Processing acoustic features of animal vocalizations using classification procedures. Free

Remote methods for classifying age, sex, group membership, or individual identification of animals that live in visually obscured environments are extremely valuable tools for field biologists, but reliable identification of individual callers still presents important challenges. Acoustic features of animal vocalizations can be processed to extract caller identification using a variety of sophisticated classification techniques, but the exact classification process can be difficult to justify rigorously and challenging to repeat on novel data. A feature extraction and classification process that is clear, simple, and repeatable would be a major benefit to wildlife studies. Classification and regression trees (CART) generate intuitive and clear processes for handling multidimensional acoustic information. Examples of CART applied to Mexican spotted owl (Strix occidentalus lucida) and humpback whale (Megaptera novaeangliae) vocalizations will be provided. These CART results will be compared to other classification techniques, particularly neural networks. CART performance was comparable, but had the advantage that it yielded explicit classifiers used to categorize vocalizations, making it easy to integrate into acoustic surveying systems. This promises to be a valuable tool for conservation and management of these and other endangered species.