On the optimization and selection of wavelet texture for feature extraction from high-resolution satellite imagery with application towards urban-tree delineation

Integration of spectral and multi-scale texture is proposed in order to improve the detection and classification of urban-trees from QuickBird imagery. Arguing that spatial-structure semantic information exits at a hierarchy of scales and that texture is a consequence of objects in the hierarchy, multi-scale wavelets decomposition is proposed for the extraction of vertical, horizontal and diagonal texture components. Pre-selection of texture sub-bands is achieved via mean, entropy, variance and second angular moment. The resulting sub-bands are analysed for separability between trees and similarly reflecting features, such as rice-paddy, grass/lawns, open ground and playground, based on Kullback-Leibler (KL) divergence and Battacharyya distance. The results are ranked and classified with k-means. In comparison with the field data, KL gave the best results with omission and commission error of 4.4%. The proposed methodology has the ability to capture the increased natural variability in reflectance and improved the accuracy by 23.6%, in comparison with spectral-only.