A Hybrid Road Identification System Using Image Processing Techniques and Back Propagation Neural Networks
Ko Kuo-Tu · 1995
Road identification of remotely sensed satellite images have been used for many different purposes, e.g., military, map publishing, and census, etc. In this paper, I present a method which combines traditional image processing techniques with backpropagation neural networks to detect roads in a Landsat TM satellite image. This image is first processed to extract all the line features which are the candidates of roads, then the supervised neural network is employed to tell whether a candidate road pixel is really a road pixel or not by its spectral characteristics. Furthermore, some postprocessing steps are taken to refine the result. The resultant image showed this hybrid identification system performs better than using only image processing techniques or only neural network techniques. If we are interested in a particular kind of road (i.e., interstate, urban area road, bridge, etc.) we may construct specialist detectors; this is also discussed in this paper.