Machine Learning-Based Assessment and Prediction of Psychoacoustic Indicators of Road Traffic Noise

This study explores the use of machine learning techniques to evaluate and predict psychoacoustic noise characteristics associated with pass-by road traffic events. Using a dataset comprising various acoustic and psychoacoustic parameters such as LAeq, FS50, L10, L90, and spectral indicators, a comprehensive analysis was conducted to assess their predictive potential. Gradient Boosting, Random Forest, and ARIMA models were employed for different tasks, including both classification and time-series forecasting. In addition, feature engineering techniques were used to create composite variables and enhance model input quality, while sequence-based learning methods allowed for temporal dynamics to be captured. The best-performing Gradient Boosting model achieved R² = 0.63 and MAE = 0.122 in predicting LAeq and FS50 indicators. The dataset used consisted of 1,200 pass-by noise events from an open-access repository, including both acoustic (LAeq, L10, L90) and psychoacoustic (FS50, R50, N50, S50) metrics. The results highlight the capability of machine learning not only to improve the accuracy of psychoacoustic modeling but also to support real-time, perception-aware urban noise monitoring systems. Such approaches can enable more responsive and adaptive noise management strategies in smart city planning. These findings demonstrate the potential of ML-based models to inform proactive urban noise management and public health strategies.