Machine learning for predicting soundscape: From individual-level deterministic models to group-level probabilistic models

Urban environments are used by a large number of diverse people, but existing soundscape prediction models are focused on perceptual outcomes of an idealised average individual. With respect to developing group-level soundscape prediction models, it remains unclear which factors are important for predicting soundscapes and which types of models perform better for that task. Therefore, by relying on the International Soundscape Database, this study aims at determining which factors can be used to predict soundscape and which model performs better at the group level. In this study, methods, such as correlation analysis, are used to select demographic, acoustic, visual, and geographic information factors that are significantly correlated with soundscapes. Subsequently, this study compares the performances of four models—linear regression, random forest, XGBoost, and gaussian process regression (GPR)—in soundscape prediction tasks conducted at the individual and group levels. The results show that the equivalent sound pressure level (|r|>0.31), roughness (|r|>0.34), total harmonic distortion (|r|>0.31), relative approach (|r|>0.30) and vegetation (|r|>0.48) are important to the soundscape prediction. The performance of the GPR model is better than the other three models at the individual level (R_ISOPleasant²=0.36, MAE_ISOPleasant=0.26, RMSE_ISOPleasant=0.33, R_ISOEventful²=0.18, MAE_ISOEventful=0.23, RMSE_ISOEventful=0.29). At the group level, the performance of the GPR model is also relatively high (KL_ISOPleasant=0.81,DME_ISOPleasant=0.26,DME_ISOEventful=0.38). This study identifies the key acoustic and visual factors of soundscape perception and demonstrates the advantages of GPR. The introduction of a probability distribution–based framework is expected to predict soundscape at the group level and offer guidance for urban sound environment design.