Deep-Learning-Based Source Localization Using a Local 3-D Acoustic Intensity Field in a Sound-Speed Mismatched Deep-Ocean Environment

This article proposes a deep-learning (DL)-based method for accurate acoustic source localization in the deep ocean, which leverages a local 3-D acoustic intensity field (AIF) with the dimensions of 1.26 km
$\times 1.26$
km
$\times 504$
m to estimate the 3-D position of a source within a range of 20 km. The local AIF can be constructed from measurements from an underwater acoustic sensor network (UASN) using techniques, such as optimal interpolation. To achieve accurate 3-D localization, a 3-D residual convolutional neural network (ResCNN) is designed for processing the AIF. Moreover, the method incorporates a 3-D water column sound-speed field (SSF) that covers the entire sound propagation path to enhance localization accuracy. To incorporate the environmental information for source localization, two approaches are introduced. One is to implicitly utilize the environmental information through network training based on training data generated in various environments. The second approach explicitly includes the 3-D water column SSF to the network input. Since the environmental mismatching problem is inevitable, we discuss the impact of deviations between the adopted SSF and the real SSF on source localization accuracy. The training, validation, and test datasets are generated by the BELLHOP3D acoustic simulation model. Once the proposed DL-based localization networks are fully trained, their performance outperforms that of a conventional matched-field processing (MFP) localization algorithm for the case where the mismatching factor of SSF obtained through common ocean model estimation ranges from 18% to 29%.