Automatic Anuran Species Recognition via Vocalization
MetadataShow full item record
Species recognition of Anuran vocalization is an important problem in ecology, which is costly and time consuming. Recently, biologists have turned to machine learning techniques to automate the process. An automated process can be made even more helpful if available to use on a smart phone, which can be used in the field by scientists and citizen scientists alike. The problem of automatic frog call recognition is challenging because features need to be chosen that can properly separate different species’ very similar sounding vocalizations. This thesis offers the following contributions: 1. Three novel feature extraction methods, which do not require segmentation or background noise reduction; 2. Survey of seven classifiers and combinations of these; 3. Automatic classification on large database of 736 calls from 45 species which is a larger database of individual anurans than any previous classification work, and the largest database of calls ever complied. Methods were testing using leave one out and 10 fold cross validation. Results from the model that classified 43 species shows that 15 species were identified with 100% accuracy. This is the highest number of accuracy for the most species in published literature.