Abstract
Marine mammal calls play a vital role in navigation, localization, and communication. Effectively classifying these calls is essential for ecological monitoring, species conservation, and military biomimetic applications. However, traditional machine learning methods struggle to capture complex acoustic patterns, while most existing deep learning approaches rely solely on frequency-domain features and require large datasets, which limits their performance on small-scale marine mammal datasets. To address these challenges, we propose a hybrid architecture combining a time-attention Long Short-Term Memory (LSTM) network and a multi-scale dilated causal convolutional network. The model comprises three modules: (1) a frequency-domain feature extraction module employing dilated causal convolutions at multiple scales to capture multi-resolution spectral information from Mel spectrograms; (2) a time-domain feature extraction module that inputs Mel-frequency cepstral coefficients (MFCCs) into an LSTM enhanced with a time-attention mechanism to highlight key temporal features; and (3) a classification module leveraging transfer learning, where a pre-trained neural network is fine-tuned on real marine mammal call data to improve performance. Extensive experiments were conducted on vocalizations from four marine mammal species. Our proposed method outperformed existing baseline models across four evaluation metrics: accuracy, precision, recall, and F1 score, with improvements of 3%, 7%, 2%, and 4%, respectively. The results confirm the effectiveness of combining frequency- and time-domain features along with attention mechanisms and transfer learning. This hybrid approach enhances the accuracy and robustness of marine mammal call classification, especially under limited data conditions.