Adapting a U-Net to model road traffic noise from geographic data

Abstract

Urbanization and road traffic noise are closely linked with each other. The impact of excessive noise levels on human’s health is of major concern for organizations such as the World Health Organization and the European Environmental Agency. However, precise exposure maps are scarce, as highly accurate sound propagation models are constrained by huge computational efforts and high costs. Alternative large-scale Land-Use Regressions (LUR) are limited by a lack of available training data and model-specific constraints. Despite the recent contributions of deep learning to geo-spatial data analysis, the portfolio of statistical models for noise modeling has not been expanded to include deep learning. By leveraging cost-efficient geodata and artificial intelligence, accurate and efficient noise exposure mapping can be achieved. Therefore, in this study, Sentinel-2 satellite images, topographic radar data and a building model are used to model road traffic noise (Lden). This noise indicator Lden, is common in health and environmental exposure sciences and shows the equivalent sound level over a 24-hour period with a penalty for evening and nighttime noise. Averaged over a whole year, it is – compared to the fluctuate nature of sound – static and therefore comprehensive to map using geoinformatic tools. At the same time, sound is spatially distributed very unevenly and a high granularity of at least 10 x 10 Meter is obligatory. With respect to the amount of data needed for deep learning, our experiments scope 70 German cities with an overall area of 10,956 km². Eleven input features related to noise emission and sound propagation are used to evaluate the performance of different input feature combinations using a Resnet50-U-Net model. Our best performing model, using all eleven features, achieved an overall accuracy of 89% and an Intersection over Union (IoU) score of 0.68 which indicates a good fit of the model. Furthermore, we conclude that future research should continue investigations on conducive input features, learning strategies, and geographical transferability. Only then, the ambitious goal of producing highly accurate but also cost-efficient, large scale and high-resolution exposure maps can be met.