The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
616 
Attention is focused on MRF, which are examined concerning 
their use for texture-based extraction of streets from aerial 
photographs. Besides the theoretical principles, also the 
methods and aspects of implementation are discussed such as 
simulated annealing, parameter analysis according to Maximum 
Pseudo Likelihood (MPL) and Conditional Least Square (CLS). 
In addition, different influences on the derivation of texture 
parameters are researched including the response to noise, 
scaling as well as to size and configuration of neighbourhood 
systems. This is referred to familiar textures (Brodatz-textures). 
In this paper, the application of MRF on aerial images is 
specified. Evidence is provided regarding the applicability of 
these approaches to distinguish between urban areas and 
vegetation. 
Afterwards the MRF are successfully applied to real image data. 
The paper ends with a summary and an outlook. 
2. STANDARD APPROACHES 
Figure 2: Segmentation results using Laws texture feature. Left: 
Original image, Middle: texture image, Right: classified image. 
Top and bottom line show different combinations of couples of 
Laws energy vectors. 
An useful compilation of standard methods can be found in 
(Tomita & Tsuji, 1990). 
Well-known approaches are based on the co-occurrence matrix 
(see e.g. Haralick, 1973) and derived parameters (e.g. contrast, 
energy, entropy). Single texture features are unsuitable for 
texture segmentation, caused by different viewing and 
illumination conditions as well as shadows, etc. Since single 
statistical texture features don’t permit image segmentation 
examinations were carried out using couples of texture features. 
All combinations were examined based on the first order 
statistics and the statistics were derived from the second order 
co-occurrence matrix. Only a few feature combinations allow 
good picture segmentation. Best results could be achieved using 
the couple of texture features entropy and energy (see figure 1). 
Segmentation results from single images cannot be generalized. 
Completely different results are obtained using different images 
with the same features. 
Unfortunately, the known problems appear also at the Laws 
features. The results vary strongly between the images, shade 
leads to problems and single objects (such as vehicles) are 
wrongly classified. 
It is to conclude that the classic attempts at texture-based 
segmentation are not sufficient. 
3. RELATED REPRESENTATIONS AND MRF 
3.1 MRF-based Segmentation 
For segmentation and classification in remote sensing 
applications, MRF-based methods have been established for 
several years. The MRF-model offers several advantages for 
image segmentation and classification. A finite number of 
parameters characterize spatial interactions of pixels in order to 
describe an image region. 
Figure 1. Segmentation results of a single image using entropy 
and energy. Left: Selection of the training sub-areas vegetation 
(green) and road (blue). Middle: The entropy over energy 
diagram. Right: Segmentation result. 
Similar results were also achieved by using Laws filter and 
Fourier approaches with a circle shaped spatial filter. The 
energies correspond to higher order statistics. Laws has 
introduced a uniform approach to derive many different texture 
features from an image. He defines linearly independent base 
vectors of the size 3, 5 or 7. These base vectors reflect 
respectively certain qualities such as intensity, edges or 
waviness. A designated couple is chosen from these base 
vectors and calculates a square mask using the dyadic product. 
The image is then convolved by this mask. Figure 2 shows 
some results of different combinations of couples of Laws 
energy vectors (Laws, 1980). 
Texture description and texture generation are essential steps 
for scene understanding and scene segmentation. Zhao et al. 
implemented an infrared battlefield simulation to test infrared 
weapons (Zhao et al., 2007). The MRF-model was used to 
synthesize various kinds of visible textures. Gaussian-Markov 
random fields (GMRF) were firstly used to sample the 
distribution of temperature fields on terrain surfaces. 
Images with both fine spectral and spatial sampling 
(hyperspectral images) can be used to perform a joint texture 
analysis in both discrete spaces. To achieve this goal a 
probabilistic vector texture model, using a GMRF was proposed 
(Rellier et al., 2004). The parameters allow the characterization 
of different hyperspectral textures. An application of this work 
is the classification of urban areas. 
Descombes et al. focus on GMRF and propose therefore 
different estimation methods (Descombes et al., 1999). They 
demonstrated the potential of the estimated parameters to carry 
out texture differentiation for remote sensing data. 
3.2 MRF modelling 
The image is called Markovian, if the probability distribution of 
grey levels for each pixel Zj merely depends on the grey levels 
of the neighbouring pixels Z‘. These images can be described by