Abstract by Mitchell Cutler
Physics and Astronomy
Using Machine Learning to Guide Geospatial Data Collection
Geospatial models for ambient “soundscapes” have wide-ranging applications from influencing urban planning, to public health, to social behavior. These models use machine learning to correlate acoustic data with local geography and then predict sound levels at other locations. The scope of this project focuses on the applications of geospatial models to better predict acoustical environments across the contiguous United States (CONUS). Current models, while being fairly accurate for describing general trends, are limited in their resolution because of a lack of data diversity. In this paper, several methods will be explored to increase model resolution and accuracy by determining where new data should be collected. These data collection locations will be determined by analyzing which geospatial features are underrepresented. Methods to find these underrepresented points will include k-means clustering and hierarchical clustering and will be validated against previous models.