Domain Adaptation of Few-Shot Bioacoustic Event Detection in Different Environments

Few-shot Bioacoustic Event Detection (FS-BAED) aims to monitor and analyse events with few labelled samples. Across different environments, the feature distribution of background events may diverge, which is considered a domain shift problem. This paper identifies two distinct forms of bias induced by such shifts: the familiar sampling bias of few-shot learning, and a under-recognised semantic feature bias, whereby background events in novel environments are systematically misclassified as target events. To address both, this work introduces a unified FS-BAED framework. First, a contrastive learning backbone augmented with a multi-class classification strategy mitigates sampling bias by enforcing instance-level feature separability. Second, an environment-shift detector flags and filters out background samples whose acoustic features diverge significantly from known distributions, thereby suppressing false positives caused by semantic feature bias. The proposed framework is evaluated on the DCASE 2022 Task 5 dataset under two settings: (i) a fixed environment benchmark common to existing works, and (ii) an altered environment scenario in which foreground events from DCASE dataset are tested with simulated backgrounds of ESC-50 dataset. Compared to prior methods, our framework yields a 1.47% F-measure gain in the fixed setting and over 21% in altered environments, validating the importance of explicitly modeling semantic feature bias.