Abstract
In marine ecology research, it is crucial to accurately identify the marine mammal species active in the target area during the current season, which helps researchers understand the behavioral patterns of different species and their ecological environment. However, the difficulty and high cost of collecting marine mammal calls, coupled with limited publicly available datasets, result in insufficient data for support, making it difficult to obtain accurate and reliable identification results. To address this problem, we propose MarGEN, a deep learning-based augmentation method for marine mammal call signal data. This method processed the call data into Mel spectrograms, then designed a self-attention conditional generative adversarial network to generate new samples of Mel spectrograms that were highly similar to the real data, and finally reconstructed them into call signals using WaveGlow. The classification experiments on the calls of four Marine mammals show that MarGEN significantly enriches the diversity and volume of the data, increasing the classification accuracy of the model by an average of 4.7%. The method proposed in this paper greatly promotes marine ecological protection and sustainable development, while effectively advancing research progress in bionic covert underwater acoustic communication technology.