Investigation of Unsupervised Models for Biodiversity Assessment

Significant animal species loss has been observed in recent decades due to habitat destruction, which puts at risk environmental integrity and biodiversity. Traditional ways of assessing biodiversity are limited in terms of both time and space, and have high cost. Since the presence of animals can be indicated by sound, recently acoustic recordings have been used to estimate species richness. Bioacoustic sounds are typically recorded in habitats for several weeks, so contain a large collection of different sounds. Birds are of particular interest due to their distinctive calls and because they are useful ecological indicators. To assess biodiversity, the task of manually determining how many different types of birds are present in such a lengthy audio is really cumbersome. Towards providing an automated support to this issue, in this paper we investigate and propose a clustering based approach to assist in automated assessment of biodiversity. Our approach first estimates the number of different species and their volumes which are used for deriving a biodiversity index. Experimental results with real data indicates that our proposed approach estimates the biodiversity index value close to the ground truth.