Data-driven source localization using shipping sources of opportunity

Underwater source localization is often achieved with a purely model-based approach such as matched-field processing, which relies on simulated replica-field. However, such approaches only yield reasonable predictions if the complex and dynamic ocean environment is sufficiently known – often a daunting task. Alternatively, it has been suggested that channel impulse responses (CIRs) estimated from measurements of sources of opportunity (such as commercial shipping vessels) can feed a data-driven approach to source localization that forgoes the need for precise model-parameters [Durofchalk et al., JASA 146(4), 2691–2691 (2019)]. In this presentation, vertical line array (VLA) data from the SBCEx16 experiment conducted in the vicinity of shipping lanes in the Santa Barbara channel (580 m depth, downward refracting profile) are first used to a construct a library of estimate CIRs between selected locations along shipping tracks and VLA receivers using ray-based blind deconvolution (RBD) [Byun et al., JASA 141(2), 797–807 (2017)]. Subsequently, this library of data derived CIRs is used to train a machine learning algorithm to localize other surface sources. The average localization error and computational efficiency obtained with this machine learning approach is compared to that of traditional matched field processing techniques.