Research On The Land Cover Classification For Multi-polarization Sar Images

In recent years, with the rapid development of synthetic aperture radar (SAR) technology, SAR data source has shown many new characteristics, such as multi-polarization and multi-band, and multi-temporal. Multi-source SAR images indicate that the different properties of land cover from different point of view. So it is helpful to improve the accuracy and feasibility of SAR classification by using the image information from different sources. Multi-source SAR data processing has become one of the hotspots in the research of microwave remote sensing.Researching on the land cover classification for multi-polarization SAR images, along with utilizing the image processing and analysis, statistic analysis, pattern recognition and information fusion, this thesis addresses the fusion techniques of the multi-polarization SAR images in aspects of the speckle suppression, multi-polarization SAR images segmentation, texture feature extraction and the classification fusion. The main contents are summarized as follows:An adaptive edge-preserved speckle suppression algorithm is proposed. In order to improve edge-preserved which is a shortage of the classic statistics filtering algorithm, the algorithm distributes weights for the convolution mask based on three information of SAR image: relative deviation, distance and edge discontinuity. Experimental results of the speckle suppression show that the improved algorithm has better performances in both speckle smoothing and edge protecting.A new ratio of averages (ROA) calculation method is improved to yield the gradient of SAR image, and a fusion segmentation algorithm for multi-polarization SAR images is proposed based on the new ROA. The new ROA method can obtain the exact gradient of every direction by using four sub-windows to produce six kinds of average-pair to compute gradient. The segmentation algorithm is separated into three parts: gradient computation, watershed segmentation and region merging of over segmentation. In the gradient computation and region merging steps, all of the SAR images are applied. Experiments show that the fusion algorithm based on the new ROA can get exact region segmentation.Two kinds of texture feature extraction method for the segmentation region are presented. One is based on stationary wavelet analysis theory, and the other is based on wavelet transform and texture spectrum analysis. Combining the results of the segmentation, the energy texture feature of every region of one SAR image is extracted to every sub-image of stationary wavelet decomposition. And then all of the SAR image region feature yield the finial feature set. In addition, based on the texture spectrum analysis, one kind of rotation invariant texture is researched by using eight texture unit and local quaternary pattern to describe the microcosmic content of texture from eight directions. Combining stationary wavelet analysis and texture spectrum analysis, texture spectrum feature can be extracted from the sub-images of wavelet.A feature-level fusion classification algorithm based on polarization-choose is proposed for multi-polarization SAR image. The fusion principle on feature-level is established by studying the characteristic of multi-polarization images. The K-means clustering algorithm is applied as the classifier. And adaptive threshold strategy and feature-weighted strategy are adopted. The classification results of multi-polarization SAR images can obtain height classification ration, and indicate the fusion algorithm is effective.