Automated extraction of buildings from IKONOS imagery by integrating spectral and spatial information

Very High Resolution (VHR) imagery is becoming indispensable in urban mapping and land management. But owing to its characteristics, traditional spectral-based building extraction methods are not appropriate. This thesis presents an automated spectral-spatial integration extraction approach to extract two-dimensional roof outlines of urban buildings from IKONOS imagery. It substantially improves accuracy, compared with the well-known Extraction and Classification for Homogenous Objects (ECHO) and Iterative Self-Organizing Data Analysis Technique (ISODATA). With spectral-based Multi-peak Supervised Segmentation (MSS), the approach generates first stage classification. In the second stage, the original colour image and the first stage result are fed into a spatial structure classifier. With analysis of overlapped neighborhood's degree of membership to classes, homogeneity and cross-neighborhood structure, the first stage classification is significantly improved. Finally, buildings are extracted with post-processing. Compared with ISODATA, Signature Editor and ECHO, unique contributions are three new algorithms: MSS, Overlapped Neighborhood Analysis, and Cross-neighborhood Structure Detection.