Research On Brightness Temperature Reconsturcion Method And Target Detection Algorithm Of Synthetic Aperture Interferometric Radiometers Based On Machine Learning

Synthetic aperture interferometric radiometers (SAIRs) hold the same favorableconcealment and survival ability, and are applicable for most weather conditions andbattlefield environment in day and night. And high spatial resolution and instantaneousimaging without scanning is achieved by adopting the technique of aperture synthesis. Sothe researchers pay more and more attention to the novel detection method.The high-quality real-time brightness temperature reconstruction method and theeffective target detection algorithm are the major problem of the synthetic aperturemicrowave radiometer target detection. In this paper, based on machine learning theory, atfirst, the nature of the SAIR brightness temperature reconstruction is analyzed, then basedon a detailed analysis of the model selection of the SAIR inverse problem, the real-timehigh-quality SAIR brightness temperature image reconstruction methods are addressed; atlast, based on the statistical properties of the SAIR image, we propose the backgroundsuppression and target detection of SAIR target detection. The main contents include:The existing SAIR brightness temperature reconstruction methods are analyzed in theviewpoint of statistics. It is pointed out that the nature of the existing the SAIR inversionmethod is model selection problem. The difficulties of the optimal regularizationparameter determination and the actual risk minimization are also discussed.The existing SAIR brightness temperature reconstruction methods are analyzed in theBayesian framework. The potential statistical properties of the brightness temperaturedistribution and visibility samples for the classical SAIR brightness temperaturereconstruction methods are presented. Based on the detailed analysis of the statisticalmodel of the SAIR inverse problem, a real-time brightness temperature reconstructionmethod based on Bayesian linear regression is proposed. Using the Bayesian modelselection, the priori probability is introduced into the learning process of the SAIRbrightness temperature reconstruction. The Bayesian estimation model of the SAIRinverse problem is established. The methods for model paramter estimation arepresented. Compared with the existing SAIR brightness temperature reconstructionmethods, this method can real-time reconstruct the brightness temperature image without reducing the quality of image.The bound of the actual risk of SAIR brightness temperature reconstruction isanalyzed. It is pointed out that structural risk minimization is more suitable for SAIRinverse problem than empirical risk minimization. Based on the sparse regression modelestablishment to the SAIR inverse problem, two brightness temperature reconstructionmethods based on structural risk minimization, in the viewpoint of statistics and Bayesianframework, are proposed, respectively. Using the sensitive factor to control the structuralrisk minimization, the former is similar to the traditional regularization inversion method,and also need to artificially select the optimal model parameters; using the prioriprobability to achieve the structural risk minimization, the latter can automatic determinethe model parameters. Compared with the SAIR image inversion methods based onempirical risk minimization principle, these methods have a smaller expansionperformance of the noise and lower computational complexity.A mathematical model of the the SAIR target detection is established. Using thestatistical peculiarities of the SAIR brightness temperature image, the Gaussian mixturemodel and kernel regression method, a robust kernel regression algorithm for SAIR targetdetection is proposed. To suppress complex background and detect small target, gaussianmixture model is used to establish the optimal loss function of the SAIR backgroundestimation, and kernel regression is used to minimize the risk function in the feature spaceof the SAIR background estimation.In this paper, the SAIR brightness temperature reconstruction methods and targetdetection methods based on machine learning are undergone by a rigorous theoreticalanalysis, tested by simulation and experiment, and have a good practical prospect.’...