Machine learning-based 2D and 3D traffic noise modelling in campus environments: a framework for vertical façade exposure assessment

Traditional noise mapping methods often overlook the complex spatial dynamics and vertical variability associated with urban noise propagation. In this study, we present a machine learning (ML) framework that integrates both 2D and 3D modelling for traffic noise prediction in a campus environment. We employed multilayer perceptron (MLP), extreme gradient boosting (XGB), and the traditional CoRTN model to predict equivalent continuous sound levels (LAeq) based on traffic, topographic, and land use factors. The XGB model achieved the highest performance with an R² = 0.95, MAE = 0.91, MSE = 2.75. Model interpretation and robustness were examined using SHAP and global sensitivity analysis, identifying road proximity as the dominant predictor, with traffic composition and spatial context variables acting as secondary, interaction-dependent factors. High-resolution 2D noise maps captured spatial patterns aligned with observed data, while the ML-based 3D voxel model produced detailed façade-level noise profiles. A consistent vertical attenuation trend was observed, with high exposure (> 65 dB) decreasing from 55% at ground level to less than 8% by the fourth floor. These findings provide planners and designers with a data-driven tool to identify façade-level noise exposure and apply targeted mitigation strategies. This study represents an early application of a voxel-based ML approach for 3D façade noise prediction in a campus environment.