Object-based VHSR image classification using multiband compact texture unit descriptor

Abstract

This paper evaluates the capability of a recently proposed method named multiband compact texture unit. This method extracts texture by characterizing simultaneously spatial relationship in the same band and across the different bands. This evaluation is performed in the context of object-based classification paradigm using WorldView-2 image of a forest area. For that image-objects were generated through superpixel segmentation. Classification in the object-feature space is performed suing K nearest neighbor algorithm. The proposed approach is compared to two groups of methods. The first group includes texture methods that use only spatial relationships in the same band: Gabor features wavelets and Granulometry. The second group includes methods that use intra-band and inter-band spatial relationships: integrative gray-level co-occurrence matrix, opponent Gabor features and opponent local binary patterns. Experimental results show that texture extracted using both intra-band and inter-band spatial relationship improves the classification accuracy compared to when it is extracted in each spectral band independently. Among the methods of the second group that use both intra-band and inter-band spatial relationships, the multiband compact texture unit method produces the best results.