Robust DOA estimation via covariance calibration and multi-loss complex-valued CNN

Conventional DOA estimation methods perform poorly under low SNR and limited snapshots. To address this problem, we propose a robust method based on an improved deep complex-valued convolutional neural network in this paper. The proposed method integrates a covariance calibration layer for input denoising, a multi-scale attention module to aggregate spatial features, and a multi-loss supervision strategy that jointly optimizes spectrum fitting, covariance recovery and angle error. The experiments show that this proposed method outperforms traditional algorithms such as MUSIC and ESPRIT, as well as existing deep models under low SNR and few snapshots conditions, achieving the lowest RMSE across tests. The ablation study confirms the effectiveness of the calibration module and the multi-loss strategy, providing an effective high-precision DOA solution for noisy environments, which has practical application value and is applicable to radar and communication fields.