Traffic noise assessment in urban Bulgaria using explainable machine learning

• Land-use regression models for traffic noise were developed using machine learning. • Extreme gradient boosting (XGB) performed best, explaining 68.9 % of the noise variability. • Shapley additive explanations unpacked the black-box nature of the XGB. • Road traffic noise estimates are openly available. • Almost 97 % of the urban population was at risk of experiencing harmful traffic noise. Fine-grained noise maps are vital for epidemiological studies on traffic noise. However, detailed information on traffic noise is often limited, especially in Eastern Europe. When acoustic noise propagation models are unavailable, rigid linear noise land-use regressions are typically employed to estimate noise levels; however, machine learning likely offers more accurate noise predictions. We innovated by comparing the predictive accuracies of supervised machine learning models to estimate traffic noise levels across the five largest Bulgarian cities. In situ A-weighted equivalent continuous sound levels were obtained from 232 fixed-site monitors across these cities. We included transport- and land-use-related predictors using 50–1,000 m buffers. Extreme gradient boosting (XGB) had the highest ten-fold cross-validated fit ( R² =0.680) and the lowest root mean square error (RMSE=4.739), insignificantly besting the random forest-based model ( R² =0.667, RMSE=4.895). Support vector regression ( R² =0.633, RMSE=5.358), elastic net ( R² =0.568, RMSE=5.625), and linear regression ( R² =0.548, RMSE=5.569) performed significantly worse. Shapley values for the XGB showed that the length of major roads within 100 m buffers, footways within 50 m buffers, residential roads within 50 m buffers, and the number of buildings within 50 m buffers were important non-linear predictors. Our spatially resolved noise maps revealed striking geographic noise variations and that, on average, 96.8 % of the urban population experiences harmful noise levels.