International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
4.2 The IKONOS image. 
   
Figure 5. The IKONOS image of a forestry zone 
The goal of this work is to perform an efficient classification of 
forestry scenes and more particularly to segment the forest in 
tree density classes. The classes of density of trees are visually 
very close from one to another; they differ only by their high 
frequency distribution. 
A test was carried out on a IKONOS panchromatic image of 
635x563 pixels which represents a forest scene of Labrador 
(Figure 4). On this image we can clearly distinguish different 
tree density classes, two lakes and non-stocked zones. We chose 
to fix the number of classes to 5: 3 different tree density classes 
(from dense to sparse), a class for the clear land and a class for 
the lakes. The results are given in the Figure 6, Figure 7 and 
Figure 8. 
Figure 6. 
  
zr» B » 
Figure 7. The Laws Filters 
36 
    
! Figure 8. The LMS method 
  
From these results, it clearly appears that the proposed 
algorithm gives more homogeneous segments and that the 3 tree 
density classes can easily be differentiated (Figure 8). 
Furthermore, the results given by the analysis based on the 
Haralick texture parameters are very heterogeneous. The lakes 
are not detected and the density classes as well as the "non- 
stocked" class are completely mixed (Figure 6). Generally, the 
obtained classes do not seem to correspond to those which one 
can visually detect in the IKONOS image. The Laws filter 
approach gives homogenous results but the lakes are missed and 
confusion exists between the tree density classes. 
5. CONCLUSION 
We have seen that the method based on the multifractal 
analysis, that we propose, gives good results in the case of 
IKONOS image, but also with brodatz textures. The tree density 
classes appear clearly and the non-wooded zones and the lakes 
are well detected. The LMS method uses only the high 
frequencies to classify the image. It would be interesting to 
integrate additional information, such as low frequencies, to the 
approach. This could help to identify regions with smooth 
textures, and to differentiate classes having very little local grey 
level variability but very different mean values. The 
multifractal analysis is an interesting tool for the texture 
analysis because it enables to characterize the singularities in a 
local and global way. However, the parameters required for the 
algorithm are not easy to compute automatically. A study on the 
automatic estimation of these parameters will be considered in a 
future work. It is also envisaged to use other classification 
methods than the K-means in order to see the possible profits in 
term of percentage of classification. Preliminary tests on the use 
of the Legendre spectrum are also in hand and give promising 
results. 
A ground truth image is to be produced by a photo-interpreter 
so that it will be possible to quantitatively measure the 
effectiveness of the method on the IKONOS images at our 
disposal. 
Internationa 
Abry, P., G 
Ondelettes € 
Arduini, F. 
multifractal. 
Internationc 
Processing, 
Berroir, J. ! 
doctorat de 
Chellappa, 
Textures 1 
Transaction 
ASSP-33, p 
Daubechies 
Industrial ai 
Frisch, U. a 
intermittenc 
on Turbul 
Dynamics à 
Grassberget 
strangeness 
208. 
Haralick, 1 
Textural fe 
on Systems, 
Kam, L. 
characteriz: 
Concepts fc 
Kaplan, L. 
Classificati 
Processing 
Laws, K. | 
238, pp. 37 
Lévy Veh 
Multifracta 
Pattern Re 
Mallat, S. : 
Processing 
Theory, Vo 
Muzy, J., 
formalism 
versus the 
of Statistic 
Pentland, 
Scenes. /E 
Intelligenc 
Turiel, A 
application 
Journal on 
Turner, M 
functions. .