Incidence of data augmentation in machine learning using broadband spectrograms

One of the main problems in ocean acoustics is the lack of training data that assist networks to generalize better understand undersea environments. In deep learning, data augmentation is commonly used to add variability to the data by applying small transformations to the samples with the purpose of improving the performance and generalization capabilities of the networks. Here, this definition is extended to broadband spectrograms from merchant ships collected during the Seabed Characterization Experiment in 2017. In this study, residual neural networks were used as the machine learning algorithm for performing source localization and environment classification. Additionally, nine different data augmentation techniques were considered for studying their performance predictions. Each of the transformations was applied to the data set during the training stage, and the results were compared to establish the effects in the performance for both regression and classification tasks. The metrics used were the root mean squared error for regression and accuracy in the case of seabed classification. The results show a favorable potential of residual-based deep learning models to differentiate the seabed types and the source position based on these types of data, in addition to the improvements in performance after applying complex transformations during the training. [Work supported by the Office of Naval Research, Contract No. N00014-19-C-2001.]