Comparisons between physics-based, engineering, and statistical learning models for outdoor sound propagation.

Area of Science:

• Acoustics
• Environmental Science
• Computational Modeling

Background:

• Outdoor sound propagation modeling is crucial for noise assessment and environmental planning.
• Existing models range from complex physics-based simulations to simplified engineering methods and statistical learning approaches.
• A need exists to evaluate and compare the accuracy of these diverse modeling techniques.

Purpose of the Study:

• To benchmark and compare the predictive accuracy of various outdoor sound propagation models.
• To evaluate engineering models (ISO 9613-2, Harmonoise, Nord2000) against a physics-based Crank-Nicholson parabolic equation (CNPE) model.
• To assess the performance of statistical learning models (bagged decision tree, random forest, boosting, artificial neural network) using CNPE as a reference.

Main Methods:

• A physics-based Crank-Nicholson parabolic equation (CNPE) model was used as a benchmark to generate simulated sound propagation data.
• Simulated data encompassed diverse conditions: downward/upward refraction, hard/soft boundaries, and low frequencies.
• Engineering models (ISO 9613-2, Harmonoise, Nord2000) and statistical learning models were compared against CNPE predictions using skill scores.

Main Results:

• Statistical learning models demonstrated superior performance, with skill scores consistently above 99.5% (bagged decision tree, random forest, boosting, artificial neural network).
• Engineering models showed varied performance: Nord2000 achieved 83.8%, while ISO 9613-2 (0.6%) and Harmonoise (-7.1%) had significantly lower skill scores.
• The study highlights the high accuracy of statistical learning methods in simulating complex sound propagation scenarios.

Conclusions:

• Statistical learning models offer a highly accurate and flexible approach for outdoor sound propagation prediction.
• These advanced models significantly surpass traditional engineering methods in predictive performance across various acoustic conditions.
• The findings support the adoption of machine learning techniques for more reliable environmental noise modeling and assessment.