Building Extraction From High Resolution SAR Imagery

With the demand on urban SAR image interpretation, techniques of building extraction from high-resolution SAR images are investigated in this thesis. Aiming at dealing with high-resolution and large-scene images in practical applications, a hierarchical procedure for building extraction is proposed. First, the built-up areas are detected from the large-scene image; then, building detection is carried out in the detected built-up areas; finally, the geometrical information of the buildings is extracted. According to this procedure, this thesis thoroughly studies the techniques of built-up area detection, building detection and geometrical information extraction. The main work includes the following aspects.(1) The phenomena of buildings imaged from a SAR sensor is analyzed and simulated. The main scattering mechanisms encountered in built-up areas, the typical imaging effects of buildings and their mapping relationship to a SAR image are analyzed. Then the characteristics of buildings in SAR images are concluded and validated by simulated images. All the analysis and simulation is the base of the subsequent research. Especially, simulation of building SAR images provides the testing data for the study of geometrical information extraction.(2) According to the requirement of providing ROI for building detection, an effective and efficient method of detecting the built-up areas from SAR images using the variogram texture feature is proposed. First, the variogram texture feature can describe the strong dissimilarity of the built-up areas in SAR images, and thus discriminates the built-up and non-built-up areas. Second, according to the different variogram curve patterns of the built-up and non-built-up areas, the best texture lag for the binary classification can be easily determined. Third, the variogram texture feature can be computed efficiently. Based on the theoretical analysis of the redundant computation, a fast recursive algorithm for computing variogram is designed, which makes built-up areas detection more practical and can meet the application demands of providing ROI for building detection.(3) Since the existing methods of building detection from SAR images are not robust for images with complex scene or different appearances of buildings, a method of building detection using the marker-controlled watershed transformation is proposed, aiming at detecting buildings with their whole and accurate boundaries from the built-up area. By introducing the marker-controlled watershed transformation, this method can make use of not only the characteristics of the building, which are strong scattering and high gray values, but also the characteristics of the surrounding background, which are the black netlike structures formed by roads and shadows. The combination of the characteristics of buildings and background can overcome the problems of linking neighboring buildings in complex scene or dividing a building into several parts when its gray values fluctuate greatly. Besides, the new method can get the closed boundaries of the buildings. Since the ROEWA edge detector, an edge detector for SAR images with good localization performance, is used, the detected building boundaries are also accurately localized. Furthermore, according to the typical shapes of the building in SAR images, a shape analysis method called direction-correlation analysis is proposed to remove the false alarms. The experiments invalidate that the new method is effective with high detection rate, low false-alarm rate and good localization performance. The detection results can be used in the process of extracting the buildings'geometrical information.(4) Since the existing methods of extracting buildings'geometrical information have defects in accuracy or practicality, a new frame of geometrical information extraction is proposed via matching the geometrical model of a building and the real SAR image. On one hand, the frame maps the model into the image to determine the typical regions such as the layover and the shadow, which contain the geometrical information of a building. On the other hand, with the features such as the building boundary obtained from the building detection, the best model parameters can be found by matching the model and the detected features. The new frame is a general one, because the building model, the typical feature regions and the methods of detecting these regions can be adjusted in different applications. An implement of the frame is carried out using the common flat building as an example. First, the method of mapping a building model into the layover and the shadow regions with different imaging conditions are discussed. Second, the matching functions of the layover boundary and the shadow boundary are designed, and they form the whole matching function. Third, the algorithm of estimating the model parameters using the genetic algorithm is given. When the matching function reaches its maximum, the corresponding parameters are the best estimation for the building geometrical information. The experiments of both the simulated and the real images show the accuracy and practicality of the new frame: the use of the mapping relationship of the geometrical model to the image ensures the accuracy; the transform of extracting the geometrical information to maximizing the matching function and the automatic parameter estimation ensure its practicality.