PT - JOURNAL ARTICLE AU - Oisin Mac Aodha AU - Rory Gibb AU - Kate E. Barlow AU - Ella Browning AU - Michael Firman AU - Robin Freeman AU - Briana Harder AU - Libby Kinsey AU - Gary R. Mead AU - Stuart E. Newson AU - Ivan Pandourski AU - Stuart Parsons AU - Jon Russ AU - Abigel Szodoray-Paradi AU - Farkas Szodoray-Paradi AU - Elena Tilova AU - Mark Girolami AU - Gabriel Brostow AU - Kate E. Jones TI - Bat Detective - Deep Learning Tools for Bat Acoustic Signal Detection AID - 10.1101/156869 DP - 2017 Jan 01 TA - bioRxiv PG - 156869 4099 - http://biorxiv.org/content/early/2017/06/29/156869.short 4100 - http://biorxiv.org/content/early/2017/06/29/156869.full AB - Passive acoustic sensing has emerged as a powerful tool for quantifying anthropogenic impacts on biodiversity, especially for echolocating bat species. To better assess bat population trends there is a critical need for accurate, reliable, and open source tools that allow the detection and classification of bat calls in large collections of audio recordings. The majority of existing tools are commercial or have focused on the species classification task, neglecting the important problem of first localizing echolocation calls in audio which is particularly problematic in noisy recordings.We developed a convolutional neural network (CNN) based open-source pipeline for detecting ultrasonic, full-spectrum, search-phase calls produced by echolocating bats (BatDetect). Our deep learning algorithms (CNN FULL and CNN FAST) were trained on full-spectrum ultrasonic audio collected along road-transects across Romania and Bulgaria by citizen scientists as part of the iBats programme and labelled by users of www.batdetective.org. We compared the performance of our system to other algorithms and commercial systems on expert verified test datasets recorded from different sensors and countries. As an example application, we ran our detection pipeline on iBats monitoring data collected over five years from Jersey (UK), and compared results to a widely-used commercial system.Here, we show that both CNNFULL and CNNFAST deep learning algorithms have a higher detection performance (average precision, and recall) of search-phase echolocation calls with our test sets, when compared to other existing algorithms and commercial systems tested. Precision scores for commercial systems were reasonably good across all test datasets (>0.7), but this was at the expense of recall rates. In particular, our deep learning approaches were better at detecting calls in road-transect data, which contained more noisy recordings. Our comparison of CNNFULL and CNNFAST algorithms was favourable, although CNNFAST had a slightly poorer performance, displaying a trade-off between speed and accuracy. Our example monitoring application demonstrated that our open-source, fully automatic, BatDetect CNNFAST pipeline does as well or better compared to a commercial system with manual verification previously used to analyse monitoring data.We show that it is possible to both accurately and automatically detect bat search-phase echolocation calls, particularly from noisy audio recordings. Our detection pipeline enables the automatic detection and monitoring of bat populations, and further facilitates their use as indicator species on a large scale, particularly when combined with automatic species identification. We release our system and datasets to encourage future progress and transparency.