Validating two geospatial models of continental-scale environmental sound levels.

Area of Science:

• Environmental Science
• Acoustics
• Machine Learning

Background:

• Accurate modeling of outdoor environmental sound levels is crucial but challenging.
• Machine learning models offer a potential solution for predicting soundscapes at continental scales.

Purpose of the Study:

• To validate two continental-scale machine learning models for predicting outdoor sound levels.
• To assess the performance of models using geospatial data and A-weighted L50 as a metric.

Main Methods:

• Utilized geospatial layers as input features for two distinct machine learning models.
• Validated model predictions against A-weighted L50 (sound level exceeded 50% of the time) during summer daytime.
• Compared model performance, identifying areas with significant prediction errors.

Main Results:

• Observed validation errors exceeding 20 dBA in certain areas.
• Attributed large errors to insufficient acoustic training data and spatial dissimilarity between training and validation sites.
• Identified the need for models to extrapolate beyond their training data.

Conclusions:

• Current machine learning models for environmental sound levels require improvement.
• Optimal data collection strategies and robust uncertainty quantification are essential for future model development.
• Further research is needed to enhance the accuracy and reliability of large-scale acoustic models.