Contextual fusion by genetic approach applied to the classification of satellite images

摘要

Contextual fusion consists of the combination of multisource and multitemporal data by taking into account the information given by spatial or temporal context. In this paper, we propose a contextual fusion process for satellite image classification based on genetic concept. A Genetic Algorithm (GA) seeks the extreme of a function defined on data space called "initial population". Each individual of this population is a chromosome characterised by a whole of genes. Genetic programming uses this data structure and makes evolve iteratively the initial population through a specific number of generations by successively applying three genetic operators: selection, crossover and mutation. Searched solution is a final population obtained when a termination criterion is satisfied. To develop our contextual fusion process based on a genetic approach we have used two GA. The first one is applied during the training step to extract a data base of best chromosomes where each best chromosome characterises one information class. The second one is applied for the classification of each chromosome (pixel) taken in its spatial neighbourhood. Test data set available is TM5 multispectral image of Algiers city (Algeria). This image acquired in 1996, contains nine main classes illustrated in two thematic maps; one given by a genetic approach and the other by a Markovian approach. A comparative study is then carried out between these two approaches.