Deep learning for high‐resolution estimation of clutter angle‐Doppler spectrum in STAP

Abstract Space‐time adaptive processing (STAP) methods can provide good clutter suppression potential in airborne radar systems. However, the performance of these methods is limited by the training samples' support in practical applications. To address this issue, a deep learning framework for STAP is developed. First, the clutter space‐time data and their exact clutter covariance matrices (CCMs) are simultaneously modelled via simulation, in which various non‐ideal factors such as aircraft crabbing, array errors, and internal clutter motion with all possible levels in practice are all considered. Then, a multi‐layer two‐dimensional convolutional neural network (CNN) is developed. In this CNN, low‐resolution angle‐Doppler profiles estimated by a few training samples are used for the input and the high‐resolution counterpart obtained by the exact CCMs are utilized for the labels. Once trained, the CNN can be used to predict the high‐resolution angle‐Doppler profile using a few measured data in near real time. The high‐resolution clutter spectrum can be further calculated using the space‐time steering dictionary and the above obtained profile. Finally, the CCM of the measured data can be constructed and the space‐time weight vector can also be achieved. Compared with recently developed sparsity‐based STAP methods, the performance of the proposed method is better and the computational load of it is far fewer, and therefore more suitable for real‐world implementation. The simulation results have demonstrated the superiority of the proposed method in both clutter suppression performance and computation efficiency.