An integrated system for automatic road mapping from high-resolution multi-spectral satellite imagery by information fusion

Abstract

Here we present an integrated system for automatic mapping of urban and suburban roads from high-resolution satellite imagery. Road extraction strategies can be enriched and improved by multi-source data fusion. Fusion can occur by combining road extraction results from different image sources, or by applying multiple detectors to a single source image and then fusing the multi-detector road extraction results. Our proposed system models roads differently for urban vs. suburban environments. In suburban areas, roads are modeled as curvilinear and homogeneous regions with nearly constant width. Roads are extracted by integrating the output of two road detectors followed by a road tracker. The first detector is based on fine-scale image segmentation and grouping where road centerlines are identified and extracted using shape and structure information from the segments. This detector generates a very reliable road extraction but is incomplete as many road segments are not captured. A multi-scale curvilinear structure detector based on differential geometry is used as a complementary detector. This detector utilizes a fuzzy decision algorithm based on a 1D road profile model. The final suburban road centerlines are generated by integrating the results from the two detectors using an optimum path search algorithm. In many dense urban areas in the US, road networks are composed primarily of straight lines that form a grid structure. Here we use directional morphological filtering to mask out dark and bright structures shorter than a city block. We then use a spatial signature weighted Hough transform to generate a road grid hypothesis. The modified Hough transform incorporates spatial information derived from surrounding image pixels and can be applied to multi-spectral images. Each piece of the hypothesized road grid is then verified using homogeneity, shape, and vegetation information from the local surrounding area. Missing pieces of the road network are added using a road tracker based on profile matching. The integrated road extraction system based on the fusion of multi-detector results is tested using IKONOS multi-spectral imagery of the City of Columbia, Missouri. Evaluation of the extracted road networks using representative test sites show completeness values that range between 70% and 86% and correctness values that range between 70% and 92%.