Soft Computing Approach to Pattern Recognition and Image Processing - Part I: Pattern Recognition - Multi-objective Variable String Genetic Classifier: Application to Remote Sensing Imagery

摘要

The problem of partitioning different landcover regions from multispectral satellite images is treated as one of designing a classifier that can distinguish the pixels belonging to the different classes, given their intensity values in the different bands. Multi-objective genetic algorithm is used for designing such a classifier, CEMOGA-classifier, which can approximate well arbitrary complex/overlapping class boundaries using a variable number of hyperplanes. The concept of variable length real-encoded chromosomes is utilized for this purpose. The classifier is non-parametric and is designed in such a way that the number of misclassified training points and the number of hyperplanes are minimized, while the product of the class wise recognition scores is maximized. In addition to an existing way of implementing multi-objective genetic algorithms, a technique that takes into account the domain specific constraints is also proposed. The superiority of the classifier is demonstrated, both visually and quantitatively, on data from remote sensing imagery: a numerical Landsat data set, and a SPOT satellite image of a part of the city of Calcutta.