October 1 | 2018

Capturing the spatial variability of noise levels based on a short-term monitoring campaign and comparing noise surfaces against personal exposures collected through a panel study.

Fallah-Shorshani M, Minet L, Liu R, Plante C, Goudreau S, Oiamo T, Smargiassi A, Weichenthal S, Hatzopoulou M.

Environ Res. 2018 Aug 17;167:662-672. DOI: 10.1016/j.envres.2018.08.021 



Environmental noise can cause important cardiovascular effects, stress and sleep disturbance. The development of appropriate methods to estimate noise exposure within a single urban area remains a challenging task, due to the presence of various transportation noise sources (road, rail, and aircraft). In this study, we developed a land-use regression (LUR) approach using a Generalized Additive Model (GAM) for LAeq (equivalent noise level) to capture the spatial variability of noise levels in Toronto, Canada. Four different model formulations were proposed based on continuous 20-min noise measurements at 92 sites and a leave one out cross-validation (LOOCV). Models where coefficients for variables considered as noise sources were forced to be positive, led to the development of more realistic exposure surfaces. Three different measures were used to assess the models; adjusted R2 (0.44-0.64), deviance (51-72%) and Akaike information criterion (AIC) (469.2-434.6). When comparing exposures derived from the four approaches to personal exposures from a panel study, we observed that all approaches performed very similarly, with values for the Fractional mean bias (FB), normalized mean square error (NMSE), and normalized absolute difference (NAD) very close to 0. Finally, we compared the noise surfaces with data collected from a previous campaign consisting of 1-week measurements at 200 fixed sites in Toronto and observed that the strongest correlations occurred between our predictions and measured noise levels along major roads and highway collectors. Our validation against long-term measurements and panel data demonstrates that manual modifications brought to the models were able to reduce bias in model predictions and achieve a wider range of exposures, comparable with measurement data.