Development of an interpretable machine-learning model for capturing nonlinear dynamics of multisensory interactions in public open spaces

Abstract

This study developed an interpretable machine-learning model using SHAP (SHapley Additive exPlanations) to predict the soundscape quality in urban open spaces by integrating auditory, visual, and especially thermal perception features. Data collected from a field survey involving 24 public open spaces in Hong Kong, including perceptual responses from 1,882 participants and physical measurements, formed the basis of the analysis. The prediction model was constructed using the eXtreme Gradient Boosting (XGBoost) algorithm, as it outperformed the Random Forest (RF) algorithm and significantly exceeded the performance of traditional multiple linear regression (MLR) models in handling datasets with high contextual complexity. To circumvent the shortcomings of non-transparency and non-interpretable in machine-learning models, this study incorporated SHAP analysis to identify the most influential factors for open space soundscape and to investigate their interactions, which has not been revealed in earlier soundscape research. LAeq, perceived dominance of road traffic noise, greenery percentage and PET, were identified as the most influential factors. LAeq and perceived dominance of road traffic noise had significant negative effects, especially when LAeq exceeded 68 dBA. Greenery percentage showed a threshold effect, with dominant and positive contributions only when coverage exceeded 40 %. Notably, PET and its interaction with LAeq had a significant impact, with high temperatures (> 35 °C) intensifying the negative impacts of high sound pressure levels. These findings underscore the importance of “sound-visual-thermal” interactions and highlight that urban design should minimize the coexistence of high sound pressure levels, elevated temperatures, and heightened road traffic noise perceptions.