Study Of Knowledge-Aided Clutter Covariane Matrix Estimation In STAP

Space-time adaptive processing(STAP) attracts a lot of attention due to its ability in improving effectively the performance of clutter suppression and target detection in airborne radar. A crucial implementation in STAP is estimating the clutter covariance matrix(CCM) in the cell under test(CUT), whose methods may be divided into two categories. On kind of technique exploits the statistical property of clutter and accomplishes estimation according to some criteria such as maximum likelihood principle, with the help of some training data which is independent and identically distributed(i.i.d) with the data in CUT. Another kind of technique takes advantage of the clutter structure property and finishes estimation by employing the clutter model and estimating the parameters in the model.The first kind of method can acquire good performance under the assumption that there are enough homogeneous training data. However, this assumption is always violated in realistic environment since the environment is always heterogeneous and using the heterogeneous training data to estimate the covariance matrix will cause estimation error, and thus lead to the degradation of detection performance. The second kind of approach can obtain good performance under the condition that the clutter model matches well to the real clutter covariance matrix and all the parameters are approximately estimated accurately, which is always computationally complicated. Researchers find recently that the performance of clutter suppression can be improved dramatically by exploiting the prior knowledge and accomplish intelligence information processing. In this background, this paper focuses on the study of knowledge-aided clutter covariance matrix estimation in STAP.Chapter two primarily analyzes the properties of the clutter covariance matrix, which is composed of characteristic spectrum, power spectrum and effects of realistic factors. Simulation results show that realistic factors including channel error, internal clutter motion and airplane crab can increase the degree of freedom(DOF) of clutter and widen or deform the power spectrum.Chapter three proposes a novel approach for training data selection, which is based on the geometrical property(the distance between covariance matrices). A lot of distance criteria containing Euclidean distance, Riemannian distance, spectrum distance, physical Euclidean distance and physical spectrum distance has been compared and three kinds of methods calculating the distance has been considered, which include the distance between adjacent training data covariance matrices, the distance between training data covariance matrix and sample covariance matrix and the distance between training data covariance matrix and knowledge-aided covariance matrix. Simulation results indicate that using Riemannian distance physical Euclidean distance and physical spectrum distance between training data covariance matrix and knowledge-aided covariance matrix can select training data effectively.Chapter four centers on the clutter covariance matrix estimation based on the prior SAR images. Under the assumption that the scattering properties of clutter patch does not change with the azimuth angle, the covariance matrix estimation using perfect SAR images possesses small estimation error and good detection performance. However, the scattering properties of some strong discretes such as high-voltage telegraph pole, bridge and household change dramatically with the change of azimuth angles, which violates the assumption. In addition, due to the effect of weather and climate, the scattering properties of SAR images may be not identical to the true scene. To solve these two problems, the paper proposes using sub-look SAR images to acquire scattering properties in some specific angle and uniting the SAR images and training data to carry out “color diagonal”, respectively. Simulations verify the effectiveness of the proposed approach.