Detection and classification of marine mammal sounds using AlexNet with transfer learning

In this study, AlexNet with transfer learning was employed to automatically detect and classify the sounds of killer whales, long-finned pilot whales, and harp seals with widely overlapping living areas. Transfer learning was used to overcome the overfitting problem of deep network as the training samples was insufficient. A challenging dataset containing both target (the three marine mammal sounds) and non-target (ambient noise including ship noise, pulse interference, and man-made sounds, etc) sounds collected from different recording times, locations and devices was used to examine the performance of the proposed method, and the sounds used in the test dataset were completely independent of the training dataset. The overall accuracy of the trained detection and classification models reached 99.96% and 97.42% respectively. Importantly, each trained model took only 1.3 ms to detect or classify a single image. Furthermore, feature visualizations and strongest activations demonstrated that the proposed method learns the true differences between different marine mammal sounds rather than differences between different recording environments and devices. Therefore, all results show that the proposed method has excellent performance and great potential for practical application.

Automated summary

In this study, AlexNet with transfer learning was used to automatically detect and classify killer whale, long-finned pilot whale, and harp seal sounds with high accuracy. The challenging dataset evaluation confirmed the excellent performance of the proposed method in real-world scenarios.