ReconUNet: Learning to Reconstruct a Clean Covariance Matrix Under Low SNR, Multipath, and Array Imperfections

Direction-of-arrival (DoA) estimation degrades under low signal-to-noise ratio, coherent multipath, and array imperfections. Coherent replicas are especially damaging because they collapse the rank of the signal covariance. We propose ReconUNet, a U-Net that maps a stack of multi-lag autocorrelations to a covariance matrix. It is trained with a single reconstruction loss toward the closed-form clean covariance on the nominal manifold, and its output feeds unmodified classical estimators. In simulations with an eight-element array, we compare it to six classical estimators, spatial smoothing, and three learned frontends trained on the same corpus. With a Root-MUSIC backend, ReconUNet attains a pooled RMSE of $4.23$${}^{\circ }$ on the held-out split, against $7.91$${}^{\circ }$ for the best learned baseline and $12.43$${}^{\circ }$ for Root-MUSIC on the sample covariance. It has the lowest error for every source count, and one trained model serves MUSIC, Root-MUSIC, ESPRIT, and Unitary-ESPRIT. The network acts as a learned regularizer of the covariance, not as an array calibrator. Its multi-lag input depends on the temporal correlation of the sources. When trained at one bandwidth, it fails for wideband sources; when trained over a range of bandwidths, it is the most accurate method tested across that range. Code and weights are released with the paper.