Research On Adaptive Clutter Suppression And Target Detection In Heterogeneous Environment

The fundamental mission of radar is to decide the presence of the interested target in the scene.The typical way to perform this task is to transmit electromagnetic signal and process the received echo.Generally,the return signal not only contains the possible target signal,but also contains signals from other scatterers and the thermal noise.The signals from other scatterers are referred to as clutter and the target signal is usually embedded in clutter.As a consequence,the clutter suppression is normally required before target detection in order to improve the signal-to-clutter-ratio of the target.The clutter is often complicated and varying,so the adaptive clutter suppression technique is often exploited to mitigate the clutter effectively.The adaptive clutter suppression technique often requires a certain amount of homogeneous training data to estimate the clutter statistical property in real time.However,the realistic environment is normally heterogeneous.As a consequence,how to adaptively suppress the clutter and detect the target in heterogeneous environment is an important research direction in the field of radar.This dissertation focuses on this direction and the main contribution is summarized as follows:1.Investigate the effect of heterogeneous environment on adaptive clutter suppression and target detection.This dissertation takes STAP as an example and analyzes the effect of amplitude heterogeneity,spectrum heterogeneity and moving target in training data.The results show that the factor which has the most severe effect on the STAP performance is generally the target-like signal in training data,which means the signal sharing similar angle and Doppler with the true target.2.Study the method of removing heterogeneous samples according to the generalized likelihood function criterion.The heterogeneous sample is composed of homogeneous component and non-homogeneous component,where the homogeneous one shares the same statistical property.When the non-homogeneous component is unknown,we have devised the maximum likelihood estimate of the set composed of indices of heterogeneous samples according to the generalized likelihood function.Based on the estimate,we have proposed the maximum likelihood(ML)method for censoring outliers.However,the computational load of the ML method is heavy.To circumvent this problem,we have proposed the approximate maximum likelihood method.We have also considered the case when the non-homogeneous component is target-like signal and derived the corresponding ML estimate of the outlier subset under the assumption that the homogeneous component is complex Gaussian distributed with mean zero and covariance matrix known.The objective function of this ML estimate is coincident with the adaptive power residue criterion.However,the traditional method based on the adaptive power residue criterion often replaces the true interference covariance with the sample covariance matrix,which is sensitive to target-like signals.In order to circumvent this problem,we have proposed the reiterative censored normalized adaptive power residue algorithm.3.Study the method of censoring heterogeneous samples according to the regularized generalized likelihood function.The ML method and its approximate algorithms require that the number of homogeneous samples should be no less than the system degrees of freedom,which might be hard to meet in practical applications.In order to deal with the scenario with limited homogeneous samples,we have proposed to enforce a regularization constrain on the generalized likelihood function.We have derived the regularized maximum likelihood of the outlier subset and put forward the regularized maximum likelihood algorithm for excising outliers.Moreover,to reduce the computational burden of the RML method,we have proposed three approximate algorithms.The performance of RML and its approximate algorithms relies on the selection of the regularization parameter.We have considered two ways to select the regularization parameter: expected likelihood and cross validation.The proposed techniques can effectively select proper regularization parameter and when the number of training data is limited,the RML and its approximate algorithms outperform the ML and its approximate algorithms.However,when the number of training data is sufficiently high,the two kinds of algorithms almost share the same performance.4.Study the techinque of SR-STAP.SR-STAP generally exhibits a good performance with 4 to 6 homogeneous samples,so it is very suitable for the scenarios with limited training data.However,the main problem of SR-STAP is that its computational burden is significantly heavy.In order to circumvent this problem,we have proposed a reduceddimension SR-STAP method based on the power spectrum.Specifically,this technique firstly decides a region in the space-time plane composed of possible clutter scatterers according to the power spectrum and then derives the corresponding clutter amplitude using the sparse recovery method.After that,the clutter covariance matrix is obtained.The reduced SR-STAP method can reduce the computational load significantly as compared with the plain SR-STAP with small performance loss,as a result,it is more suitable for the real-time processing.SR-STAP also requires selecting homogeneous training samples in heterogeneous environment.We have proposed to select the training data accroding to the locations of clutter scatterers and distance between clutter covariance matrix,respectively.Since the proposed methods have exploited the sparse property of training data,their performances are superior.5.Study the design of tunable detectors based on the sparse property of target.Target detection is the ultimate task of adaptive clutter suppression.The main problem of traditional detectors is that the strong targets or coherent interferers might also trigger a detection and result in a false detection.The problem can be solved with the information of target azimuthal position.Based on the sparse property of target in the azimuth dimension,we have proposed to obtain the target amplitudes at different angles using the sparse recovery method.Exploiting the sparse amplitude,we have proposed two types of decision architectures: one is based on the two-stage paradigm and the other is based on the likelihood theory.By changing the interval between angles,the proposed detector can achieve a tradeoff between the detection performance of matched targets and suppression of mismatched signals.