Towards knowledge-based extraction of roads from 1 m-resolution satellite images

As the IKONOS satellite with 1 m-resolution camera was launched in 1999, mapping using spaceborne images will be an important issue in the computer vision area as well as photogrammetry, mainly because most major man-made objects of interest can be identifiable. One of the automatically identifiable objects of importance may be roads. Detecting roads using edge detection approaches may be very difficult because a number of edge elements from such as buildings, etc., can be generated from edge detector. In this paper, we propose a method for the extraction of approximated road regions based on region segmentation that utilizes region information. Our method consists of the following three steps. First, an image is segmented using the modified hierarchical multi-scale gradient watershed transformation. Then, the road candidates are identified using information about road gray level, elongatedness and connectedness. The identified road candidates are expanded by connecting the close-by roads knowing that roads are connected objects. Our method was tested on the simulated spaceborne images and the result shows that the automation of road extraction is quite promising.