A NOVEL APPROACH TO ROAD DETECTION FROM SATELLITE IMAGES
Humans easily identify roads in remote sensed images, but this task is difficult to automate using computers. In order to identify road segments from satellite images, human beings (experts) seem to first search for a set of linear and curvilinear features and then apply knowledge or use experience to decide (or guess) whether these linear and curvilinear features are roads are not. A automatic hybrid road detection method is proposed, based on the perception of the human beings and takes the advantage of both statistical (Gaussian Mixture Model (GMM) based method) and artificial neural network techniques.
In the proposed method, a satellite image is processed in four stages to obtain road network. In the first stage, the edges are extracted from a satellite image using an efficient method of 1-D processing. The advantages of using 1-D processing, when compared to 2-D operators are: (i) computational efficiency, (ii) weaker edge segments which are necessary for proper road extraction are detected and (iii) it is less sensitive to noise. In the second stage, this edge map is post-processed to eliminate isolated edge pixels and link discontinuous edge segments, using a set of binary morphological operations. In the third stage, a set of linear and curvilinear segments are extracted from the post-processed edge map using a GMM based extractor. In the final stage, a supervised feedforward neural network is used to identify the road segments.
Two approaches: (i) multi-resolution and (ii) iterative are proposed to improve the results of the GMM based method. The set of linear and curvilinear segments resulting from the GMM based method include edge segments which belong to field boundaries, river banks, roofs of buildings etc., along with road segments.
A feedforward neural network is trained using statistical features obtained from spectral(R,G,B)/gray-level(one band) and edge orientation values. The features are extracted over a small local window around the pixels which are candidates for road segments.
Satellite Images
Reference Road Networks
Edge Maps Using 1-D Processing
Results Using GMM Based Extractor
Orientation Maps
Training Samples
Results of the Proposed Method
For comments, mail to sdas@iitm.ernet.in ,
koshy@iitm.ac.in