Abstract by Katrina Pedersen
Physics and Astronomy
Machine learning-based prediction of outdoor ambient sound levels: Ensemble averaging and feature reduction
Outdoor ambient sound levels can be predicted from machine learning-based models derived from geospatial and acoustic training data. To improve modeling robustness, median predicted sound levels have been calculated using tuned models from different supervised machine learning modeling classes. The spread in the ensemble provides an estimate of the modeling accuracy, which can be used to guide future data collection. Additionally, an initial analysis of feature importance metrics suggests that the number of geospatial inputs can be reduced from 120 to 15 without significant degradation of the model’s predictive error, as measured by leave-one-out cross validation. However, the predictions from the reduced-feature modeling may be less physical in certain regions. These results suggest the need for more sophisticated data collection and validation methods.