Abstract: Nowadays, Machine Learning (ML) techniques are ubiquitous. They are applied to many contexts, both in production and research fields. Recently, they have shown strong performance in predicting road traffic noise levels, even when used at different sources and receiver distances. Due to their nature, they use the distance between source and receiver as any other feature, instead of considering it as one of the most important ones. This raises an important question: does distance really have a minor role, or does it influence the relationships between input variables and noise outputs in a non-easily perceivable way? In this contribution, the authors pursue an exploratory data analysis (EDA) approach to examine if and how source-receiver distance affects an ML model for road traffic noise estimation. This study highlights how the relationships between input variables and output sound levels change at different distances from the source. Statistical distributions, correlation patterns, and feature relevance indicators are investigated for two sensors recording the same noise events at different distances from the source. The results show that the relationship between variables and noise levels is different between the two cases. These findings emphasize the need to supplement performance-based evaluations with exploratory and interpretative analyses, providing new insights into the role of spatial factors in data-driven noise modeling.

Keywords: Road traffic noise, Machine Learning, Exploratory Data Analysis, Noise modelling, Source– receiver distance, Feature importance, Data-driven modelling, Environmental acoustics

Pages: 960-970