Marcos Augusto Burgos Saavedra, Samuel Wong, and Burak Ayva
Abstract – The current available approaches for predicting a range of road traffic noise level indicators rely solely on Linear Regression models that use either LA10(18H) or LA10(1H) as an input variable. However, a crucial correlation between LA10(18H) and other important indicators is missing from the literature. These indicators are often re-quired to be assessed by regulatory authorities, such as the Queensland Department of Transport and Main Roads. This paper extends the prediction scope of regression models to include important indicators such as LA10(12H), LAeq(15H), LAeq(9H), Max|LA10(1H)|, Max|LAeq(1H)|, and Max|LAmax|. The study also incorporates additional road traffic factors as input variables and compares the performance of several machine learning regression methods. The principal conclusion is that Random Forest consistently yields the lowest prediction error across all indicators, with improvements ranging from 20% to 40% compared to the current Linear Regression approach. Furthermore, averaged indicators perform best when using LA10(18H), annual average daily traffic (AADT), average traffic speed, and the percentage of heavy vehicles as input variables, while maximum-based indicators additionally require the road pavement type. Finally, the R-squared values for the different noise level indicators reach up to 98%, indicating a substantial enhancement in the accuracy of noise level indicator predictions.
Published in the Proceedings of ACOUSTICS 2025, 12-14 November 2025, Joondalup, Australia which will be made available on the Australian Acoustical Society (AAS) website: AAS Conference Proceedings
