National road traffic noise estimation with ensemble learning and multimodal geodata

Abstract

The European Noise Directive mandates the mapping of noise – high, continuous sound pressure levels considered to be a major health threat. However, the strictest rulesets apply to specific regions only and the majority of residential areas are unmapped. Transfer learning was deployed to close spatial data gaps between the official, strategic road traffic noise maps. The three most suitable hyperparameter configurations achieved weighted Kappa values (a measure of ordinal agreement) ranging between 0.889 and 0.956 during repeated cross-validation. The best model achieved an overall classification accuracy of 90.7 % when tested against held-out samples. 7.8 % of predictions exhibited minor deviations within ± 5 dB(A). The model was subsequently deployed to predict road traffic noise across Germany at 10 x 10 Meter resolution for 2017. The results suggest a total of 13.1 million people exposed to yearly averaged road traffic noise (Lden) above 55 dB(A) and stress need for improved noise policies.