Study on the extraction of urban roads from high-resolution remotely sensed imagery based on the knowledge of road features

This paper presents a novel road-extraction method focusing on a road network in an urban central area. The method introduces the knowledge of road features into the extraction process and makes full use of spectral and spatial context relationships and geometric information, thus successfully discriminates roads and spectrally similar buildings and solves the problem of urban roads inconsistent morphology in the imagery. We adopt a Decision Tree model to extract the raw roads information based on the spectral knowledge of pure pixel signatures. Then an "Eliminate & Growing" algorithm is developed based on the context spatial relationships to make the roads independent and filled and reduce the "salt and pepper" effects. Next, we retrieve more accurate road information in vector format in terms of the road's geometric characteristics. Moreover, we manage to retrieve the hidden roads blocked by the trees via utilizing the information of wayside trees. And finally we use mathematical morphology to form the road network. This method has successfully extracted all the main and sub-main roads in the study area; the result has demonstrated the method's high accuracy and usefulness in practice.