An automated approach for constructing road network graph from multispectral images

We present an approach for automatically building a road network graph from multispectralWorldView II images in suburban and urban areas. In this graph, the road parts are represented by edges and their connectivity by vertices. This approach consists of an image processing chain utilizing both high-resolution spatial features as well as multiple band spectral signatures from satellite images. Based on an edge-preserving filtered image, a two-pass spatial-spectral flood fill technique is adopted to extract a road class map. This technique requires only one pixel as the initial training set and collects spatially adjacent and spectrally similar pixels to the initial points as a second level training set for a higher accuracy asphalt classification. Based on the road class map, a road network graph is built after going through a curvilinear detector and a knowledge based system. The graph projects a logical representation of the road network in an urban image. Rules can be made to filter salient road parts with different width as well as ruling out parking lots from the asphalt class map. This spatial spectral joint approach we propose here is capable of building up a road network connectivity graph and this graph lays a foundation for further road related tasks.