Ali Akbar Abkar 
  
rest cover 
Dic 
on. 
function of 
jons in case 
is used for 
of 0.360612 
  
itive 
i-is 
'ao between 
pic expand 
  
  
Figure 8. Forest and non-forest cover map obtained from the maximum likelihood per-pixel 
classification of the Phrao Landsat-TM image of 1989. 
  
  
  
  
  
  
  
  
  
  
(a) with 19 land cover classes 
rest Cover (b) with two merged land cover classes (forest and non-forest) 
onstrained (c) after 8 applications of the majority filter. To treat isolated and mixed pixels, a post-merging 
IS operator is applied to the ML result, iteratively. 
ana Methods Overall accuracy 
1556 at soil- % 
r= 44. ML classification 69.39 
on-isotropic Filtering results 80.21 
LML classification 83.93 
rao between 
of the soil- Isotropic expand 86.29 
result. 
Non-isotropic expand 87.09 
(likelihood-based constrained expend) 
Soil-constrained isotropic expand 88.70 
Table 1. Accuracy assessment of the various models for hypothesis generation 
4. CONCLUSIONS 
In this paper a Likelihood-Based Segmentation and Classification method (LBSC) for remotely sensed images, with 
ultimate goal of describing objects in terms of their positions, sizes, shapes and geometric relations, based on a 
minimum-cost or maximum-benefit principle. A key feature of the LBSC method is to break down the image analysis 
into an evidence part and a hypothesis part; the evidence part comes from the RS data and the hypothesis part comes 
from the analyst knowledge which might have been stored in GIS. RS and pattern classification provide evidence 
(likelihood vectors) for subsequent analysis by segmentation and classification. Hypothesis evaluation provides a 
cost/benefit optimization criterion for finding the best hypothesis to match the evidence using the cost matrix, which is 
user specific. In this way the objectives and preferences of users/decision-makers are taken into account. The model 
parameters are adjusted until there is a minimum cost (or maximum benefit) value for the hypotheses and parameters. 
Therefore, instead of maximum a posteriori probability for each image sample (pixel), maximum a posteriori 
probability per object hypothesis and object parameters is derived. In other words, minimum cost of confusion per 
object over all objects and data samples is derived. 
In summary, the experimental results indicated the following main achievements: 
- A new method of image segmentation and classification to estimate the geometry and radiometry of objects using 
remotely sensed data. 
on based on 
- A framework for integration of prior/external knowledge (GIS) in image analysis, which is based on the combined 
probability of RS data and object models for the entire sample set of RS data by hypothesis generation and 
parameter estimation with a Bayes type cost/benefit function, which allows: 
- generation and prediction of the initial shape hypotheses of objects, and if relevant, incorporate constraints in 
ssults of the 
in order to hypotheses generation process 
10del comes - perform per-object segmentation and classification; 
erpretations 
- perform calculation of per-object accuracy (maximum likelihood per-object hypothesis); 
- derivation of the shape description of objects; 
- Provide a formal framework for application of available knowledge/information of objects in classification 
- Extension of the geometric model to objects of non-rectangular shape, i.e. the algorithm can create irregularly shaped 
segments; this makes the approach applicable to various cases. 
h. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 15