The problem of automated urban road network extraction is extremely complex, as roads in urban scenes interact strongly with other scene objects (vegetation, buildings and vehicles). This problem can be simplified if regions corresponding to roads were previously isolated using a classification procedure. Next, the urban road network can be extracted from these regions previously detected, leading to a geometric and semantic description. The image classification procedure can be used in order to isolate regions of road, but in complex urban scenes, the use of only spectral data may not be sufficient for a reliable separation of classes with similar spectral characteristics. To overcome the problem, it is proposed the integration of laser scanner geometric and radiometric data with high-resolution RGB aerial images in the classification using Artificial Neural Network, with main focus to isolate regions of road. The benefit of this integration was cheked using different combinations of input data. The experiments showed that the combination that integrates different sources of data allows the separations of the road class with better accuracy and the problems related to spectral similar responses were minimized.
Artificial Neural Network; Normalized Digital Surface Model; Laser Pulse Intensity Image
