• AUV
• detection and classification
• marine mammals
• marine robotics

This thesis, carried out in collaboration with ATLAS Elektronik, proposes an autonomous system for the automatic detection and classification of marine mammals. Passive acoustic monitoring is currently one of the main methods used to study cetaceans, allowing vocalizations to be detected and classified without interfering with the animals' natural behavior.
The thesis presents two different phases of work: at first, data on selected marine mammal species were collected and studied. The acoustic signals were then preprocessed and converted into log-Mel spectrograms to enhance their spectral characteristics. In the second part, two distinct neural networks were developed for the detection and classification of marine mammals. The two networks differ in the number of mammals considered and in their architecture: the first has a cascade architecture, while the second consists of a single network.
The two solutions are based on real datasets and enriched using data augmentation techniques, showing good generalization capabilities despite the limited availability of data.
Validation of the system was carried out in Germany at the Arberger Kanal in Bremen. During the experiment, acoustic signals were replayed in a controlled environment and data were collected to test the effectiveness of the neural first in realistic scenarios. The results highlighted the potential for future applications in noninvasive monitoring of marine mammals.