Di Nardo, Francesco (Università Politecnica delle Marche, Ancona), De Marco, Rocco (CNR), Li Veli, Daniel (CNR-IRBIM), Screpanti, Laura (Università Politecnica delle Marche), Castagna, Benedetta (Università Politecnica delle Marche), Novelli, Giovanni (Università Politecnica delle Marche), Lucchetti, Alessandro (CNR-IRBIM), Scaradozzi, David (Università Politecnica delle Marche)

High-Accuracy Detection of Bottlenose Dolphin Whistle Using AI

Scheduled for presentation during the Regular Session "Intelligent systems" (WeBD), Wednesday, June 11, 2025, 15:20−15:40, Room C

33rd Mediterranean Conference on Control and Automation, June 10-13, 2025, Tangier, Morocco

This information is tentative and subject to change. Compiled on May 9, 2025

Abstract

The persistent interaction between dolphins and commercial fishing operations has led to ecological and socio-economic challenges, primarily through bycatch and depredation. Traditional mitigation strategies have shown limited success, needing innovative solutions. Intelligent robotic systems capable of identifying and consequently responding to dolphin vocalizations seem to be a promising approach to mitigate dolphin interactions with fishing operations. The core of this intelligent system should be an advanced algorithm or an artificial intelligence architecture capable of identifying dolphin vocalizations and distinguishing them from other underwater sounds. Thus, this study proposes a novel approach to detect dolphin whistles using a convolutional neural network (CNN) paired with advanced spectrogram processing techniques. The method utilizes audio recordings of common bottlenose dolphins (Tursiops truncatus) from Oltremare marine park in Italy. Whistle detection was enhanced by applying edge-detection filters to spectrograms, which highlights characteristic of dolphin whistles while filtering out noise. The processed spectrograms served as inputs to a CNN with a three-layer architecture optimized for binary classification of dolphin whistles. The model achieved very promising results, with accuracy, precision, recall, and F1-scores around 99% across a 10-fold cross-validation. The findings demonstrate the method robustness, offering potential applications in conservation efforts and real-time monitoring. Future research will focus on adapting the approach to field conditions where real-time processing and non-ideal whistle recording pose additional challenges.