A Deep Learning Model for Marine Mammal Call Classification

The classification of marine mammal calls is of enormous significance for the protection of marine mammals. Acoustic methods are the most effective tools for studying and monitoring marine mammals, as many species emit loud and distinctive sounds that are more reliable than visual cues. To address these requirements, this article proposes a multispecies marine mammal call classification method. First, this article develops a basic model framework for call classification. Using convolutional neural network bank (CBank) to effectively capture the characteristics of call audio on different time scales, highway networks make deep CBank network training more efficient, while bidirectional recurrent neural network (BiGRU) can consider both past and future information through bidirectionality, which improves the model's ability to perceive the dynamic changes of time series. Second, this article employs quantum computing and a chaotic algorithm to introduce an improved particle swarm optimization (IPSO) algorithm for optimizing the parameters of CBank-highway networks-BiGRU (CHBG). Finally, this article proposes an expedite attention (EA). EA reduces the amount of calculation and memory usage by block processing and sparse matrix operations, which accelerates the speed of dealing with long sequence features. We used the IPSO-CHBG-EA method to classify ten types of marine mammal calls and 1 type of marine noise. The experimental results showed that the accuracy was 9.03% higher than the parallelized artificial neural networks, the chimpanzee optimization algorithm's control parameters, and the empirical mode decomposition models, and the training time was reduced by 20%–30%.