0.08 AT. t T T T 
© 0.06 : 
o 
c 
> 
3 
o 
o 
° 0.04} 4 
© 
® 
2 
A E 
2 
z 002r 4 
F 0.00 l A 1 a. 
0 20 40 60 80 100 
Temperoture 
  
  
  
  
  
  
  
  
  
Figure 5: Top: Daedalus channel six image (0.695 — 0.750 um, recorded at 12:30 MET, flight altitude: 300 m), original class (black 
pixels) and the corresponding histogram of noon temperatures together with the appropriate Gaussian density curve (correlation 
coefficient: 0.85); below: NDVI-image (recorded at 12:30 MET) and the resulting two images after applying subdivision (step 2). A 
rectangular metal roof was isolated by this subdivision. 
with the ones based on NDVI methods (see figure 7). 
(75 + 4)% 
(79 + 8)% 
agreement(ndvi > 0.5) 
agreement(ndvi > 0.3) 
The errors represent the standard deviation of the agreement 
values. In Figure 6 two resulting vegetation maps are shown. 
It can be seen that the maps differ slightly from each other 
(for the reason of different initial seeds), but even in the worse 
case shown on the right the agreements with the NDVI-based 
maps are satisfying. 
6 DISCUSSION 
The subdividing of the classmap through the correlation 
mechanism always provides a better separation of the under- 
lying scene in comparison to a one-time unsupervised clas- 
sification. An inherent disadvantage of this method is the 
separation of data subsets which the user probably wants to 
see in one class. But this is inevitable, since man-made sur- 
face types cannot be assumed to have Gaussian distributed 
temperatures. 
Another source of misclassification is the necessary assump- 
tion that there is no change of the scene contents from one 
overflight to another. If this is not valid, the concerned ob- 
jects are probably misclassified because different parts of the 
corresponding temperature curves belong to different sub- 
stances. This effect can be seen in Figure 6, where aeroplanes 
which were only present in the early morning scene are part of 
232 
the vegetation maps. Areas which are shaded at noon show 
the same effect. 
Registration inaccuracies yield misclassification essentially at 
the borders of different surface types. Even by using many so 
called passpoints (corresponding points in different images, 
up to thirty in our case) it is not possible to get a registration 
accuracy better than one pixel. Objects with a specific height 
cannot be registered exactly if the corresponding viewing an- 
gles are different. 
The proposed algorithm is sensitive to the choice of the initial 
parameters for the first classification, as expected from the 
similar heating behaviour of many substances (see Figure 2). 
On the other hand the subsequent unsupervised classifications 
providing the subdividing of the "first guess classmap" are 
robust against the choice of its initial seeds. So the final 
vegetation map essentially depends on a one time choice of 
parameters. Providing some adequate temperature values as 
initial seeds (e.g. from regions of interest or model runs) one 
can avoid this problem. 
At one point the proposed algorithm surpasses the NDVI- 
based method: with a threshold of 0.3 the latter one clas- 
sified the metal roof as vegetation (Figure 7), the first one 
succeeded at this point (Figure 6). This is one example for 
the statement that by using thermal data one can improve 
classification results based on other methods. 
To illustrate the usefulness of histogram data the tempera- 
ture curves of all resulting classes for a typical classification 
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998