Traffic Noise Prediction: Modeling and Evaluation Across Diverse Road Networks

Traffic noise prediction. For mapping, we created a 20 m × 20 m spatial resolution grid spanning the maximum and minimum latitude and longitude of the study area. Since we were concerned with predicting noise near roadways, we extracted only the P × Q grids that contained roads, clipping the roads in each grid into 20 m segments. We assumed that all predictor variables were uniform within a 20 m grid cell. For example, each grid contained a set of weather features Wg = {wg,1, wg,2, ..., wg, n }, where n denotes a specific weather variable (e.g., humidity). A similar procedure was taken for all road features, land use, and traffic data. We trained the four models with these input variables and traffic noise in each 20 m grid, denoted as ={ }∈  × L ll l , , ..., g iP 1 2 Q , as the target output. Given this structure, the trained models were then used to predict traffic noise along the road at each 20 m × 20 m grid cell. ■ RESULTS Summary Statistics. The 16 walking routes together with their noise levels are plotted in Figure 3, where proximity to larger roadways can be seen to relate to higher noise. Average maximum traffic volume for each route is also listed in Table 1, showing that routes with higher vehicles per hour tended to have higher LAeq (Table 1). The total number of measured traffic noise points after spatial aggregation (N = 6647) had a mean LAeq of 61.9 dB, a standard deviation (SD) of 6.91 dB, a median of 61.6 dB, and an interquartile range from 56.3 to 67.6 dB. The median and mean LAeq for each route were very close, differing typically by only a few tenths of a dB, indicating that short-term aberrant noise sources such as yard care machinery, dogs barking, or people passing while talking had little contribution after spatial aggregation. Model Results. Scatterplots of measured versus predicted LAeq from the 5-fold 30% test sets are shown in Figure 4. The LR with highest adjusted R2 and best Mallow’s Cp selected 34 out of 44 variables, which resulted in R2 = 0.78 and an RMSE of 4.29 dB. NN outperformed LR with a CV R2 of 0.91 and RMSE of 2.89 dB but was not as good as the tree-based RF or ensemble method XGB, which had a CV R2 s of 0.96 and RMSE of 1.83 and 1.82 dB, respectively. The slopes of the regression lines for RF and XGB were 0.90 and 0.91, respectively, indicating that both models slightly underestimated high levels of traffic noise but were better than LR, which fitted a slope of only 0.60. The LORO CV R2 and RMSE for each of the 16 routes are shown in Table 1, revealing that traffic noise prediction ability varied markedly by route location with R2 and RMSE values ranging from 0.54 to 0.88 and 2.54 to 6.44 dB, respectively. XGB predictions had relatively less route-to-route variability indicating superior model performance and the best ability to better predict noise across a variety of street types and land uses (Table 1). LR also performed well with an average R2 of 0.70 but had higher RMSE, averaging 4.98 dB. Both NN and RF had far poorer performance metrics with lower R2 s and higher RMSEs (Table 1). The relatively good performance of LR compared to NN and RF for LORO CV was partly explainable by the exhaustive LR variable selection process, which reduced the chance of overfitting compared to fitting on all variables.