Multilevel Object Based Image Classification over Urban Area Based Hierarchical Image Segmentation and Invariant Moments

With the availability of very high resolution multispectral imagery from sensors such as IKONOS and Quickbird, it is possible to identify small-scale features in urban environment. Because of the multiscale feature and diverse composition of land cover types found within the urban environment, the production of accurate urban land cover maps from high resolution satellite imagery is a difficult task. In the present study, a multilevel object based classification method based on hierarchical segmentation and shape descriptors was proposed. Hierarchical segmentation was first performed by multichannel watershed segmentation and dynamics of the contours in watershed. Traditional watershed transformation defined for gray level image was extended to multispectral image segmentation by computing multispectral gradient image through a vector based approach, which uses extended dilation and erosion operations. The hierarchical multispectral image segmentation was then conducted by an improved method for dynamics of the contours. After segmentation, spectral features and shape features from different segmentation levels were calculated and combined in subsequent classification. The shape features used in the study were the Hu’s invariant moments, the useful shape descriptors. A hierarchical object based classification method was proposed, which combined pixel based and object based classification methods. The vegetation classes and shadow area were extracted by pixel based classification and a post classification processing. Non-vegetation classes were classified through an object based classification, which combined the spectral and shape features at multiple scales. The proposed method was compared with several classification methods, using a subset of Quickbird image covering Beijing urban area. The results showed that the proposed method produced higher overall classification accuracy, compared to other classification methods. ∗ Corresponding author