A novel multi-pollutant space-time learning network for air pollution inference

Detailed information about air pollution in space and time is essential to manage risks to public health. In this paper we propose a multi-pollutant space-time learning network (Multi-AP learning network), which estimates pixel-wise (grid-level) concentrations of multiple air pollutant species based on fixed-station measurements and multi-source urban features, including land use information, traffic data, and meteorological conditions. We infer concentrations of multiple pollutants within one integrated learning network, which is applied to and evaluated on a case study in Chengdu (4900 km², 26 April - 12 June 2019), where air pollutant (PM_2.5, PM₁₀ and O₃) measurements from 40 monitoring sites are used to train the network to estimate pollutant concentrations in 4900 grid-cells (1 km²). The Multi-AP learning network allows us to estimate highly-resolved (1 km × 1 km, hourly) air pollution maps based on pollutant measurements which cover less than 1% of the grid-cells with better accuracy compared to other approaches, and with significant computational efficiency improvements. The time-cost is 1/3 of the time-cost of modelling each pollutant individually. Furthermore, we evaluate the relative importance of features and find that the meteorological feature set is the most important followed the land use features. The proposed Multi-AP method could be used to estimate air pollution exposure across a city using a limited set of air pollution monitoring sites.