Graph Search and its Application in Building Extraction from High Resolution Remote Sensing Imagery

Abstract

Man-made object recognition from remotely sensed imagery is not only scientifically challenging but also of significant practical importance for spatial data acquisition and update of geographic information system databases, mapping, cartography, image interpretation, military activities and other applications, etc. In the literature, a large amount of work that has been done in the field of high resolution image understanding focuses on the development of efficient and robust algorithms to detect and extract typical man-made objects, such as buildings and roads. Most methods for building extraction can be classified into two categories: edge-driven approaches and region-driven approaches. The edgedriven approaches usually involve a procedure of bottom-up processing of image primitive features, trying to link or group the linear features corresponding to a building to obtain building boundary. In a region-driven strategy, various methods, such as artificial neural networks, support vector machines, machine learning strategies and other traditional classification schemes in pattern recognition are employed to categorize the regions derived by segmentation based on region features. The following section briefly reviews the methodologies of these two categories available in the literature. Edge-driven methods, such as perceptual grouping and contour tracing, are widely used for building extraction in the literature. Perceptual organization (Mohan and Nevatia 1989) was used to detect and describe buildings in aerial images. There, linear features are firstly extracted and grouped into parallel lines. Parallel lines with aligned endpoints trigger the formation of a U structure. Two U structures trigger the formation of a rectangle hypothesis which is further filtered. Katartzis and Sahli (2008) proposed a method based on a stochastic image interpretation model, using a novel contour-based grouping hierarchy under the principles of perceptual organization. The BABE (Buildup Area Building Extraction) system (McKeown 1990) performed perceptual organization of lines and orthogonal corners into chains and rectangles to form building hypotheses that were further verified using shadow information. Lin and Nevatia (1998) derived the geometric relationship between building margin lines and building shadows analytically according to a general illumination model.