CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
photographs of Alsatian façades. These images present also a 
small spectral resolution for a very high spatial resolution. 
The aim of this analysis is to extract the elements composing 
the half-timbered façades by developing specific rules, which 
should be reliable to each typical Alsatian façade. 
2. OBJECT OF INTEREST 
Object of interest is the Alsatian house, built in the 18th 
century. These typical residences can still be found in little 
villages of Alsace or in the old part of Strasbourg city. 
Only a few number of today’s carpenters are able to master the 
technique and knowledge allowing the construction of Alsatian 
half-timbered houses in the way our ancestors worked (Fig.l). 
Figure 1. First stage in the construction of an 
Alsatian half-timbered house 
Several digital photographs have been taken with a conventional 
camera. We used one of them - a typical (renovated) Alsatian 
house - for all experiments and the others for assessing the 
quality of the semantic rules. 
3. EXPERIMENTS USING 
CONVENTIONAL METHODS 
Façades of Alsatian half-timbered houses are composed of large 
and dark oak beams which constitute the structure of the house 
at the beginning of its construction. Between these beams, 
brickwork with facing replaces today the cob used by our 
ancestors. Whatever it is, the last coat is a light colour. So, in a 
first level of our experiment, we use for discrimination only the 
spectral signature of the pixels. 
3.1. Supervised classifications 
The choice of training samples is limited to six main classes of 
interest : sky, oak wood beams, freestone (sand stone), tiled 
roof, shutter, and facing (over masonry). A minimum distance 
as well as a maximum likelihood classification are performed. 
The first one uses a geometrical algorithm simply based on the 
Euclidian distance separating the pixel from the sample clusters 
in the spectral space, whereas the second one is more 
sophisticated since it is based on the probability (calculated 
according to sample statistics) for a pixel to belong to a class. It 
is certainly the most used classifier for remote sensing purposes. 
The main difference between both classifications is the tiled 
roof which appears more homogeneous in the maximum 
likelihood (see Fig.3 b) and c)). 
The other misclassifications are also mirrored by 2D scatterplots 
(Fig.2). Indeed, some classes like freestone, facing and also 
tiled roof present considerable spectral signature overlaps. 
Figure 2. 2D scatterplot of the image (Bl=red; B3= blue) 
with the 6 samples chosen for the experiment 
3.2. Unsupervised classification 
In order to test the presence of spectral classes, we start an 
unsupervised classification (isodata scheme). It is performed 
after imposing a maximum number of 6 classes. Nevertheless, 
only 5 clusters appeared (not tiled roof class) as shown in 
Fig.3a). This emphasises the main problem we are confronted 
with, i.e. attempt to define semantic classes on the basis of 
spectral classes. 
Quantitatively, in order to assess accuracy, we perform a 
confusion matrix using a reference classification and achieve: 
77% overall accuracy for the classification with minimum 
distance, 81% with maximum likelihood and 17% with 
unsupervised classification. Results are not really satisfying and 
above all they depend strongly on the samples. 
Qualitatively, it becomes clear that conventional classification 
methods are not efficient with HR imagery. Indeed, a lot of 
isolated single classified pixels are dispersed over the image 
(Fig.3 a) to c)). Although a majority filter could lessen this 
effect, the main problem remains: a spectral analysis alone is 
not powerful to discriminate classes with similar responses and 
different meanings. 
4. EXPERIMENTS USING 
OBJECT ORIENTED METHODS 
4.1. Supervised classification 
The object oriented image analysis system eCognition2.1 
(Definiens) follows the concept that semantic information 
necessary to interpret an image is not represented in individual 
pixels, but in meaningful image objects and their mutual 
relations. That’s why image objects are generated through 
multivariate segmentation where simultaneously spectral and 
shape heterogeneity criterions can be weighted. A scale 
parameter also influences the final object size by stopping the 
merging process. 
In a first step we decided to use the same training samples as in 
the previous experiment in order to assess the influence of 
objects instead of pixels. The sample-based classifier used is a 
fuzzy approach of nearest neighbour clustering. When 
activating only spectral feature as discriminating criterion, this 
classifier is nearly comparable to minimum distance. Thus the 
results are similar except for the homogeneity aspect (Fig.3 d)).