Explainable machine learning models for outdoor exceedance level prediction based on geospatial variables

Area of Science:

• Environmental acoustics
• Geospatial data analysis
• Machine learning applications

Background:

• Sound level modeling is crucial for understanding acoustic environments.
• Previous studies focused on sound exceedance levels using specific datasets.
• Diverse environments, from national parks to urban areas, present unique acoustic challenges.

Purpose of the Study:

• To develop and analyze data-driven Random Forest regression models for predicting sound exceedance levels.
• To evaluate model performance using geospatial variables.
• To assess the impact of urban data on prediction accuracy.

Main Methods:

• Utilized a dataset of acoustic exceedance levels from diverse US locations.
• Applied advanced Python libraries to train Random Forest regression models.
• Incorporated 99 geospatial variables to predict sound levels.
• Developed 3 general and 5 ancillary data-driven models.

Main Results:

• Achieved promising predictive power with R-squared values ranging from 0.54 to 0.91.
• Root mean squared error varied between 1.77 and 5.97 dB.
• Models including more urban data demonstrated superior performance.
• Performance variability was linked to dataset limitations in diverse environmental coverage.

Conclusions:

• Data-driven models show significant potential for predicting outdoor sound levels.
• Urban acoustic data integration enhances model accuracy.
• Further development requires datasets covering a wider range of natural and urban environments.
• An interactive online dashboard enhances accessibility for non-experts.