Tue, Aug 16, 2022: 1:45 PM-2:00 PM
518A
Background/Question/MethodsAs the global climate is changing, so can the phenology of living species. Birds, for example can adjust the timing of their migration or breeding in response to warmer springs. Polar regions are particularly vulnerable to climatic perturbations, as temperatures rise faster there. Unfortunately, due to their remoteness, the cost and logistics involved, few studies describe how climatic variation impacts the phenology of arctic birds. As bird vocal activity is closely linked to their breeding stage, singing activity during the summer could be used as a proxy to their breeding phenology. And thanks to the recent advances in acoustic technology, we are now able to records sounds for months in harsh conditions and at multiple locations. This introduces a new challenge: analyzing the huge amount of data collected. Yet, with the recent progress made in the field of deep learning, it now becomes possible to train models that automatically detect bird songs in acoustic recordings with record accuracy. Using data collected from 24 acoustic recorders deployed in 8 arctic sites in 2018 and 2019, as well as publicly available bird song databases annotated with a high time resolution, we trained a lightweight bird song detector using a deep convolutional network.
Results/ConclusionsWe evaluated our model on a manually annotated subset representing a full summer of recordings on an arctic site. Our model achieves an AUC score of 89,4% on this target data. More importantly, we show that the curve of vocal activity during the summer provided by our model very closely follows the one obtained with our reference dataset. Besides, with AUC scores greater than 82% on two other reference datasets of European birds, our model offer a good generalization. And since it is lightweight, it can process 5 minutes of audio recording in less than a second on a middle range laptop. As deploying audio recorders becomes easier and cheaper, we believe that our model provides a fast, easy and reliable way to study the avian phenology responses to climate change at the community level on a global scale, even in the most remote areas.
Results/ConclusionsWe evaluated our model on a manually annotated subset representing a full summer of recordings on an arctic site. Our model achieves an AUC score of 89,4% on this target data. More importantly, we show that the curve of vocal activity during the summer provided by our model very closely follows the one obtained with our reference dataset. Besides, with AUC scores greater than 82% on two other reference datasets of European birds, our model offer a good generalization. And since it is lightweight, it can process 5 minutes of audio recording in less than a second on a middle range laptop. As deploying audio recorders becomes easier and cheaper, we believe that our model provides a fast, easy and reliable way to study the avian phenology responses to climate change at the community level on a global scale, even in the most remote areas.