Research On Polarimetric SAR Imagery Filtering And Segmentation

Polarimetric Synthetic Aperture Radar (PolSAR) compared with single polarization SAR can obtain the fully polarimetric information contained in the backscattering electormagnetic waves of the terrain targets. The polarimetric information can be recorded in the form of complex scattering matrices, with which the scattering characteristic of the targets can be extracted and analyzed. At present, the airborne and spaceborne PolSAR systems are becoming more and more, obtaining more and more PolSAR data sets. However, compared with the rapid development of hardware systems, the research on PolSAR imagery analysis and automatic interpretation is relatively lagging behind. This is partly due to the extreme complexity of the earth’s surface scene, which makes it very difficult to completely understand the polarization scattering mechanism of all terrain targets in the short term. There are still many aspects of the PolSAR theory that needs further research. On the other hand, the special SAR imaging mechanism inevitably generates speckle noise, which disturbs the real information of the features in the PolSAR imagery, and greatly increases the difficulty of the PolSAR imagery interpretation.PolSAR imagery filtering and segmentation is an important means of speckle noise suppression. As the important preprocessing steps of pixel-based and object-oriented PolSAR imagery interpretation methods, PolSAR imagery filtering and segmentation have important significance to precisely and quickly interpret PolSAR imagery. Aiming to accurately restore the true information of the PolSAR imagery, this dissertation took a deep research on PolSAR imagery filtering and segmentation.In terms of PolSAR imagery filtering, this dissertation systematically summarized the existing PolSAR images filtering algorithms, divided them into four categories according to the technologies used by these algorithms:filters based on homogeneity measurement, filters based on probability distribution assumption, filters based on patch matching, and filters based on the other technologies, and pointed out the problems existing in the four categories of filtering algorithms, which laid a foundation for the subsequent research. This dissertation then summarized the basic principles and methods of PolSAR imagery filtering. On this basis, an adaptive-window filtering algorithm and a generalized balloon mean shift (GBMS) filtering algorithm were proposed. To evaluate the proposed filtering algorithms, this dissertation systematically summarized the existing qualitative and quantitative evaluation methods for PolSAR imagery filtering, and divided them into two categories: full-reference map methods and no-reference map methods. To improve the existing evaluation methods, this dissertation proposed an edge-preservation-index-based quantitative evaluation method and a polarization-target-decomposition-based quantitative evaluation method, building a roughly complete evaluation system of PolSAR imagery filtering algorithms. At last, this dissertation proved the effectiveness of the proposed PolSAR imagery filtering algorithms by using simulated and real PolSAR data.In terms of PolSAR imagery segmentation, the existing algorithms have great difference in the form of the segmentation results, but their names are vague. This dissertation divided the PolSAR image segmentation algorithms into three categories according to the difference of the segmentation results:conventional segmentation, superpixel segmentation, and hierarchical segmentation. The differences and relationships between the three types of segmentation methods were clearly expounded, sorting out the train of thought for follow-up studies. This dissertation then proposed a generalized statistical region merging (GSRM) algorithm, and used it to PolSAR imagery conventional segmentation and superpixel segmentation; extended the proposed GBMS algorithm for PolSAR imagery superpixel segmentation; proposed a superpixel-based binary patition tree (BPT) segmentation algorithm by integrating the superpixel segmentation method and the hierarchical segmentation method. To evaluate the proposed segmentation algorithms, this dissertation summarized the existing evaluation methods for PolSAR imagery segmentation. Since there is still no mathematical definition of the important concept of "segmentation scale", this dissertation proposed the mathematical definition of segmentation scale, and further put forward a segmentation result evaluation method based on segmentation scale. At last, this dissertation demonstrated the effectiveness of the proposed PolSAR imagery segmentation algorithms by using real PolSAR data.The main contributions of this dissertation are as follows:1) To solve the problem that the existing homogeneity measurements cannot distinguish the homogeneous areas and heterogeneous areas very well, and the threshold is difficult to determine, this dissertation defined a new polarimetric homogeneity measurement and proposed a threshold automatic selection method. To overcome the scallop effect and the false lines problems of the refined Lee filter, this dissertation proposed an adaptive-window filtering algorithm, which can adaptively select the size and shape of the filtering windows, making it possible to preserve the detail information in heterogeneous areas while suppressing the speckle noise in homogeneous areas.2) To avoid the problem that the probability-distribution-assumption-based filtering algorithms may have different filtering effects for different type of PolSAR imagery, this dissertation proposed to use mean shift algorithm to filter PolSAR imagery, because mean shift is a non-parameter probability density estimation method. The adaptive, asymmetric, variable bandwidth mean shift is derived and the bandwidth selection method is proposed according to the characteristics of the PolSAR data. The proposed GBMS algorithm can be applied directly to single-and multi-polarization SAR imagery, and the noise suppression effect is superior to the conventional probability-distribution-assumption-based filtering algorithms.3) To introduce the SRM algorithm into PolSAR imagery segmentation, this dissertation established a new SRM model according to the multiplicative model of SAR data, derived a new merging criteria, and finally proposed the GSRM segmentation algorithm. This algorithm does not depend on the probability distribution of SAR data, has good noise immunity, and can be applied to single-and multi-SAR imagery directly without any preprocessing steps. AIRSAR L-band data proved that the GSRM algorithm can get more accurate segmentation results comparing with the original SRM algorithm.4) The proposed GSRM algorithm and the proposed GBMS algorithm were further extended for superpixel segmentation of PolSAR imagery. Both these two new superpixel segmentation algorithms are superior to the commonly used Ncut algorithm in both execution efficiency and implementation effects. This dissertation then proposed a superpixel-based BPT segmentation algorithm by integrating the superpixel segmentation method and the hierarchical segmentation method. The proposed algorithm has obvious improvement in execution efficiency compared with the pixel-based BPT segmentation algorithm. In terms of implementation effect, they are similar when using the region-number-based pruning strategy, and the former is superior to the latter when using the homogeneity-based pruning strategy.Although this dissertation has obtained certain achievements on the basis of previous researches, there are still many aspects worthy of further research and discussion, such as the removal of the scar effect of the adaptive-window filtering algorithm proposed in this dissertation, scar effect’s influence on the subsequent processing, and the contradiction between the use of the polarization information and the multi-look preprocessing.