Urban features recognition and extraction from very-high resolution multi-spectral satellite imagery: a micro–macro texture determination and integration framework
This study presents the first experimental results on the integration of discrete wavelet transform (DWT) derived contexture (macro-texture) and grey-level co-occurrence matrices (GLCM) (micro-texture) in the recognition and extraction of the following selected urban land cover information from very-high spatial resolution Quickbird imagery: residential buildings, commercial buildings, roads/parking and green vegetation. The DWT filters capture the lower and mid-frequency texture information, whereas the GLCM captures the high-frequency textural components, for the same scene features. Besides the commonly used micro-texture (GLCM), the macro-texture (DWT) is modelled here to take care of the contextual information defined as feature edge (size and shape). This edge information is arguably derived from the multi-scale and multi-directional components of the DWT. From the statistical significance testing of the per-pixel classification accuracy results with the z-score, it was found that the integrated feature sets comprising the Quickbird spectral bands, 3×3 mean-GLCM and the first level of the vertical-DWT sub-band outperformed all the other tested input primitives, with a z-score value of 2.25. The accuracy results showed that all the three feature primitives were essential in improving the recognition and extraction of tested urban land cover in very-high spatial resolution Quickbird imagery.