Automatic standardized processing and identification of tropical bat calls using deep learning approaches

Consistent and comparable metrics to automatically monitor biodiversity across the landscape remain a gold-standard for biodiversity research, yet such approaches have frequently been limited to a very small selection of species for which visual approaches (e.g., camera traps) make continuous monitoring possible. Acoustic-based methods have been widely applied in the monitoring of bats and some other taxa across extended spatial scales, but are have yet to be applied to diverse tropical communities.

In this study, we developed a software program “Waveman” and prepared a reference library using over 880 audio-files from 36 Asian bat species. The software incorporated a novel network “BatNet” and a re-checking strategy (ReChk) to maximize accuracy. In Waveman, BatNet outperforms three other published networks: CNNFULL, VggNet and ResNet_v2, with over 90% overall accuracy and 0.94 AUC on the ROC plot. The classification accuracy rates for all 36 species are at least 86% when analysed in combination. Moreover, our library preparation and ReChk greatly improved the sensitivity and reduced the false positive rate, when tested with 15 species for which more detailed and situationally diverse records were available. Finally, BatNet was successfully used to identify Hipposideros larvatus and Rhinolophus siamensis from three different environments. We hope this pipeline is useful tool to process bioacoustic data accurately, effectively and automatically, therefore allowing for greater standardization and comparability for researchers to understand bat activities across space and time and therefore provide a consistent tool for monitoring biodiversity for management and conservation.