Deep Learning-Based Covariance Reconstruction for STAP in Subarray-Sampled Airborne FDA Radar Systems

Abstract

Airborne radar clutter exhibits pronounced spatiotemporal coupling, which makes effective suppression challenging in both the spatial and temporal domains. Space–time adaptive processing (STAP) has proven to be an effective solution, with its performance critically dependent on accurate covariance matrix estimation. However, practical constraints related to platform size and hardware resources prevent full-array reception in large-scale multichannel radar systems, such as frequency diverse array (FDA) radars. To address this challenge, this article investigates a deep neural network (DNN)–based clutter suppression approach for airborne FDA radars operating under subarray sampling. First, simulations are used to generate paired datasets consisting of subarray-sampled covariance matrices and their corresponding full-array theoretical covariance matrices. Then, two DNN architectures—a multilayer perceptron and a dilated convolutional neural network—are developed. Using subarray covariance matrices as inputs and full-array covariance matrices as labels, the proposed networks are trained to reconstruct the complete covariance matrix from severely undersampled data. The reconstructed covariance matrices are subsequently applied to STAP for clutter suppression. Simulation results demonstrate that both DNN models can accurately recover the full-array covariance matrix from subarray measurements, enabling effective clutter suppression in airborne FDA radar systems.