Research On Radio Frequency Interference Mitigation And Land Cover Classification For PolSAR

Polarimetric synthetic aperture radar (PolSAR) can collect the backscattering information of illuminated target under different combinations of transmit and receive polarization states, which can greatly improve the capability of discriminating different targets. It shows great potential for military and civil applications such as battlefield reconnaissance, precision strike, resource exploration and environmental monitoring etc. In practical situation, PolSAR is susceptible to the radio frequency interference (RFI) emitted from the electromagnetic equipment sharing the same frequency band. The presence of RFI would lead to the degradation of imaging quality, and poses a hindrance to the accuracy of image interpretation, such as feature extraction and land cover classification etc.Since the radar echoes and interference are mixed together in the time domain or frequency domain, it remains a difficulty to detect and mitigate the RFI effectively, which is a hot topic in the area of radar signal processing. Meanwhile, PolSAR image provides interpretation of target inherent property, such as shape, orientation and dielectric property, etc. It is also a hot topic to design an appropriate feature extraction and classification scheme in the purpose of qualitative interpretation of PolSAR image. Under the grant of National Natural Science Foundation of China, the Program for New Century Excellent Talents in University, as well as the support of crosswise project, this dissertation focuses on investigating the issues of RFI mitigation and land cover classification for PolSAR data. Its main contributions are listed as follows:1. Firstly, this dissertation introduces basic operation principle of the PolSAR system, and derives the signal models of useful target echoes and interference signals. Then, the mapping relation between the radar echoes and image formation process is presented. Qualitative and quantitative analysis of the adverse impact of RFI on the simulated and measured dataset are illustrated from two perspectives, i.e.,evaluation of imaging quality and interpretation of scattering mechanisms. This section provides theoretical foundations for further study of RFI mitigation methods.2. The presence of narrow band interference (NBI) would degrade the imaging quality and pose a hindrance to image interpretation.To tackle this issue, according to the statistical difference between the NBI and useful echoes, this dissertation proposes a kurtosis-based detection and independent subspace analysis (ISA)-based mitigation scheme. Firstly, the analytical statistical distribution of radar echoes with and without NBI is derived, followed by detailed analysis of NBI’s characteristics in time-domain, frequency-domain, time-frequency domain as well as the image domain. Based on the statistical difference, an adaptive detection method using the kurtosis value is proposed. This detection method can determine the threshold adaptively according to the Neyman-Pearson criterion, and can guarantee the detection performance under rather low false alarm level. Then, a NBI mitigation method based on the independent subspace analysis is proposed. It overcomes the greatest limitation that independent component analysis (ICA) could not be used for solving the single channel underdetermined blind source separation problem. According to the statistical difference between NBI and useful echoes, independent basis signals corresponding to NBI can be extracted from the time-frequency distribution, and thus can achieve the reconstruction and subtraction of independent subspace of interference. The merits of the ISA-based method are that it not only can mitigate the effect of NBI, but also can separate out each NBI component and provides a better insight into the characteristics of NBI components. The experimental results of simulated and real-measured datasets verified the effectiveness of the proposed detection and mitigation scheme.3. Wide band interference (WBI) is highly overlapped with useful echoes in time domain and frequency domain, which makes it difficult to mitigate the WBI effectively. To tackle this issue, a WBI mitigation method based on the instantaneous spectrum Eigen subspace filtering is proposed. Firstly, two classical models for WBI are established, and detailed analysis of their appearances and characteristics in time-domain, frequency-domain, time-frequency domain as well as image domain is provided. In time-frequency distribution plane, the interference and useful echoes differ in their concentration characteristic. Based on this difference, a WBI mitigation method based on the instantaneous spectrum Eigen subspace filtering is proposed. For each instantaneous spectrum, the existence of interference signal can be identified according to the negentropy-based statistical hypothesis test. The threshold is determined according to the false alarm level. Then, the interference signal is mitigated by Eigen subspace filtering, which minimizes the signal distortion introduced by the mitigation scheme. The experimental results of the simulated data, as well as real measured data sets, show that the proposed scheme is effective in suppressing the WBI and in obtaining a high-quality image.4. To deal with the problem of effective feature extraction in PolSAR land cover classification, a novel feature extraction scheme based on the tensorial independent component analysis is proposed. Various polarimetric parameters can be obtained by target decomposition techniques, and multi-feature combination is of great help for characterizing the land cover and improving the classification accuracy. Since some of the features provide similar interpretations of the scattering mechanisms, there may have a lot of redundancy and correlation among them, which would make the subsequent classification a difficult task. Common feature extraction methods use matrix linear algebra and require rearranging the original tensor into a matrix. This leads to the loss of the spatial information of the PolSAR data. Within the tensor algebra framework, the whole PolSAR image is formulated as a third-order polarimetric feature tensor by employing multi-features combination. By formulating the matrix-based feature extraction methods using tensor representation, the shortcomings of the traditional methods is analyzed, and then a novel feature extraction method based on tensor algebra is proposed. The proposed scheme consists of twofold processing:the low-rank approximation of the spatial modes and the dimensionality reduction for the feature mode. By jointly considering the overall redundancy between the spatial and feature mode, the optimal feature extraction matrices are solved in the alternate least square manner. The extracted features by the proposed method are less redundant and more discriminative, exemplified by the great improvement of classification accuracv of the simulated data and real measured datasets5. Aimed at improving the classification accuracy of PolSAR image, a classification scheme based on the multilinear principle component analysis (MPCA) is proposed by making full use of the correlation among local neighboring pixels. Firstly, based on the multi-feature combination, every pixel of the PolSAR image is modeled as a third-order tensor involving its local neighboring pixels. The training tensor samples are determined according to a priori information. MPCA determines a multilinear projection onto a tensor subspace of lower dimensionality that captures most of the variation present in the original tensorial representation. The projection matrices is optimized from the training set in an alternative least square manner, and thus improves the discriminability among different classes by considering the neighboring spatial information. Unknown labeling tensor samples can be extended by the optimal projection matrices. The classification results of both the simulated and real measured datasets demonstrates that the proposed classification scheme could greatly improve the classification accuracy even with small number of training samples.