The Evolutionary Structure of Acoustic Learnability: A Deep Learning Approach to Neotropical Birdsong
Passive Acoustic Monitoring offered a scalable solution for biodiversity assessment in the Neotropics, although classifying hundreds of sympatric species in complex soundscapes remained challenging. We developed a deep learning framework for large-scale avian classification by training convolutional neural networks on recordings from 667 Neotropical bird species across northern South America. Efficient-NetV2L performed best, achieving 94.48% accuracy, 94.30% macro F1-score, and 0.998 macro ROC-AUC. After training, we conducted a post hoc analysis of species-level F1-scores using phylogenetically informed models (PGLS and PGLMM) to test morphological, ecological, and geographic predictors under phylogenetic control. Broad traits explained little interspecific variation: the best-supported PGLS accounted for approximately 2.2% of the variance, whereas the null model received the strongest support among PGLMM candidates. Geographic range size showed the most consistent negative association with F1-score, while morphological and ecological predictors added little explanatory power. Performance nevertheless showed weak but significant phylogenetic signal, and frequent confusions tended to involve more closely related species. Grad-CAM and Monte Carlo Dropout provided complementary descriptions of saliency and predictive uncertainty. Overall, deep learning performed effectively for regional biodiversity monitoring and provided a comparative framework for assessing how much of the variation in acoustic classification performance could be explained by broad biological predictors.