Towards Domain Shift in Location-Mismatch Scenarios for Bird Activity Detection

Bioacoustic monitoring serves as a valuable tool for gaining insights into the well-being of wildlife. Sensor locations with diverse acoustic conditions pose a major challenge for deep learning-based audio classification systems. In this paper, we study unsupervised domain adaptation techniques for the task of bird activity detection in short audio segments using two bird recognition datasets with recordings from diverse locations. Furthermore, we explore various distance and divergence metrics to quantify the domain shift as a proxy to predict the expected drop in classification accuracy at different recording locations. Our results confirm the superior performance of the instancewise feature projection-based domain adaptation (IFPDA) technique across multiple audio domains and demonstrate that useful domain shift metrics can be derived from the energy distribution across frequency bands.