Adaptive statistical learning models for long-range sound propagation

Abstract

Uncertainties in source characteristics, meteorological conditions, and topographic variations present formidable challenges for accurately predicting long-range outdoor sound propagation. Numerical propagation models inherently assume perfect knowledge of these uncertain variables and are fixed in a modeling sense. In contrast, statistical learning models can incorporate new observations to update the underlying prediction model. Past work has shown that statistical learning models trained on synthetic data for predicting long-range sound propagation have, at best, an overall root-mean-square error (RMSE) of about 5 dB. This limit appears to be imposed by modeled atmospheric turbulence. It is hypothesized that this prediction limit may be lowered as observational data are incorporated into trained statistical learning models. Furthermore, data are assimilated by a Kalman filtering process for the purpose of updating knowledge of the atmospheric and source characteristics. Within the prediction phase thre...