The challenges of mapping quiet urban areas from mobile sound surveys

Detailed spatial data on noise levels are needed to identify quiet areas. Yet noise maps derived from sound propagation models based on traffic flows or sensor networks are often incomplete. Surveys using mobile sound recording coupled with machine learning modelling have offered a possible solution, but so far have focused on averaged noise levels. We argue that quiet areas should ideally be consistently quiet, rather than simply quiet on average, yet mobile survey data usually lack the full temporal profile needed to quantify consistency in noise levels. Using data from our Southampton study area, we show that modification of the loss function used in machine learning modelling of mobile survey data can yield information on the likely range of noise levels at any location across an entire city. We use this approach not only to identify locations that are consistently quiet, but also to help understand which urban features may be associated with quiet, average and noisy events at any given location.