Rapid assessment of vessel noise events and quiet periods in Glacier Bay National Park and Preserve using a convolutional neural net

A rapid assessment of vessel noise events and quiet periods in Glacier Bay National Park and Preserve (GBNP) reveals significant diurnal and seasonal patterns, with shorter noise-free intervals (NFIs) and higher hourly proportions of vessel noise during peak tourism months (June–August). Using a machine-learning model trained on 18 hours of labeled audio data, researchers analyzed nearly 9,000 hours of hydrophone recordings from 2017 to 2020. The convolutional neural net transfer-learning model (GBNP-CNN) identified vessel noise with an F1 score of 75%, though some confusion with harbor seal roars led to NFI lengths being minimum estimates. Despite this, the model effectively captured qualitative patterns, showing that in 2018, daytime NFIs during peak tourism were less than half the duration compared to May and September (1.3 min vs. 2.9 min). In 2020, reduced large-vessel tourism resulted in a decrease in vessel noise proportion during June–August. This study highlights the potential of ML models for monitoring soundscape metrics, providing critical data for managing acoustic habitats and conserving sensitive species in GBNP.

Haver, S. M., Gustafson, K. B., & Gabriele, C. M. (2023). Rapid assessment of vessel noise events and quiet periods in Glacier Bay National Park and Preserve using a convolutional neural net. Environmental Data Science, 2. https://doi.org/10.1017/eds.2022.36