Reducing Low-Rank Interference to Improve Environmental Estimation from Ambient Sound

Hypothesis testing presents an efficient means of partitioning the seabed into regions of similar geoacoustic properties. Ambient sound collected by an autonomous vehicle towing a vertical line array (VLA) can be used to characterize the seabed over large regions of interest. However, additional sound sources such as passing ships result in low-rank interference that can make accurate hypothesis testing impossible. We present a non-convex optimization procedure for removing low-rank interference from covariance matrices formed from a VLA. We describe an alternating optimization algorithm and an intelligent initialization procedure that allows us to efficiently recover the uncorrupted covariance matrix. On synthetic data, we show that our approach can allow for reliable hypothesis testing even in the presence of very strong interference. We also demonstrate that recordings from the New England Shelf Break Acoustics experiment contain low-rank interference that disrupts hypothesis testing, and that our method can adequately mitigate this interference, restoring the area under the receiver operating characteristic curve to 1.0.