Seeing residential buildings from remotely sensed imagery: An object-oriented approach

Automatically extracting land use information from remotely sensed imagery is an active yet challenging topic. As the urban unit that is related most closely to the spatial distribution of population, the buildings for residential land uses are of special interests for broad applications. This paper presents a three-step approach to identify residential buildings using light detection and ranging (LiDAR) data, aerial photographs, and road maps. A multidirectional ground-filtering algorithm first separates ground from LiDAR data to produce a digital surface model, a digital terrain model, and the height of objects above ground. Then, a morphology-based building-detection method extracts buildings by gradually removing other objects (especially trees) based on the difference in the first and last returns of LiDAR data, building height, vegetation indexes from aerial photograph, and the morphological characteristics of building footprints. Finally, residential buildings are separated through the classification based on seven land use indicators: area, height and compactness of buildings, the distance to major roads, the percentage of green space and parking space surrounding a building, and building density within a block. The method was tested in an area in Austin, Texas. The results showed that the method successfully extracted buildings from LiDAR and aerial photographs and identified 81.1% of the residential buildings.