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

BACKGROUND: In modern societies, noise is ubiquitous. It is an annoyance and can have a negative impact on human health as well as on the environment. Despite increasing evidence of its negative impacts, spatial knowledge about noise distribution remains limited. Up to now, noise mapping is frequently inhibited by the necessary resources and therefore limited to selected areas. OBJECTIVE: Based on the assumption, that prevalent noise is determined by the arrangement of sources and the surrounding environment in which the sound propagates, we build a geostatistical model representing these parameters. Aiming for a large-scale noise mapping approach, we utilize publicly available data, context-aware feature engineering and a linear land-use regression (LUR) model. METHODS: Compliant to the European Noise Directive 2002/49/EG, we work at a high spatial granularity of 10 × 10-m resolution. As reference, we use the day–evening–night noise level indicator L(den). Therewith, we carry out 2000 virtual field campaigns simulating different sampling schemes and introduce spatial cross-validation concepts to test the transferability to new areas. RESULTS: The experimental results suggest the necessity for more than 500 samples stratified over the different noise levels to produce a representative model. Eventually, using 21 selected variables, our model was able to explain large proportions of the yearly averaged road noise (L(den)) variability (R(2) = 0.702) with a mean absolute error of 4.24 dB(A), 3.84 dB(A) for build-up areas, respectively. In applying this best performing model for an area-wide prediction, we spatially close the blank spots in existing noise maps with continuous noise levels for the entire range from 24 to 106 dB(A). SIGNIFICANCE: This data is new, particular for small communities that have not been mapped sufficiently in Europe so far. In conjunction, our findings also supplement conventionally sampled studies using physical microphones and spatially blocked cross-validations.