A new method is presented to extract road networks in urban area from high-resolution satellite images. The first step is
to compare the consistency of gray value in different directions, so that the most consistent direction at every pixel is
found. The texture descriptions in this direction consist of the feature vector of segmentation. An improved MRF model
combined with the Expectation-Maximization algorithm is then applied to the feature space to separate roads from other
objects unsupervisedly. After removing the irregular nonroad speckles in the segmented image, road segments may be
detected easily. Connecting road segments finally produces the road networks. Experiments demonstrate that the
proposed method may extract networks efficiently.