Machine Learning Reveals Microdialect Variations in Salish Sea Melospiza melodia

Understanding microdialects, or distinct variations in songbird (oscine) vocalizations, is critical to identifying populations that may act as environmental indicators for anthropogenic impact. Although vocal variation in avian species has been studied for decades, microdialect classification can be improved using the emerging fields of passive acoustic monitoring (PAM) and machine learning. This study aims to quantify the ability for neural networks to detect fine-scale geographical variations among contiguous populations through predicting an individual’s location. A lightweight ~0.92M parameter Convolutional Neural Network (CNN) was designed to identify the presence of microdialects in five populations of Salish Sea Melospiza melodia (song sparrow) from the Puget Sound and British Columbia using citizen science recordings (364 hr.) and date metadata. This was achieved with two complementary models: one to predict geographical coordinates through regression and another to classify recordings into discrete regions. The regression model achieved an average error of 35.16 km, a 60.95% improvement over the baseline average obtained by chance. The classification model performed with an accuracy of 63.02%, an increase of 43.02 percentage points over the baseline accuracy. These results demonstrate the capability of deep learning models to detect subtle variations among contiguous populations, allowing feasible implementation in terrestrial autonomous recording units (ARUs) with limited computational resources and variable field conditions. This suggests that deep learning can enable real-time, efficient monitoring of oscine populations in ecological studies.