Predicting traffic noise – a scalable approach using land-use regression

Abstract

Notable noise mapping obligations exist in the European Union. However, they are limited to designated areas such as large agglomerations and main traffic infrastructures and therefore exclude certain populations from exposure assessments and consecutive noise action planning. Existing maps are not spatially congruent to epidemiological cohorts like the German National Cohort (NAKO). As noise mapping is frequently confined by the necessary resources and data, we searched for economic alternatives for area-wide noise mapping making use of spaceborne earth observations. Using remote sensing methods, we built a geostatistical model embracing the arrangement of sources of noise and the surrounding environment in which the sound propagates. In our experimental set-up, we relied on publicly available noise data, context-aware feature engineering and a machine learning model. Eventually, the scalable approach explained 78% of the variations and can be deployed for predictions at a high spatial granularity of 10x10 meters. With it, we aim to spatially close the blank spots in existing noise maps allowing to assess noise exposure beyond already mapped urban populations in suburban and rural areas as well.