DOA Estimation for Low-Angle Targets Using Time–Frequency Deep Learning

Abstract

Low-angle direction-of-arrival (DOA) estimation has become increasingly important in radar systems due to unmanned aerial vehicles (UAVs) operating at low altitudes. However, low-angle DOA estimation remains challenging because the direct and ground-reflected signals are highly correlated and angularly close. Existing methods often rely on simplified signal models or multiple-snapshot assumptions. Therefore, their robustness is limited under model mismatch, array imperfections, and single-snapshot moving-target scenarios. To address these challenges, this paper exploits frequency and temporal diversity to reduce the effect of multipath distortions. For this purpose, a radar operating simultaneously at multiple carrier frequencies is employed, and neighboring temporal snapshots of the moving target are used. Accordingly, neural-network-based estimators are developed, including a multilayer perceptron (MLP) that exploits frequency diversity, a bidirectional long short-term memory (BiLSTM) that exploits temporal diversity, and an MLP-BiLSTM fusion network that jointly exploits time and frequency diversity for DOA estimation. To train these networks, a model-selection strategy is proposed in which limited real measured data are used to evaluate networks trained with different candidate signal models. The best model is then used to generate synthetic training data. In addition, a constrained maximum-likelihood (ML) estimator is developed using low-angle prior information and extended to multi-frequency and consecutive-snapshot cases. The corresponding conditional Cramér–Rao bound (CRB) is also derived. Simulations and real-data results show that the proposed neural-network-based estimators achieve robust and accurate DOA estimation with low computational complexity in practical low-angle multipath scenarios.