This technological tool drastically reduces the time required for manual review of acoustic recordings. According to Neus Pérez, a UCA researcher and co-author of the study, the system can process up to 500 hours of underwater recordings in a single day, with a reliability close to 88%. This information was disseminated via a press release by the Fundación Descubre.
The methodology employed is adaptable to acoustic monitoring programs in other regions, even in challenging marine environments. The details of this research have been published in the article 'Iterative deep learning for cetacean whistle detection in the Strait of Gibraltar' in the journal Engineering Applications of Artificial Intelligence.
The Strait of Gibraltar, due to its geographical location, is a crucial habitat for numerous marine mammal species, including large cetaceans such as the sperm whale (Physeter macrocephalus) and the fin whale (Balaenoptera physalus), in addition to being a vital migratory route for dolphins and orcas. However, studying the signals emitted by these animals is hampered by environmental noise, predominantly anthropogenic in origin, which complicates species identification in recordings.
The study and classification of the signals emitted by cetaceans in this context is subject to various difficulties, such as environmental noise, dominated by anthropogenic sounds, and the consequent difficulty of identifying species in the recordings.
Given that the Strait of Gibraltar is a particularly noisy environment due to intense maritime traffic and high species density, the findings of this study and the applied methodology—which uses two capture systems to differentiate between whistles and noise—have great potential for use in other high-acoustic-density areas worldwide. It is estimated that between 50% and 70% of ambient noise levels in oceanic waters exceed 90 decibels, especially in areas of high human impact, potentially reaching over 120 dB in naval corridors or industrial zones.
The application of these detection methods is considered "especially promising" in high-acoustic-impact zones that are also ecologically significant for cetaceans.
The researchers installed passive acoustic monitoring systems near the island of Tarifa, collecting over 1,300 hours of audio. These recordings were made over a month and a half at different times of the year, using hydrophones or underwater microphones to continuously record marine sounds without disturbing animal behavior. This allows for nocturnal monitoring, in poor visibility conditions, during storms, or at great depths.
In parallel, a system was developed to automate the process and intelligently select fragments most likely to contain cetacean vocalizations. To train this system, existing internet audio of cetacean sounds was used, and artificial intelligence models, originally designed to recognize bird songs, were adapted to the marine environment through transfer learning techniques.
