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

• DOI: 10.1121/1.4919993
• Publication Date: 2015-05-01T12:07:43Z
• Authors: Nathan J. Reznicek; Carl R. Hart; D. Keith Wilson; Chris L. Pettit; Edward T. Nykaza
• Keywords: 0103 physical sciences; 01 natural sciences

Many outdoor sound propagation models exist, ranging from highly complex physics-based simulations to simplified engineering calculations. More recently, highly flexible statistical methods have become available, which offer the capability to predict outdoor sound propagation, given a training dataset. Among the variety of modeling approaches, the range of incorporated physics varies from none, as in the statistical learning methods, to comprehensive considerations for physical models. Engineering methods vary in the level of incorporated physics, oftentimes resorting to heuristic or approximate approaches. In order to compare the capability of engineering and statistical learning models, one particular physics-based model is used for outdoor sound propagation predictions, namely, a Crank-Nicholson parabolic equation (CNPE) model. Narrowband transmission loss values predicted with the CNPE, based upon a simulated dataset of meteorological, boundary, and source conditions, act as simulated observations. Among the engineering models used in the comparisons are the Harmonoise propagation model and the ISO 9613-2 method. Among the statistical learning methods used in the comparisons is a random forest regression model. Metrics such as the root-mean-square error and the skill score are computed for both the engineering models and statistical learning models.