Sparsity-based space-time adaptive processing using convolutional neural network

In this paper, a deep learning framework for space-time adaptive processing is developed. Firstly, a set of clutter covariance matrixes (CCMs) are modeled based on the prior parameters of radar and navigation system with respect to all possible levels of non-ideal factors, and the columns of each CCM is formulated as undersampled noisy linear measurements of the sparse coefficients corresponding to angle-Doppler spectrum. Then the original spectrum coefficients, obtained by least-square estimation from the modeled CCMs and known space-time steering dictionary, are used as input to train the convolutional neural network (CNN). Meanwhile, the corresponding labels can be obtained by the exact spectrum of modeled CCMs via minimum variance distortionless response algorithm. Once trained, the CNN can be used to predict angle-Doppler spectrum coefficients that corresponds to a new measurement vector in near real time. Simulations results have demonstrated the superiority of the proposed method in both clutter suppression performance and computation efficiency.