Underwater Acoustic Monitoring Using MFCC for Fuzzy C-Means Clustering, Naive-Bayes and Hidden Markov Model-Based Classifiers

摘要

The whale sounds help researchers in population assessments and to follow the migratory path of whales. Acoustics is the best way to study and observe cetaceans since it is automatic and non-invasive. A technique capable of differentiating between whale songs, other marine sounds and man-made sounds would be very useful for the scientific community. This paper presents a system that can classify signals collected from humpback whales vocalization data through acoustic monitoring from Puerto Rico coastal marine habitats. The fuzzy c-means algorithm is used to cluster, in an unsupervised manner, sounds from the same marine source. The Naive-Bayes classifier and a Hidden Markov Model are used to classify marine sounds such as whale songs, man-made sounds and natural background noise from hydrophone recordings under water. The MFCC (Mel-Frequency Cepstral Coefficients) approach is used to extract the most important characteristic of the signals and a vector quantization and clustering approach is used to obtain the states for training the HMM. Satisfactory results are obtained through HMM classifier which outperforms the Naive-Bayes classifier for whale songs and marine sounds.