Feature Selection with Ant Colony Optimization and Its Applications for Pattern Recognition in Space Imagery

This paper presents a feature selection (FS) algorithm using Ant Colony Optimization (ACO). It is inspired by the particular behavior of real ants, namely by the fact that they are capable of finding the shortest path between a food source and the nest. There are considered two ACO-FS model applications for pattern recognition in remote sensing imagery: ACO Band Selection (ACO-BS) and ACO Training Label Purification (ACO-TLP). The ACO-BS reduces dimensionality of an input multispectral image data by selecting the "best" subset of bands to accomplish the classification task. The ACO-TLP selects the most informative training samples from a given set of labeled vectors in order to optimize the quality of a supervised classifier. The proposed ACO-FS model applications have been evaluated using the dataset of a LANDSAT 7 ETM+ multispectral image. The experimental results have confirmed the effectiveness of the presented approaches.