Mapping ambient sound levels using physics-informed machine learning

Abstract

Mapping the spatial and temporal distribution of ambient sound levels is critical for understanding the impacts of natural sounds and noise pollution on humans and the environment. Previously, ambient sound levels have been predicted using either machine learning or physics-based modeling. Machine learning models have been trained on acoustical measurements at geospatially diverse locations to predict ambient sound levels across the world based on geospatial features. However, machine learning requires a large number of acoustical measurements to predict ambient sound levels at high spatial and temporal resolution. Physics-based models have been applied to predict transportation noise at high spatial and temporal resolution on regional scales, but these predictions do not include other anthropogenic, biological, or geophysical sound sources. In this work, physics-based predictions of transportation noise are combined with machine learning models to predict ambient sound levels at high spatial and temporal resolution across the conterminous United States. The physics-based predictions of transportation noise are incorporated into the machine learning models as a geospatial feature. The result is a physics-informed machine learning model that predicts ambient sound levels at high spatial and temporal resolution across the United States. [Work funded by an Army SBIR]