A Probabilistic Fusion Concept for Road Extraction from Multiple SAR Views

Abstract:
In this article, a probabilistic fusion concept for road extraction from multi-aspect SAR images, which incorporates sensor geometry and context information, is proposed. Before fusion, the uncertainty of each extracted line segment is assessed by means of Bayesian probability theory. This assessment is performed on attribute-level and is based on predefined probability density functions learned from training data. In the first part the importance of global and local context information and the benefit of incorporating sensor geometry within the fusion module are discussed. The second part concentrates on the analysis of the uncertainty assessment of the line segments. Finally, some results regarding the uncertainty assessment of the line segments using real SAR images are presented.