ECE PhD Proposal Review: Hamed Mohebbi Kalkhoran

PhD Proposal Review: Machine Learning Approaches for Classification of Myriad Underwater Acoustic Events Over Continental-shelf Scale Regions with Passive Ocean Acoustic Waveguide Remote Sensing

Hamed Mohebbi Kalkhoran

Location: Zoom Link

Abstract: Underwater acoustic data contain a myriad of sound sources that include bioacoustics related to marine life such as marine mammals and fishes; man-made such as ships, sonar, and airguns; as well as natural geophysical processes such as earthquake, hurricane, and volcanic eruption. Among underwater acoustic events, marine mammal vocalization classification is one of the most challenging problems due to their transient broadband calls, high variation in the calls of a specie (intra-class variation), and high similarity between the calls of some species. In this thesis, we investigate machine learning approaches for classifying marine mammal vocalizations for real-time applications. We utilize acoustic data from a 160-element coherent hydrophone array and employ the passive ocean acoustic waveguide remote sensing technique to enable sensing and detections over instantaneous wide areas more than 100 km in diameter from the array. A variety of computational accelerating approaches, combining hardware and software, that make the methods desirable for real-time applications are also developed.
Humpback whale behavior, population distribution and structure can be inferred from long term underwater passive acoustic monitoring of their vocalizations. Here we employ machine learning approaches to classify humpback whale vocalizations into song and non-song calls. We use wavelet signal denoising and coherent array processing to enhance the signal-to-noise ratio. To build features vector for every time sequence of the beamformed signals, we employ Bag of Words approach to time-frequency features. Finally, we apply Support Vector Machine (SVM), Neural Networks, and Naive Bayes to classify the acoustic data and compare their performances. Best results are obtained using Mel Frequency Cepstrum Coefficient (MFCC) features and SVM which leads to 94% accuracy and 72.73% F1-score for humpback whale song versus non-song vocalization classification.
To classify a large variety of whale species by their calls, we extracted time-frequency features from Power Spectrogram Density (PSD) of the beamformed signals. Then we used these features to train three classifiers, which are SVM, Neural Networks, and Random forest to classify six whale species: Fin, Sei, Blue, Minke, Humpback, and general Odontocetes. Best results were obtained with Random forest classifier, which achieved 95% accuracy, and 85% F1 score. To detect transient sound sources, first we applied Per-Channel Energy Normalization (PCEN) on the PSD of the beamformed signals. We applied thresholding on the PCEN data followed by morphological image opening to find potential sound sources and reduce noisy detections. Then we applied connected component analysis to obtain the final detected sounds for each bearing. To estimate the Direction of Arrival (DoA) of detected sounds, we applied non-maximum suppression (NMS), which is widely used in object detection applications in computer vision, on the detected sounds. We used mean power of each detected sound as the scores for NMS. To speed up the data processing, we investigated a variety of accelerating approaches, such as analyzing the effect of floating point precision, applying parallel processing, and implementing fast algorithms to run on GPU.