Underwater Acoustic Signal Classification Based on Sparse Time–Frequency Representation and Deep Learning

For classification of underwater acoustic signals, we propose a novel sparse anisotropic chirplet transform (ACT) to reveal fine time-frequency structures. The signal features in the form of a time-frequency map are fed into a deep convolutional neural network, referred to as a time-frequency feature network (TFFNet), which brings flexibility to signal classification. The TFFNet is based on a novel efficient feature pyramid enhancing feature (EFP) maps by aggregating the context information at different scales. To remove the gridding artifacts on enhanced feature maps, a form of aggregating transformation, a forward feature fusion, is utilized to merge the forward feature maps. Main contributions of this work are a novel sparse ACT, a TFFNet classifier, and an EFP with forward feature fusion. Experimental results demonstrate that the sparse ACT provides a high-resolution time-frequency representation of underwater signals and the TFFNet improves the classification performance compared to known networks and two machine learning methods (random forest and support vector machine with radial basis function kernel) on two real data sets, an underwater acoustic communication signal data set and whale sounds data set.