Avian bioacoustics research was greatly assisted by the introduction of autonomous recording units, which not only allow remote monitoring but also make large-scale studies possible. However, manual inspection of acoustic recordings becomes more challenging with increasingly larger datasets. In this study, we developed a logistic model to predict the probability of bird presence in audio recordings using sound frequency percentiles. The acoustic recordings covered bird songs and calls in a wide range of environments (e.g. grassland, forest, urban areas) along with the presence of noise due to weather, traffic, insects, and human speech. Based on leave-one-out cross-validation, our final logistic model resulted in a 75% overall accuracy and a 16% false negative rate using the optimal cut-off of 0.35 (i.e. probability ≥ 0.35 indicates the presence of birds). Compared with a convolutional neural network model using the same dataset, the logistic model was about seven times faster in terms of the processing time, but achieved slightly lower overall accuracy. This bird sound detection model using sound frequency percentiles in a logistic model opens up promising approaches to aid in automatic, accurate, and efficient analyses of large audio datasets for monitoring wildlife communities.
Autonomous recording units (ARUs), bird song recognition, biodiversity monitoring, convolutional neural networks, IEEE bird sound detection challenge