RAIL : Road Recognition from Aerial Images Using Inductive Learning

Roads, together with buildings, are one of the major man-made surface features of interest in remote sensing imagery. Most knowledge-based road recognition systems use a priori heuristic rules to enable recognition. These rules place explicit constraints on road properties and scene content. However, in remote sensing imagery, scene content varies, with roads taking on a wide variety of geometrical, radiometric, topological and contextual properties. The limitation of using a priori rules is that the systems are restricted to the kind of imagery for which these explicit constraints are valid. These systems typically fail on imagery outside of these conditions. In this paper, we propose an adaptive and trainable road recognition system. This s ystem i s able to perform semi- automatic e xtraction of roads from aerial im agery, using inductive learning techniques from t he field of Machine Learning within Artificial Intelligence. The system is called RAIL, for Road Recognition from Aerial Images using Inductive Learning. Instead of using a p riori rules, RAIL uses s upervised, multi-level l earning to derive rules that may be applied du ring feature e xtraction and recognition in images. This technique is quite general, making few assumptions about t he scene or its contents, and applicable to images of different scales, content, complexity and quality. RAIL has been tested on reasonably complex high resolution aerial scenes and has achieved encouraging results.