Interpretable Machine Learning of Urban Noise Levels from Street-Scale Morphology and Functionality

• Data-driven, non-causal framework links street morphology to urban noise. • Daytime in situ noise measurements capture street-scale spatial variability. • Street functional hierarchy dominates noise exposure across cities. • Interpretable ML reveals nonlinear effects and key urban design indicators. • Results support noise screening and planning in data-constrained cities. Urban noise exposure is shaped not only by traffic intensity but also by the spatial configuration and functional hierarchy of streets. Traditional prediction approaches, however, often prioritize traffic flow while underrepresenting the role of urban form. This study examines how street-scale morphological and functional characteristics are associated with environmental noise levels across four Chilean cities with contrasting urban structures. A total of 530 daytime in situ measurements of the equivalent continuous sound level L A e q were combined with 71 urban predictors derived from municipal records and open-source data within 150 m buffers around each sampling location. Three ensemble learning models (Random Forest, XGBoost, and LightGBM) were optimized using Bayesian hyperparameter search and evaluated under intra-city validation, achieving strong explanatory power ( R ² = 0.65–0.95) and high accuracy ( MAE ≈ 1.5–2.2 dBA). Cross-city evaluation using a leave-one-city-out scheme revealed moderate performance degradation when predicting unseen urban contexts, highlighting both the potential and limits of territorial transferability. Model interpretability analyses based on LightGBM gain, SHAP values, and LIME explanations consistently identified street functional hierarchy as the dominant indicator associated with urban noise exposure, followed by roadway capacity and land-use intensity. Spatial patterns revealed elevated noise levels (>65 dBA) along higher-order arterial corridors, while quieter conditions were typically associated with lower-category residential streets. Overall, the results show that interpretable machine learning provides a transparent, data-driven, and non-causal framework for linking urban form and function to street-level noise patterns, supporting noise screening and urban planning in data-constrained contexts.