Classification markovienne pyramidale adaptation de l'algorithme ICM aux images de télédétection
Date de publication1997
El Ghouat, Mohamed Abdelwafi
In this thesis, a method of pyramidal markovian classification was adapted to satellite image data based on Markov random fields and the Gibbs distribution. The Markov random field model and the Gibbs distribution exploit the dependence between neighboring pixels in images. The incorporation of new energy functions into the Gibbs distribution improves the precision of satellite image classification. These energy functions permit both better homogeneity in the class field and the preservation of principal discontinuities.In addition to new energy functions, certain methods of statistical analysis were also examined in order to take into consideration the spectral information provided by the image channels as a whole. The general modifications brought to the classical ICM algorithm (Iterated Conditional Modes) of markovian classification improves classification accuracy both qualitatively and quantitatively. However, the complexity of energy functions leads to an increase in processing time. This drawback was eliminated by applying the concept of pyramidal analysis to the classification of satellite images. The performance of the proposed energy function models is examined through the application of the modified ICM algorithm to Landsat and SPOT multispectral images acquired respectively over agricultural and forest areas. The algorithm is compared to a series of conventional classification algorithms such as the maximum likelihood, Isodata, SEM (Stochastic Estimation Maximisation) and classical ICM (before modification). The analysis of results shows that a robust function, among the energy functions adapted to the ICM algorithm, provides an improvement in classification accuracy. This function permits the preservation of principal discontinuities and the introduction of perfect homogeneity in the class field. We can conclude that the discontinuity preservation model is quite useful in the classification of multispectral satellite images. This conclusion is based on the basic theoretical notions of the new energy function models introduced by the Gibbs distribution and the Markov random field, and on the encouraging experimental results of this study.