A research team from the Marine Research Institute (Inmar) of the University of Cádiz has developed an artificial intelligence (AI) system to detect cetacean whistles in the Strait of Gibraltar, “one of the noisiest and most complex marine environments in the world”. According to the researchers, the tool drastically reduces manual review time, facilitating the monitoring of populations, the identification of periods of greater activity, and providing objective information for maritime traffic management in sensitive areas. The main innovation of the study lies in applying an iterative process; that is, a progressive training where the model analyzed the recordings and pointed out possible whistles. Using transfer learning techniques, they adapted AI models originally designed to recognize bird songs to the marine environment. The researchers installed passive acoustic monitoring systems near the island of Tarifa and collected over 1,300 hours of audio in four surveys conducted over a month and a half in different seasons. This technique uses hydrophones to continuously record underwater sounds without interfering with animal behavior, allowing for monitoring at night, in poor visibility, during bad weather, or at great depths. In parallel, they developed a system to automate the process and intelligently select fragments with the highest probability of containing vocalizations. “With the model, it takes one day to process 500 hours,” explained Neus Pérez, a researcher at the UCA and co-author of the study. The study also highlights the importance of the marine soundscape, composed of biophony (sounds of living beings), geophony (such as waves and currents), and anthrophony (human-origin sounds). “Cetaceans are known for their communicative ability, but many marine organisms also generate sound,” Pérez pointed out, adding that “analyzing this set allows us to assess a marine area and water quality, especially in areas like the Strait, where the human component is dominant.” This validation was re-incorporated so that the model could learn the acoustic particularities of the area. Therefore, the result is a tool capable of analyzing thousands of hours of underwater recordings with a reliability close to 88%. Beyond the Strait, the methodology is transferable to acoustic monitoring programs in other regions, not only in difficult marine environments, as explained in the article ‘Iterative deep learning for cetacean whistle detection in the Strait of Gibraltar’, published in the journal Engineering Applications of Artificial Intelligence.
