Detecting bird sounds in a complex acoustic environment and application to bioacoustic monitoring

Abstract

Trends in bird population sizes are an important indicator in nature conservation but measuring such sizes is a very difficult, labour intensive process. Enormous progress in audio signal processing and pattern recognition in recent years makes it possible to incorporate automated methods into the detection of bird vocalisations. These methods can be employed to support the census of population sizes. We report about a study testing the feasibility of bird monitoring supported by automatic bird song detection. In particular, we describe novel algorithms for the detection of the vocalisations of two endangered bird species and show how these can be used in automatic habitat mapping. These methods are based on detecting temporal patterns in a given frequency band typical for the species. Special effort is put into the suppression of the noise present in real-world audio scenes. Our results show that even in real-world recording conditions high recognition rates with a tolerable rate of false positive detections are possible.

Introduction

In order to evaluate the impact of human activities on populations of wild animals and to decide on the most effective actions for nature conservation, we need fundamental information on the extent of changes in the living environment. Birds are a good indicator for changes in biodiversity because they are distributed over a wide range of landscapes, are easy to detect in comparison to other animal groups and we have a good knowledge on the biology of most of the species. It is a fortunate fact that, at least in most European countries, we have huge associations of skilled and experienced birdwatchers, who willingly give their knowledge to non-profit service and support monitoring programs. Due to the activity of birdwatchers, data regarding trends in population sizes for certain European bird species has been recorded and made available since 1980 (Gregory et al., 2005). Different standardised methods for bird census have been developed (Bibby et al., 1992). Most of them are based on the mapping of singing males, assuming that the number of territorial males is equal to the number of breeding pairs. The most widely used method for estimating the number of breeding birds is based on point counts where all individuals heard and seen from stationary places are estimated (Klvaňová and Voříšek, 2007).

Complementing traditional approaches, vocalisations of birds serving for territory maintenance and mate attraction can be used for an automated acoustic monitoring of bird populations (Brandes, 2008, Frommolt et al., 2008). The main advantage of such an automated bioacoustic approach, as compared to previous methods, lies in the long-term recording in the absence of an observer. It allows to estimate bird numbers in ecologically sensitive areas (like nature reserves) or in areas that are difficult to access (for example large reed habitats). Even nocturnal birds and birds with low vocal activity could thus be effectively counted. In addition to the need of an applicable autonomous recording device, the greatest challenge is the development of appropriate pattern recognition algorithms giving reliable results even in complex acoustic environments. In order to apply acoustic methods for the monitoring of bird species, we have to solve two problems. We need pattern recognition algorithms for the automatic detection and identification of bird species and we need appropriate techniques for the estimation of the number of individuals.

Acoustic pattern recognition algorithms have been successfully applied to the study of nocturnal migrations of birds (Evan and Mellinger, 1999, Farnsworth et al., 2004, Farnsworth, 2005, Farnsworth and Russel, 2007, Hüppop et al., 2006, Mills, 2000, Hill and Hüppop, 2008, Schrama et al., 2008, Marcarini et al., 2008). However, these works mostly deal with relatively simple signals without background noise. The number of detected calls was used as a criterion to assess the occurrence of migrating birds. For noisy environments, hidden Markov models were successfully applied to frequency modulated sounds (Brandes, 2008, Trifa et al., 2008).

In this work we report about an approach using pattern recognition techniques for continuous bird monitoring. Our techniques are based on event detection and repetition rate estimation of bird song elements. Additionally, the algorithms use noise estimation from frequency bands known not to contain the bird’s vocalisations and apply them to an effective noise reduction. They can be used in parallel with traditional monitoring methodology to yield methods with improved speed and reproducibility in those cases where reliable detectors for bird vocalisations are available. In contrast to other applications our algorithms have been designed specifically for certain endagered target species. This approach was chosen in order to improve recognition rates even under poor acoustic conditions and it leads to a larger range covered by a single sensor unit in comparison to human observers and algorithms requiring high signal-to-noise ratios. Moreover, we apply this methodology to a quantitative survey of a real bird population yielding automatic map generation of breeding territories for one of the considered bird species.

In Section 2 we summarise related work in the field of pattern recognition for animal sounds. Approaches for the detection of the vocalisations of two endangered bird species are developed in Section 3. These algorithms have been evaluated in a study described in Section 4, which also gives evaluation results. These results are sufficient for automatic map generation of breeding territories for one of the bird species investigated. Results of a corresponding study are presented in Section 5. Finally, conclusions are given in Section 6.