Long-term deep learning-facilitated environmental acoustic monitoring in the Capital Region of New York State

Abstract

The effect of anthropogenic activity on animal communication is of increasing ecological concern. Passive acoustic recording offers a robust, minimally disruptive, long-term approach to monitoring species interactions, particularly because many indicator species of environmental health factors such as biodiversity, habitat quality, and pollution produce distinct vocalizations. Machine learning algorithms have been used in recent decades to automatically analyze the large quantities of audio data that result. In this study, a microphone array was used to collect continuous audio data at a site in the Capital Region of New York State for twelve months, resulting in over 8000 h of recordings. A 19-class database containing a variety of bio- and anthrophony was used to train a convolutional neural network in order to generate a reliable record of species-specific calling activity for the entire study period. These results were used to calculate an acoustics-based pseudo-species richness and abundance distribution. Additionally, heatmap plots were used to visualize (i) the time of day (x), sound category (y), and predicted number of sonic events for an average 30-day period and (ii) the day of the year (x), time of day (y), and predicted number of sonic events for each sound category. The correlations between these sonic events and various abiotic factors such as number of daylight hours, temperature, and weather activity were also examined.