Deep learning-based automated detection reveals temporal patterns of calling in tarsiers in Sulawesi, Indonesia

Dena J. Clink , Johny S. Tasirin

Tarsiers are small, haplorrhine primates that occur in Southeast Asia. Tarsiers on the island of Sulawesi range from Vulnerable to Critically Endangered, and many are data deficient, which means there is a great need for improved monitoring approaches. Sulawesi tarsiers are pair- living, territorial, and engage in duets within human hearing range, which makes them ideal candidates for passive acoustic monitoring (PAM), an approach that relies on autonomous acoustic recording units. Here, we provide a proof-of-concept workflow that combines PAM and automated detection to monitor Gursky's spectral tarsier (Tarsius spectrumgurskyae) in Tangkoko National Park, North Sulawesi, Indonesia. We used a custom trained BirdNET model deployed over ~520 hours of PAM data and manually verified all detections. Similar to previous work, we found that the majority of tarsier duet vocalizations occurred around sunrise, with a few vocalizations emitted in a non-duet context other times during the night. To our knowledge, this is one of the first applications of deep learning-based automated detection for the study of tarsiers. We advocate for future PAM studies across land use gradients, different forest types, and under different management regimes, to improve tarsier conservation efforts across Sulawesi. We provide the labeled training data and the trained model to facilitate future work.