Genetic Programming as a Preprocessing Tool to Aid Multi- Temporal Imagery Classification

Classification-based applications of remotely sensed data have increased significantly over the years. Very often, these data are gathered from different sources and in different formats causing the classification process to be scenespecific. Alternatively, spectral band indices have been developed to emphasize some elements based on spectral characteristics and therefore improving the final classification accuracy. This research applies a multi-disciplinary approach in which genetic programming (GP) and standard unsupervised algorithms are integrated into a single iterative process to develop spectral indices for each element being investigated (such as water, impervious surfaces, dense vegetation, etc). A set of indices formed by mathematical and logical operations of the spectral bands are evolved using genetic operations. The application of non-linear indices enhances the relative spectral difference among the elements investigated improving the clustering capability of the data. The algorithm’s ability to generalize provides an alternative to classify multi-temporal data with a single methodology. An example application is given for the water and impervious surface delineation using Landsat MSS, Landsat TM, and Landsat ETM+ imagery. Initial results are comparable to more labor intensive scene-specific supervised classification.