Few-shot transfer learning enables robust acoustic monitoring of wildlife communities at the landscape scale

Pre-trained classifiers for bioacoustic species detection have emerged as accessible tools for conservation and automated biodiversity monitoring at scale. Despite their wide use, these models are trained primarily on high-quality, weakly labeled species recordings from disparate regions that can fail to account for variation in passive ambient soundscape data due to poor signal-to-noise ratio and the presence of novel sounds such as abiotic noise and concurrent vocalizations. This domain shift may limit a pre-trained model's reliability and preclude its use in applied acoustic monitoring. Although relatively underexplored in bioacoustic monitoring, transfer learning may help overcome the limitations of existing pre-trained classifiers by enhancing predictive performance while avoiding the time, resources, and expertise required to develop entirely new models. We demonstrate the use of few-shot transfer learning as a strategy to efficiently adapt a pre-trained BirdNET source model to a new target domain with minimal local training data. We present an open-source workflow with guidelines to evaluate performance with ecologically meaningful metrics and train a target model with improved or missing classes for biotic and abiotic signals of interest. This process is demonstrated with an applied case study of community level acoustic monitoring across a managed temperate rainforest landscape in western Washington, United States. The target model achieved a mean precision-recall AUC of 0.94 at the audio segment level and a mean accuracy of 92 % at the site use level across trained avian species classes, an increase of 0.15 and 13 % over the pre-trained source model, respectively. These improvements increased the probability of individual species detection and produced more accurate estimates of local species richness. Several species of high conservation and management priority that could not be reliably detected by the source model achieved high performance. Superior performance was achieved with an average of only 8 local training examples per species class. Training novel classes for other signals of interest (e.g., anurans, insects, aircraft, logging, rain) reduced confusion rates and mitigated the impact of the local environment on performance. Few-shot transfer learning leverages the learned knowledge of a pre-trained classifier to enable robust monitoring at the community level across the landscape and offers an approach that can be readily applied to other environments and taxa. • Transfer learning enables community monitoring by adapting pre-trained BirdNET model. • Improved probability of detection and species richness estimates across environments. • Superior performance achieved with average of only 8 local training examples per class. • Supports detection of diverse audible taxa and abiotic signals of interest. • Guidelines and open-source workflow provided for transfer learning practitioners.