| With the rapid development of polarimetric synthetic aperture radar(PolSAR)system,more and more high-resolution images can be acquired.They can describe more detailed information of land-cover types,while the pixel complexity is also increased.Traditional image processing methods,which take image pixels as basic processing unit,can only use pixel-level image features,and are no longer suitable for high-resolution images.Therefore,focusing on OBIA technology,this paper studies the object construction of PolSAR images,the object-based image classification and buildings recognition methods.First,in view of the problem that it is difficult to select appropriate spatial scale for PolSAR image segmentation due to the multiple catogories of land-cover types,this paper proposes an adaptive spatial bandwidth selection method based on statistical semivariogram for mean shift(MS)algorithm.Based on the analysis of the polarimetric scattering process for different land-cover types,the spatial bandwidth parameter is adaptively selected in mean shift algorithm by incorporating the mathematic statistical semivariance for PolSAR image segmentation.The experimental results show that the proposed spatial bandwidth selection method based on the combination of polarimetric and statistical image characteristics can adaptively select the optimal spatial analysis scale without manual participation,which effectively improves the accuracy of PolSAR image object for object-based image interpretation.Second,with the increase of PolSAR resolution,the complexity of image processing is increased.Conventional statistical distribution models are mainly used in low or median resolution images,and are no longer suitable for high resolution images.Based on the construction of PolSAR image objects,this paper proposes an object-based markov random field model with multiscale polarimetric auxiliary label field(PA-OMRF).The polarimetric auxiliary label fields are introduced to describe the polarimetric information of objects in different scales.Based on the interaction and iteration of label fields,the PA-OMRF model is established to realize the classification of PolSAR images.Experimental results show that the proposed PA-OMRF model can effectively describe image statistical information and take advantages of the spatial semantic features between image objects to improve the image classification performance.In the meantime,image objects have rich characteristics of polarimetric,spatial,semantic and topological features.Aiming to solve the problem that the conventional object-based image analysis cannot make full use of the abundant features of image objects,this paper proposes a supervised classification method which combines the spatial and semantic features of image objects for high-resolution PolSAR image classification.The spatial and semantic features are firstly extracted based on the construction of the hierarchical image object network.Then,the spatial-semantic model is constructed to realize the comprehensive utilization of spatial and semantic features of image objects for PolSAR image classification.The experimental results show that the comprehensive utilization of multiple object features can effectively improve the accuracy of PolSAR image classification.Finally,due to the scattering characteristics of PolSAR,buildings often present similar characterics with other targets such as forest areas,which decrease the extraction accuracy.This paper puts forward an object recognition framework for buildings recognition in PolSAR images.Based on image object generation and multiple features extraction,the proposed method simulates human image cognition process and establishs the "object perception-logicical reasoning-cognitive decision" framework,which combines the fuzzy logical mathematics,neural network theory and human knowledge.Experimental results show that the proposed object recognition model can effectively distinguish building targets from other target types in complex background,which improves the extraction accuracy of building targets and reduces the false alarm rate. |
