Urban-trees extraction from Quickbird imagery using multiscale spectex-filtering and non-parametric classification

Due to the more heterogeneous spectral-radiometric characteristics within urban land-use/cover units in very-high spatial resolution imagery, the traditional pixel-wise statistical and monoscale based classification approaches do not lead to satisfactory results. The main drawback of these methods is that they neglect the shape and context aspects of the image information, which are among the main clues for information extraction at very-high spatial resolutions. This paper presents a pre-classification filtering strategy based on unsupervised multiresolution non-linear image filtering that combines spectral and textural (spectex) image characteristics. In a multiscale model, the local texture characteristics are extracted via wavelet decomposition. The multiscale wavelets texture is then used to control the multiresolution spectral filtering process using the non-linear anisotropic diffusion approach. From the multiresolution non-linear filtering procedure, scale sub-bands suitable for urban-trees extraction are selected. The selected bands are integrated with a normalized difference vegetation index (NDVI) and a principal components transformation (PCT) for classification using a decision-tree (DT) non-parametric classifier. The DT results are compared with the statistical maximum-likelihood classifier. It has been demonstrated that with Quickbird imagery a classification based on the filtered imagery improved the extraction accuracy of urban-trees by 11.7% using the parametric maximum-likelihood classifier, and by 22.5% using the non-parametric decision-tree classifier. This is an increase from a 70.8% extraction accuracy when the respective methods are not used. The results further indicate that the non-linear filtering approach is superior to the linear (median) filtering technique, by 20.8% with respect to classification accuracy.