Learning to Recognise Roads from High Resolution Remotely Sensed Images

Automatic Road Extraction from remotely sensed images is a fundamental step in the acquisition and maintenance of geographical databases. This paper proposes an automatic road recognition algorithm based on fusion of junction and segment information. Road segments and junctions in an image are independently acquired as features of edge pairs. Then, the classification decisions of the segment and junction classifiers are fused to learn initial seeds for the extended fast marching level set method. The decisions of segment classifiers and fast marching level set method are then combined to improve road extraction. The primary contribution of our approach is its ability to learn the seed points and the stopping criterion for fast marching level set methods. Experimental results on remotely sensed image datasets demonstrate the validity of the proposed algorithm.