Marine Mammal Protection from Acoustic Interference
Developed a machine learning algorithm that achieved 97.83% accuracy in detecting and classifying marine mammals
The Problem
Human activities such as commercial shipping, fisheries, and seismic blasting produce significant underwater noise pollution. This noise pollution causes imbalances in marine ecosystems that distrupts marine mammals, posing a threat to the health and sustainability of the oceans. We introdced a solution that could be implemented in shipping channels to identify when marine mammals are present
My Contributions
- Discovery - organized the research phase, identifying suitable machine learning models and datasets for the project
- Product Definition - Contributed to the creation of product requirements, architecture documents, and wireframes in collaboration with team members
- Technical Scoping - Defined technical requirements, tested algorithms, and implemented the custom data pipeline and machine learning model.
Context
The project was undertaken as part of my Bachelor of Applied Science in Computer Engineering at Queen’s University. Our team aimed to address the environmental challenge of protecting marine mammals from harmful human-induced noises. We developed a sophisticated machine learning (ML) system capable of detecting, and classifying marine mammal species to mitigate the negative impacts of underwater noise pollution. We intended to use a multiple microphone array and indurtsy software for localization of the mammal as well though this fell outside the scope.
Opportunities
With the growing emphasis on environmental conservation and marine protection, there was a significant opportunity to create a technological solution that could aid in real-world conservation efforts. Our project aimed to develop an accurate and efficient system for identifying marine mammals based on acoustic data, thereby providing valuable tools for conservationists, regulators, and environmental researchers
Action
Conducted extensive research on machine learning models adaptable to acoustic classification challenges and scraped open-source datasets containing approximately 15,000 annotated data samples. Developed a custom data pipeline to preprocess and augment the data, converting it to visual formats (MFCCs), and trained a convolutional neural network (CNN) model to classify marine mammal species. The model achieved a classification accuracy of 97.83% on the test dataset, demonstrating its effectiveness in identifying marine mammal species from acoustic signals.