Automated sea vehicle classification system based on neural network

Automated marine objects identification and classification based on their underwater acoustic sounds is an important task for water areas monitoring and ocean exploration. The key steps in object recognition are the features extracted from the marine object underwater sounds and classifier selection. Paper describes acoustic features analysis, which contains a comparison between the use of one feature and the use of various features sets. Classification results of classical machine learning methods and artificial neural network are given. Research was performed on a large volume of real acoustic records of sea vessels underwater sounds, recorded in the waters of Peter the Great Bay. Database contains about 20 GB of “.wav” files with different sea vessels classes. Developed automated classification system based on pre-trained neural network and proposed acoustic features set allows to achieve 80% prediction accuracy (F score). Features set consist of GFCCs, chroma features, CQT, spectral bandwidth and spectral centroid. The system successfully predicts the ship class even in the presence of noises.