Using Machine Learning to Measure Biodiversity from Sound Recordings
Biodiversity is an incredibly challenging metric to measure. This project aims to classify a soundscape and use that knowledge to help classify 500,000 minutes of sound data to understand broad, landscape scale patterns of biodiversity, human impact through noise pollution, and areas of quiet. All of these are indicators of ecosystem and community quality - essential measures for conservation, monitoring, and land management decision making. The main classification categories are Anthrophony (e.g., cars, airplanes, human voices), Biophony (e.g., birds, insects, amphibians), Geophony (e.g., wind, rain, running water), and Other. The main tools used to accomplish this task are mel spectrograms (e.g., visual representation of sound), convolutional neural networks (CNNs), transfer learning, ensemble learning, support vector machines (SVMs), and uniform manifold approximation and projection (UMAP). With these techniques we are able, to get braod category accuracies of 87%, and with confidence thresholding, we get accuracies of broad classification of 96%, and subcategory classification accuracies of 86%, 89%, and 100% for each subcategory classifier.