Evolutionary Mining for Image Classification Rules

In this article, an approach for creating image classification rules using evolutionary operators is described. Classification rules, discovered by application of a genetic algorithm on remote sensing data, are able to identify spectral classes with comparable accuracy to that of a human expert. Particular attention is given to mutation operators and their efficiency in the creation of robust classification rules. Genetic operators and the fitness function are detailed, and then validated for hyperspectral images (more than 80 spectral bands). In our case studies, the hyperspectral images contain voluminous, complex and frequently noisy data. The experiments have been carried out on remote sensing images covering zones of Lagoon of Venice and the city of Strasburg, France. It has been shown that the evolution-based process can not only detect and eliminate noisy spectral bands in remote sensing images but also produce comprehensive and simple rules which can be also applied to other images.