Características espectrais dinâmicas extraídas de sinais acústicos submarinos para a classificação de embarcações utilizando algoritmos de aprendizagem supervisionada

Abstract

The classification of vessels plays a fundamental role in monitoring maritime traf- fic and identifying sound sources in the underwater environment, whether they are of anthropogenic or biological origin. However, classifying a vessel based solely on its emitted sound is challenging due to the complexity of the ocean as an underwater waveguide. This work aims to develop a robust and effective feature extraction algo- rithm to capture dynamic aspects of underwater signals over time. Additionally, a classification system composed of different machine learning algorithms is evaluated. The proposed algorithm extracts dynamic spectral features (Spectral Dynamic Fea- tures - SDF), combining them with Mel Frequency Cepstral Coefficients (MFCCs) to form a combined feature vector. The performance of the proposed method is evaluated using two datasets containing vessel sounds, the first collected off the northwest coast of Spain and the second collected near the municipality of Arraial do Cabo – RJ. Furthermore, to broaden the analysis of the proposed method, the two datasets contain different vessels and environmental conditions. The results demonstrate that the proposed method achieves a classification accuracy superior to traditional methods based on static features.