Underwater Sounds Classification for Effective Smart Marine Monitoring Systems

Underwater sound classification is commonly referred to the process of identifying and categorizing different types of sounds in underwater environments, such as those produced by marine animals, human activities, or natural phenomena. Deep learning models have revealed to be able to powerfully classify underwater sounds when conditions are particularly favorable. However, a practical monitoring system must contend with environmental noise and the need to classify diverse sounds that can vary significantly, even within the same class. In this work, we conduct focused experiments to evaluate the efficacy of modern, lightweight deep learning models with an eye toward designing intelligent buoys capable of monitoring the sea for signs of malicious activities or the presence of marine animals. We assemble a large collection of audio samples by combining publicly available datasets to create a diverse set of audio sources that robustly tests the classification performance of different models. The experimental results show that an optimized deep learning model can achieve promising binary classification accuracy. As expected, accuracy decreases in the more challenging multi-class classification scenario but remains above 60%, demonstrating the model’s potential for real-world applications.