Urban Land Cover Classification With Airborne Hyperspectral Data: What Features to Use?

This paper investigates the potential effects of spectral, shape, textural, and height information and their combinations on the classification of urban areas using airborne hyperspectral data. Based on analysis of the spectral, shape, textural, and height characteristics of urban land covers, the first ten spectral principal components, eight shape components, one height component, and seven textural components were selected to examine their performance on the classification accuracy. Correlation analysis was conducted to exclude correlated components. A support vector machine (SVM) was employed to determine the significant components affecting the urban hyperspectral classification through comparison of the classification accuracy. Different combinations of these components were then tested to estimate their contributes. The classification results showed that all these components contribute to the result of urban land cover classification, but different land cover classes benefit from the inclusion of different components. The experiment further revealed the effect of significant components on the classification of urban land cover in terms of area, convexity, elongation, form factor, rectangular fit, roundness, textual factors, and mean relative height. It is suggested that the inclusion of shape, texture, and height, together with the spectral components, significantly improved the classification accuracy of urban land cover.