Integrating modeled environmental variability into neural network training for underwater source localization

Supervised machine learning (ML) is a powerful tool that has been applied to many fields of underwater acoustics, including acoustic inversion. ML algorithms depend on the existence of extensive labeled datasets, which are difficult to obtain for the task of underwater source localization. A feed-forward neural network (FNN) trained on imbalanced or biased data may end up suffering from a problem analogous to model mismatch in matched field processing (MFP), that is, producing incorrect results due to a difference between the environment sampled by the training data and the actual environment. To overcome this issue, physical and numerical propagation models can act as data augmentation tools to compensate for the lack of comprehensive acoustic data. This paper examines how modeled data can be effectively used for training FNNs. Mismatch tests compare the output from a FNN and MFP and show that the network becomes more robust to various kinds of mismatches when trained on diverse environments. A systematic analysis of how the training dataset's variability impacts a FNN's localization performance on experimental data is carried out. Results show that networks trained with synthetic data achieve better and more robust performance than regular MFP when environment variability is taken into account.