Road Extraction In High-resolution SAR Images

As the typical man-made land object, roads are essential parts of moderntransportation system, which have important geographical, political, economic andmilitary values. As a kind of microwave remote sensing system, Synthetic ApertureRadar (SAR) data acquisition could operate during both day and night, and isindependent on the influence of sunlight and clouds. Becaese of these merits, roadextraction from SAR images has attracted attentions of researchers all over the world.More land object details can be described in high-resolution SAR images compariedwith which in low-resolution images. Ideally, roads may be modeled as dark elongatedareas surrounded by pairs of parallel bright edges in high resolution SAR images.However, the successive road areas are frequently broken by obstacles and shadows,such as vehicles on the road, trees along the road, building shadows covering the road,etc. Road extraction is still a difficult task in high resolution SAR.Considering the difficulties of current road extraction methods and according to theradiate and geometric properties, the thesis lucubrates on the automatic andsemi-automatic road extraction problems respectively, and automatic road junctions’extraction problems, which can be suitable for different applications.In the research on the automatic extraction of road junctions from high resolutionSAR images, a new method is proposed for directly detecting and identifying roadjunctions. Firstly, based on the junctions gray feature, global searching is done for thecenter positions of the road junctions’ candidate regions, by using morphologicaltransformation methods. Secondly, the center positions are set as the local windowscenters, road targets are segmented by using the multi-threshold Otsu method in thelocal windows. Thirdly, according to the geometric characteristics of junctions, weobtain the angle-mean figure in the rectangular template rotation process, and then getthe number of the roads connected to a junction. Finally, the style of the junction isrecognized. In1m high-resolution airborne SAR image experiment, the results indicatethat this method is effective to detect and identify the junctions with variousinterferences.In the research on the automatic extraction of road networks from high resolutionSAR images, Markov random field (MRF) model can make full use of the imagerycontextual characters and priori knowledge, which have been widely used to extractroad networks. However, there exist some problems such as slow solution and manyparameters setting of these type methods. In order to reduce the computation ofsubsequent iterative solution of MRF, pre-linking is firstly introduced to removenumerous false line elements based on the spatial relationship among them. Then, theimproved road networks Markov function model is established to label road networks. SAR images with1meter resolution are tested in the experiment. And the results showthe effectivity of the method mentioned above in high resolution SAR imagery roadnetwork extraction.In the research on the semi-automatic extraction of road junctions from highresolution SAR images, a new road center-point extraction method is proposed bycombined using local detection and global tracking. In local detection phase, twowindows are set. The outside window is used to obtain the local road direction by usingnon-linear structure tensor, based on the fact that fences, green belts and otherdisturbances point to a consistent direction with the edge of roads. The inside windowadjusts its direction by the result of non-linear structure tensor. Then it searches for theroad areas, and determines the width and center of roads. In global tracking phase,particle filter of variable-step is used for solving the problem of tracking brokenfrequently by occlusions on the road and shadowing alongside the road. In1mhigh-resolution airborne SAR image experiment, the results indicate that this method iseffective.As mentioned above, this dissertation has explored the methods of road extractionin high-resolution SAR images. Also, these methods would be helpful to enginerringapplications posterior.