Facilitation of bird call classification using self-organising maps—A pilot study on bird calls in New Zealand

Abstract

Automatic detection and classification of bird calls from audio recordings attract interests of various users from amateur bird watching hobbyists to professional biologists and park rangers studying biodiversity and conservation. Previous studies utilize machine learning techniques many of which require supervised learning; therefore, they cannot detect bird calls that are not included in the supervisory data. Their performance also severely degrade when a segment of a bird call is overlapped by other birds’ calls or contaminated with various ambient noise. This study proposes an alternative approach for bird call classification, which is facilitated by an algorithm based on the self-organising maps (SOM). Since the SOM is an unsupervised learning algorithm, which classifies similar input features into neighboring clusters, it is expected that segments of different bird species will be classified into clusters located apart while segments with multiple bird calls and/or noise will lie down on the clusters in the middle. The classified segments are then labelled by measuring the distance between the feature vectors in each cluster and that of known bird call segments given a priori. The performance of the proposed method is evaluated by applying the algorithm to data recorded in a national park in New Zealand.