Abstract
Animals worldwide are facing ecological pressures from global climate change and increasing anthropogenic activities. To transition to a renewable energy future, extensive offshore wind development is planned globally. In the North Atlantic, future development sites overlap with the migratory range of critically endangered North Atlantic right whales (NARW) and will lead to increased risk of ship strikes, pile driving impacts, and other population risks. New methods to accurately detect cetaceans and provide real-time feedback for mitigation will be increasingly important to enact sustainable management actions to facilitate the recovery of the NARW. Recent developments in acoustic event detection made possible by deep learning have shown significantly improved detection performance across many different taxa, but such models tend to be too computationally expensive to run on existing wildlife monitoring platforms. Here, we use model compression techniques combined with an autonomous acoustic recording platform integrating an ESP32 microcontroller to bring real-time detection with deep learning to the edge. We test if edge-based inference using a compressed network running on a microprocessor entails significant performance loss and find that this loss is negligible. We leverage large, open-source datasets of noise from the NOAA SanctSound project for generating semi-synthetic training datasets that encourage model generalization to novel noise conditions. Our compressed model achieves improved performance across all tested recording sites in the Western North Atlantic Ocean, demonstrating that deep learning powered wildlife monitoring solutions can provide reliable real-time data for mitigation of human impacts and help ensure a sustainable green energy transition.