canada. In WMO, Urban 
Climates. Technical Note No.108. Geneva. 
Seger, M. & P.Mandl 1985. Strahlungstemperaturbilder 
als Beitrag zur Stadtklimatologie. In M.Seger (ed.) ( 
Forschungen zur Umweltsituation in Klagenfurt, 
p.59-93. Klagenfurt. 
Table 6. Crosstable representing the number of macro 
pixels classified in the clusters of Fig. 4 and 5. 
Clusters 
Fig. 4 
Clusters 
1 2 
Fig. 5 
3 
4 
5 
sum 
1 
45 
388 
2 
16 
29 
480 
2 
O 
185 
8 
7 
0 
200 
3 
44 
0 
0 
0 
0 
44 
4 
0 
0 
1 
116 
1 
118 
5 
0 
0 
13 
0 
0 
13 
6 
0 
0 
0 
0 
153 
153 
sum 
89 
573 
24 
139 
183 
1008 
When we do the same comparision using figures 3, 4 
and 6 no trend is discernible. We can see this fact 
also in the different descriptions of the clusters 
from Fig. 3, 4 and 6. That means that the different 
levels of spatial generalization represented by the 
different macropixelsize are more important for an 
urban-land-cover-type adequate classification of the 
thermal infrared image than the actual number of 
clusters in a certain range. For this reason the veri 
fication of the cluster-images (figures 3 to 6) were 
only done qualitatively and not quantitatively by 
correlating the maps with a handdrawn ground truth 
land use map. 
Symp 
Smal 
and; 
P.Misra 
Human ' 
ABSTRAi 
in India, 
planners 
from SF/ 
The SF, 
eg. slums 
Case st 
is highly 
ing with 
impact-a: 
A simpl 
1. INTRC 
Conventii 
cameras 
thousand 
their len: 
of much 
nology i; 
(aerial re 
lity is no 
high res- 
up 'dwell 
nation. 
Further 
format a 
as. If th 
'slums' a 
mission t 
out is no 
photogra 
The use 
lications 
uses are 
a) mon 
b) upda 
c) loca 
phys 
slum 
d) supp 
e) traf: 
f) vege 
tion 
It may 
by the sc 
Ministry 
However 
India. O 
of small 
the Hun 
of Remo 
has been 
started i 
the first 
urban pic