Decoding the soundscape using deep learning models: a system to support long-term acoustic monitoring of wildlife occupancy with citizen science input

Bioacoustic surveys have gained popularity in recent years, largely due to their versatility in rapid biodiversity assessments and long-term monitoring. Improved detection probabilities in acoustic surveys provide value in documenting occupancy of ecologically-cryptic, rare, and secretive species. Today the need for such monitoring is paramount, given rapid, global-scale decline in biodiversity. To address long-term bioacoustic monitoring in a wetland landscape, we present an experimental system in which deep-learning models (Tidzam) are applied to multiple 24 x 7 microphone capture, performing continuous, real-time species-level call identification and subsequent analyses. Tidzam is an open source, open data project that employs an iterative learning approach to segment and identify the biophony from geophony and anthrophony. Compared to earlier tools that use clustering technology to aid the analysis of recorded soundscapes, Tidzam detects, identifies and geo-localizes acoustic events at the species level at the time of capture, thus providing the temporal (daily and seasonal) variation of wildlife occupancy and activity within the landscape. With better acoustic detection and temporal range, Tidzam data can be used to estimate species occupancy and create spatially-explicit density maps. Such automated systems are critical for species identification, as the voluminous data generated exceeds human capacity to annotate and evaluate. Tidzam’s annotation platform provides a crowdsourcing mechanism inviting novice and expert observers to refine the iterative process. This platform can be extended in multiple ways for scientific inquiry, education and public engagement. Using the notation platform, experts can debate the source of an acoustic event. A Tidzam tutorial allows students of all ages to learn to identify species-specific calls. Tidzam drives a novel sound listening system that extends auditive perception of the landscape; a time machine feature invites comparisons of the current soundscape with that of earlier specified date/time/location allowing listeners to explore the spatiotemporal variability of wildlife dynamics.