A two-stage approach to automatically detect and classify woodpecker (Fam. Picidae) sounds.

Abstract

Inventorying and monitoring which bird species inhabit a specific area give rich and reliable information regarding its conservation status and other meaningful biological parameters. Typically, this surveying process is carried out manually by ornithologists and birdwatchers who spend long periods of time in the areas of interest trying to identify which species occur. Such methodology is based on the experts’ own knowledge, experience, visualization and hearing skills, which results in an expensive, subjective and error prone process. The purpose of this paper is to present a computing friendly system able to automatically detect and classify woodpecker acoustic signals from a real-world environment. More specifically, the roposed architecture features a two-stage Learning Classifier System that uses (1) Mel Frequency Cepstral Coefficients and Zero Crossing Rate to detect bird sounds over environmental noise, and (2) Linear Predictive Cepstral Coefficients, Perceptual Linear Predictive Coefficients and Mel Frequency Cepstral Coefficients to identify the bird species and sound type (i.e., vocal sounds such as advertising calls, excitement calls, call notes and drumming events) associated to that bird sound. Conducted experiments over a data set of the known woodpeckers species belonging to the Picidae family that live in the Iberian peninsula have resulted in an overall accuracy of 94,02%, which endorses the feasibility of this proposal and encourage practitioners to work toward this direction.