Traffic Noise Prediction: Model Performance and Variable Importance

Table 2 lists the relative importance of predictor variables for each model. Motorway and secondary road length, residential road length, and residential area contributed most to both LR and XGB models but to varying degrees. There was less importance placed on roadway variables for NN, and none of them were in the top 10 importance for RF. Traffic volume contributed highly in both RF and XGB (43 and 34%, respectively) but far less so in LR and NN, where it was only ranked as the third most important variable, contributing 9 and 6%, respectively. Compared to RF and XGB, commercial area was not an important predictor in LR and it appears that the ability of a model to account for this variable is an important factor in prediction performance (discussed below). Meteorology played an important role in the models with atmospheric pressure ranking second most important in the XGB model, contributing 6%.

In our simulated fixed-site monitoring comparison, we found the performance metrics for this 'pseudo-fixed site' sample to be distinctly poorer than those obtained from mobile monitoring (Supporting Information, Figure S6 and Table S2), suggesting that the spatial information gathered with a mobile approach allowed a more complete utilization of the prediction variables when building prediction models. However, this 'pseudo-fixed site' test is limited by the lack of temporal information that would be present in true fixed-site sampling.

Prediction Maps. Traffic noise predictions on all roads with 20 m spatial resolution from LR (Figure 5a), NN (Figure 5b), RF (Figure 5c), and XGB (Figure 5d) are presented on the same LAeq (dB) scale for comparison. As expected, motorways have the highest predicted traffic noise levels, followed by primary and secondary roads; areas that are primarily commercial were also louder. Overall, LR and NN, which relied less on traffic volume, predicted well on motorways but did not capture spatial variability well on residential and secondary roads in the study area. RF captured spatial variability in the predictions on residential and secondary roads but under predicted LAeq on motorways. The best-performing model, XGB, captured the heterogeneity in different road types and land uses and accurately predicted higher LAeq on motorways.

Examining specific routes (Table 1), we note that road lengths and traffic volume played a significant role in a model's route performance. For example, route 12 had a mix of residential streets with low traffic volumes and motorways with high traffic volumes (Supporting Information, Figure S7). Because of this, RF, which had no road length variables of importance (Table 2), performed very poorly (R2 of 0.46 and RMSE of 5.31 dB) compared to XGB (R2 of 0.81 and RMSE of 3.83 dB). In a sensitivity test, when motorway length was removed from the XGB model, the R2 of route 12 dropped from 0.81 to 0.41, with RMSE increasing from 3.83 to 4.54 dB. Interestingly, routes 10 (Supporting Information, Figure S8) and 13 are also close to a motorway, but the RF model performance was good (R2 = 0.76 and 0.78, respectively), likely because the traffic volumes on these routes were uniformly higher.

We also found that XGB better captured higher traffic noise over commercial areas, mostly located on the west side of Long Beach. Since commercial area was not among the top 10 predictor variables in LR, it did not capture this spatial aspect and tended to underestimate high LAeq levels in commercial areas. While RF captured a general west-to-east spatial trend of traffic noise due to the effect of commercial area, it tended to underestimate high LAeq levels near motorways, since it did not consider roadway lengths based on the variable importance results. In a sensitivity test, when the commercial area variable was removed from XGB and RF, the LORO CV for all the routes on the west side of Long Beach, including routes 5, 8, 9, 13, 14, and 15, all had higher RMSE than shown in Table 1.

In general, we believe that the wider mix of important variables in the XGB model (road features, land use, meteorology) allowed better overall model performance across varying route types than the other three models that used more limited sets of variables. This may also be reflected in Figure 5 where XGB results show the greatest spatial heterogeneity.