Using CNNs on Sentinel-2 data for road traffic noise modelling

Abstract

Urbanisation and road traffic noise go hand in hand. While the WHO and the European Environmental Agency are concerned about high noise levels and the respective adverse effects on health, detailed exposure maps are scarce. Utilizing highly accurate sound propagation models is expensive and scalable Land-Use Regressions (LUR) are often limited by the lack of available training data. Also, the portfolio of statistical models used in LURs so far has not been extended towards deep learning despite their recent contributions in urban remote sensing. By challenging a semantic segmentation network with the noise mapping problem, we aimed to test their capabilities. Different input channels, scoping road data, Sentinel-2 images, topographical data and a building model are compared against each other. The best performing model utilizes all eleven features and has an overall accuracy of 0.89. We suggest that future studies shall intensify experiments on input channels, learning strategy and spatial application.