A comparison of methods for modeling soundscape dimensions based on different datasetsa).

Area of Science:

• Environmental acoustics
• Psychoacoustics
• Statistical modeling

Background:

• Soundscape studies face challenges due to diverse methodologies and metrics.
• Assessing the suitability of statistical modeling techniques in soundscape research is difficult.

Purpose of the Study:

• To compare five statistical methods and two performance metrics for modeling soundscape Pleasantness and Eventfulness.
• To evaluate model performance across different soundscape datasets and predictor variables.

Main Methods:

• Applied linear and nonlinear regression, including machine learning approaches (random forest, extreme gradient boosting), to three soundscape datasets.
• Utilized seven acoustic and three sociodemographic predictors.
• Employed fixed-effects and mixed-effects models, assessing performance using out-of-sample R2.

Main Results:

• Models generally performed better for predicting Eventfulness than Pleasantness.
• Nonlinear methods consistently outperformed linear regression, with comparable performance among nonlinear techniques.
• Mixed-effects models yielded more generalized predictions but slightly lower performance on unseen groups.

Conclusions:

• Nonlinear machine learning approaches are superior for soundscape modeling, particularly for Eventfulness.
• Cross-validation with specialized splitting is recommended for small, imbalanced soundscape datasets.