Can statistical propagation models be saved by real 3D city data?: A regionalized study of radio coverage in New York City

Accurate radio propagation models are critical for interference and spectrum management, especially in scenarios involving dynamic spectrum access and/or inter-system coexistence in dense heterogeneous networks, due to the effect of aggregate error margins in the interference estimate. Despite their established limited accuracy, statistical path loss models often continue to be used to predict radio coverage for emerging wireless networks. In reality, urban radio coverage is highly influenced by the complex idiosyncratic structure of a given city propagation environment. In this paper, we present a joint study of the radio coverage - via detailed propagation simulations in WinProp - and statistics of the urban layout from 3D city data for seven regions of New York City (NYC). Our results reveal a very large variability in the received signal strength (RSS) among the regions (difference of up to 13.5 dB in the median RSS), demonstrating the inadequacy of simple urban propagation models for setting interference limits with bounded safety margins. A few works in the early literature had proposed to improve the accuracy of statistical propagation models by incorporating correction terms that are site-specific, as derived from 3D building data. However, the validity of such an approach is supported solely by a few localized measurement campaigns. The results of our detailed regionalized study of radio coverage over 250 km2 of NYC suggest that urban coverage prediction errors depend on the city layout in a much more nuanced manner. We show that the distribution of building heights shows no apparent correlation with RSS, and that although building density is coarsely correlated with RSS for different areas of Manhattan, this does not hold for the Bronx. We thus argue that much more sophisticated correction factors would need to be developed to substantially improve the accuracy of statistical propagation models; alternatively, if this is not feasible, site-specific coverage estimate calibration is unavoidable.