Open-set classification approaches to automatic bird song identification: towards non-invasive wildlife monitoring in Brazilian fauna

Abstract

Bird song identification has mainly been approached as a closed-set classification problem; that is, all samples are known to be from one of the classes known by the classifier. However, wildlife monitoring using bird songs is closer to an open-set classification setting, as the classifier is required to predict if a sample comes from an unknown origin, like an environmental sound or an unrelated animal. Furthermore, current approaches to bird song classification assume that the model can access the whole dataset and build optimal projections. This is not a realistic scenario in Brazil as the country has thousands of species, and it is unfeasible to build a dataset containing a representative diversity of samples of all of them. This work analyzes algorithms that can be used for the open-set classification of bird songs. The analyzed algorithms can fit models using data from one or from only a few species. The investigation revealed many current technical difficulties and highlighted several opportunities for future work in this field.