Recognition and Classification Techniques of Marine Mammal Calls Based on LSTM and Expanded Causal Convolution

Wanlu Cheng^1,2Hao Chen³Jiaming Jiang^1,2Shuang Li^1,2Jingjing Wang^1,2*Yanping Zhou^1*

• ¹Qingdao University of Science and Technology, Qingdao, China
• ²Shandong Key Laboratory of Deep Sea Equipment Intelligent Networking, Qingdao, China
• ³Technische Universität Ilmenau, Ilmenau, Thuringia, Germany

The calls of marine mammals are crucial tools for navigation and localization between individuals.Effectively classifying these calls holds significant importance for ecological monitoring, species conservation, and covert underwater acoustic communication in military biomimetics. However, the classification method based on traditional machine learning makes it difficult to deal with complex patterns, which affects the performance and classification effect of the model. Existing deep learning methods typically use a single frequency-domain feature extraction approach, which cannot fully capture important time-domain features, leading to incomplete extracted features. Additionally, these methods require large amounts of data to train the model, resulting in insufficient classification performance when the training data is limited for marine mammal calls. To enhance the accuracy and robustness of recognizing and classifying marine mammal calls, we propose a method based on a time-attention Long Short-Term Memory(LSTM) and a multi-scale dilated causal convolutional network, which consists of three main components. The first component is the frequency-domain feature extraction module, which uses a multiscale dilated causal convolutional network to extract features from the Mel spectrogram of the audio. This module connects dilated causal convolutional networks at three different scales, with each scale incorporating a distinct dilation factor. Second, is the time-domain feature extraction module, which inputs the Mel-frequency cepstral coefficients(MFCC) features of the audio data into an LSTM network for extracting time-domain features, thereby supplementing the temporal information in the audio. A time-attention mechanism is then introduced to enhance the focus on important time-domain features. Thirdly, we adopt transfer learning, where a pre-trained neural network is further trained on real marine mammal call data to replace the convolutional network used for classification. Extensive experiments on four species of marine mammals demonstrate that our method achieves the best performance across four metrics: accuracy, precision, recall, 1 Wanlu Cheng et al.