Research On Techniques Of Moving Target Detection And Recognition For Synthetic Aperture Radar In Complex Electromagnetism Environment

Synthetic aperture radar(SAR)is one of the most popular radar systems in recent years.No matter the airborne SAR systems or the spaceborne SAR systems generates the highresolution SAR image of the interested area with long-time synthetic aperture.Then the power of the ground moving targets can be accumulated during this period and a high signalto-noise ratio(SNR)can be achieved at the receiver,which brings another feasible way to detect and recognize the moving target.Nowadays,the main tasks of SAR systems focus on high-resolution SAR images,moving target detection in interested areas,and target recognition.However,in order to realize the aforementioned functions,as a wideband radar system,SAR systems can be affected by complex electromagnetic enviroments,like strong interferences,strong clutters,and strong speckle noises,in their frequency bands.The narrowband interference,like radio-frequency interference(RFI),is one of the most common interferences,because communication systems,TV signals,and other wireless systems are almost everywhere in cities and villages,which will extremely affect the SAR imaging and the following moving target detection.After suppressing the complex electromagnetism environment,similar to the traditional multi-channel airborne radar systems,SAR received stationary clutter and moving targets for different receivers,hence the difference of moving targets in different receivers can help to detect moving targets.However,SAR ground moving target indication(GMTI)system suffers from the channel mismatch problem,which results in big clutter difference between channels and may severely affects the GMTI results.Moreover,high-resolution wide-swath(HRWS)SAR system becomes a popular trend recently,but it is restricted by the range and Doppler aliasing problem.Many researchers used multiple-input multiple-output(MIMO)SAR system to solve this problem.However,MIMO SAR systems suffer from the orthogonality of transmitting signals and matched filter at the receiver.The dismatched signal will interfere with the matched signal and pollute the final GMTI results.Besides,SAR systems can generate super-high-resolution images,such as 0.1m*0.1m images,in which the shape features of moving targets can be clearly obtained and used to recognize the target classes.Nevertheless,the speckle noise,a special noise that is produced by the SAR system,affects the local features of moving targets a lot and results in a bad recognition accuracy.Therefore,in this complex electromagnetism environment,based on the channel mismatch problem,MIMO SAR waveform design and matched filter problem,speckle noise problem,and RFI problem,the main work of this paper is summarized as:1.For the SAR RFI suppression problem,we design a joint sparse and low-rank model for the narrowband RFI signal.Based on this model and with the help of auxilary variables,we propose a joint sparse and low-rank(JSLR)optimization method to suppress the RFI.However,the JSLR method has high computational complexity with slow convergence.By detailed observation,the RFIs perform like some sinusoids with stable frequencies,therefore,we can restrain the row sparsity of RFI instead of joint sparsity and low-rank property.It avoids the singular value decomposition(SVD)to update the low-rank term and also does not introduce any auxilary variables.Hence,the row sparse(RS)method is a fast method,which considers both the sparsity and low-rank property.Furthermore,since the sampling frequency is commonly larger than the main bandwidth,we can down-sampling the signal to decrease its scale and reduce the computational complexity.Moreover,the truncated signal can remove the out-of-mainband RFI as well,which may improve the final performance.At the end,the real data are provided to demonstrate the effectiveness of the proposed method.2.For the same SAR RFI suppression problem,the nuclear norm minimization problem,which is a traditional low-rank recovery problem,over-penalizes the large singular values of RFI and results in a severe pollution to the useful signal.Besides,the solution of nuclear norm minimization problem requires huge computational complexity.Hence we propose two algorithm,the reweighted matrix factorization(RMF)and the matrix factorization decomposition(MFD)algorithms,to efficiently suppress the RFI based on two different approximations of the rank function.Benefit from the matrix factorization technique,the proposed methods can speed up the convergence without degrading the recovery performance.We strictly derive the closed-form solutions of the proposed methods and employ the real data to demonstrate their effectiveness.3.For the same SAR RFI suppression problem,the semi-parameter methods usually need to finely tune the hyperparameters in the optimization problems.Hence,we propose a “lowrank + sparse” decomposition model to transfer the hyperparameter into the rank and cardinality.We replace the singular value decomposition(SVD)by the bilateral random projection(BRP)to update the RFI and employ the soft-thresholding method instead of the hard-thresholding method to recover the complex radar signal.This method is termed revised traditional decompsition(RTD)-BRP method.Besides,we employ CLEAN&BIC method to estimate the rank of RFI and sparsity estimation method to estimate the sparsity of strong scatterers in each snapshot,and this method is termed parameter-free decomposition(PFD)method.According to the estimated hyperparameters,the proposed PFD method can recover the useful signal precisely.Moreover,single-snapshot estimation will generate better performance because the RFI and the useful signal are changing with different snapshots.Finally,the real data is provided to demonstrate the effectiveness of the proposed methods under different input SINRs.4.For the channel mismatch problem,we propose a time-Doppler chirp varying(TDCV)filter to generate a new image based on the single channel SAR system,and then we compare the new image with the traditional range-Doppler(RD)image.The clutter in the scene center keeps motionless,while the moving target has range offset due to the radial velocity.Through this range offset,we can estimate the radial velocity of moving target and increase the detectable ability of slow-moving target.The numerical simulations and examples demonstrate the estimation accuracy of motion parameter and increase the detectable ability of ground moving targets.5.For the MIMO SAR system,we employ the orthogonal frequency division multiplexing(OFDM)chirp signals as the transmitting signal.Based on this waveform,we prove that the robust principal component analysis(RPCA)can help to directly extract the moving target in the matched signal without the interference of the unmatched signal and the other components in the matched signal.Besides,in order to release the computational complexity,we fully employ the characteristic of the SAR GMTI system and propose a fast interferometry RPCA(FI-RPCA)algorithm for GMTI.The proposed method extremely decreases the computational complexity without introducing any large-scale operation and avoids iterations.The performance of trational RPCA method is restricted by the noise level and hyperparameter.Hence in order to improve the GMTI performance,the proposed method employs a novel two-step magnitude and phase(M&P)along-track interferometry(ATI)method.As a result,we have more robust performance and extend the value range of the hyperparameter,which is very useful for the practical applications.Moreover,the proposed method can also estimate the motion parameter and relocate moving targets to the correct place.6.For the speckle noise in the SAR automatic target recognition(ATR)system,we propose a joint low-rank and sparse multi-view denoising(JLSMD)method by using the RPCA method to extract de-speckle-noising target images from multiple similar views of training data.With the JLSMD dictionary or training data,the SVM and SRC methods are employed to recognize targets with higher recognition accuracies.Through the three-class and ten-class targets recognition tasks based on the MSTAR public dataset,the proposed JLSMD+SRCl1 method outperforms most of the state-of-the-art deep learning methods.It not only achieves very high recognition accruacy but also avoids long training time and high-quality GPU requirement,which is an advantage for practical applications.