Two-Dimensional DOA Estimation Based on Dual-Branch CNN

Abstract

Direction-of-arrival (DOA) estimation is a core research topic in array signal processing, and two-dimensional (2-D) DOA estimation can jointly acquire the azimuth and elevation of incoming signals, which bears great practical value. Traditional subspace and sparse reconstruction algorithms are plagued by heavy computation and deteriorated accuracy under imperfect array manifolds, low signal-to-noise ratios (SNRs) and insufficient snapshots. To enhance estimation robustness and inference speed simultaneously, this paper presents a dual-branch convolutional neural network (CNN) for 2-D DOA estimation based on uniform rectangular arrays. The network takes the sample covariance matrix of array received data as input. A shared feature encoder with residual blocks and channel-attention modules extracts common spatial features, followed by two prediction heads with independent parameters for elevation and azimuth estimation. Because each branch has a 61-dimensional output while two sources may be simultaneously present, the angle estimation is formulated as multi-label classification using sigmoid outputs and weighted binary cross-entropy. Simulations covering diverse SNRs, snapshot counts, angular intervals and off-grid cases verify that the proposed network obtains smaller root mean square errors than methods with multiple signal classification (MUSIC), estimation of signal parameters via rotational invariance techniques (ESPRIT) and ordinary CNN methods, with millisecond-level inference latency. This framework offers an efficient, high-precision real-time 2-D DOA estimation scheme for complicated electromagnetic scenes.