Parameters Optimization for Urban Vegetation Classification Using KPCA

It is important that urban vegetation can be extracted automatically and precisely from high spatial resolution satellite imagery. However, the data is subject to spectral shortage and has strong nonlinear characteristics in the original spectral space. In this paper, the kernel principal component analysis (KPCA) is used to analyze the separability of the vegetation features and to select features using IKONOS data. The parameter of the kernel function is determined and the sample number is optimized with mean B-distance (MBD) using our test dataset. Experiments have shown that the maximum MBD is found to be 6.8 and 2.5 after and before KPCA respectively for six vegetation types in 50 dimensions with the power of 0.01. The KPCA-based spectral angle mapping (SAM) is also compared with both the original SAM and classic classifiers. It is demonstrated that the proposed approach is able to achieve more accurate classifications thus produce better urban vegetation map than the conventional methods