Figure 3: Two-dimensional histograms of temperatures recorded at 4:30 MET (x-axis) and 12:30 MET (y-axis). Note that the axes 
have different ranges. For the definition of the regions see Figure 2. 
Gaussian distribution leads to similar good modeling results 
providing a much easier computability (see Section 4). 
4 CLASSIFICATION METHOD 
To separate the pixels of the underlying scene in appropriafe 
classes we used unsupervised k-means-classification (Duda & 
Hart 1973) which provides a well known algorithm for remote 
sensing applications. As data set we used four-dimensional 
temperature vectors described in Section 2. 
A striking disadvantage of unsupervised classification tech- 
niques is the unknown number of classes which should be 
used for an appropriate separation of the underlying scene. 
Therefore the user must preestimate the distinguishability of 
surface types with regard to their spectral properties. The 
230 
latter are in general unknown. Furthermore, surface types 
with similar spectral behaviour will probably not be separated 
into different classes. In some cases this could be avoided by 
providing appropriate starting parameters to the unsupervised 
classification algorithm, the so called seeds, instead of using 
random values. Again a preestimation is required. 
To guarantee a high degree of independence from human 
assessment we apply the following method (see Figure 4): 
1. start the classification with an arbitrary number of 
classes (6 is an appropriate choice for a first guess clas- 
sification in most scenes) 
2. compute the cross correlation of the histograms of early 
morning and noon temperatures of each resulting class 
and the Gaussian density function defined by the cor- 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998