Few-shot transfer learning for endemic and data-scarce species: A scalable framework for bioacoustic analysis and monitoring

Acoustic communication pervades the animal kingdom, spanning taxa from insects to whales, with many species employing structured sequences of acoustic elements for interaction. Bioacoustics, the study of sound production, transmission, and reception in animals, plays a critical role in understanding wildlife behavior, communication, and ecological dynamics. Many endemic species are elusive, nocturnal, or inhabit remote regions, and some are acoustically active only during limited temporal windows. These factors make large-scale labeled data collection challenging. Furthermore, many deep learning studies exclude species with fewer than 50 annotated recordings during dataset curation, thereby limiting representation of rare and low-resource taxa. This highlights the need for data-efficient learning strategies capable of operating under severe data scarcity. In this study, we investigate episodic meta-learning using PANN-derived embeddings and Prototypical Networks for Few-shot species classification under low-resource conditions. Using the BirdCLEF 2024 dataset as a benchmark, we evaluate the proposed framework across varying N-way and K-shot settings to assess generalization to Novel species. Experimental results demonstrate that metric-based meta-learning enables robust adaptation to unseen and underrepresented species with minimal supervision. Under the Base versus Novel generalization evaluation, the proposed network achieves Novel species classification accuracies of 61.1 ± 2.13%, 47.7 ± 1.89%, and 38.0 ± 1.93% under 3-way, 5-way, and 8-way evaluation settings respectively with only five support samples per species, all substantially exceeding their corresponding random chance baselines of 33.3%, 20%, and 12.5%. Compared to a transfer learning baseline that utilizes the same PANN embedding backbone but relies on episode-level fine-tuning rather than episodic meta-learning, the proposed FSL model consistently outperforms across all settings, achieving improvements of +6.8%, +9.3%, and +10.3% at 3-way, 5-way, and 8-way respectively, with all differences statistically significant at p<0.001 and Cohen’s d increasing from 0.42 to 1.08 across settings. The model additionally demonstrates substantial computational efficiency, completing inference over 500 episodes approximately 6×, 5×, and 4× faster than the transfer learning baseline at 3-way, 5-way, and 8-way settings respectively, while achieving superior classification accuracy across all evaluation settings. These results collectively demonstrate that the proposed framework generalizes effectively to previously unseen species with minimal supervision, offering a taxonomically and computationally scalable solution for long-tail biodiversity monitoring. By emphasizing performance in data-constrained scenarios, this work contributes to the development of efficient, ecologically meaningful automated acoustic monitoring systems.