Intelligent identification of dolphin whistle in acoustic signals via deep learning.

Deep learning has been increasingly employed for identifying dolphin vocalizations. Specifically, convolutional neural networks (CNNs) proved effective in detecting dolphin whistles. The aim of the study is to introduce a novel method based on CNNs and Sobel filter to identify dolphin whistles from spectrograms extracted from underwater audio recordings. Moreover, sensitivity analysis was performed to investigate the impact of Sobel filter compared with other filtering approaches (CLAHE and Laplacian), spectrogram size, and CNN kernel shape on model performance. Three independent datasets of bottlenose dolphin vocalizations were used, recorded (1) in a dolphin pool of Oltremare marine park; (2) in the Adriatic Sea during interactions between dolphins and fishing activities; and (3) in open-sea conditions in the Hawaiian Exclusive Economic Zone (DCLDE 2022). Results showed that, in the best-case scenario, model accuracy and F1-score were 90.0% and the vertical Sobel filter significantly improved CNN performance compared with unfiltered spectrograms and with CLAHE and Laplacian filtering. Rectangular spectrograms also enhanced performance over square spectrograms. Instead, variation in CNN kernel shape had minimal impact. Model effectiveness was preserved even under challenging conditions of depredation in open sea. Moreover, computation time is compatible with real-time applications. Model suitability in different environments and the low computational time make this approach appropriate for monitoring underwater environments.