Large scale air pollution prediction with deep convolutional networks

Although considerable success has been achieved in urban air quality prediction (AQP) with machine learning techniques, accurate and long-term prediction is still challenging. One of the key issues for existing AQP approaches is that air quality monitoring stations are sparsely distributed, typically with around ten monitoring stations per city. As air quality may change abruptly in a local area, it is difficult to perform AQP accurately in areas that are far away from observation points. In addition, due to the large distance between every two monitoring stations, we cannot effectively leverage spatial relations among them to improve the AQP accuracy. In this paper, thanks to the development of low-cost air quality sensors, we are now able to collect a large-scale air quality dataset with 393 deployed air quality monitoring stations in a 120 km × 70 km region, which is more than ten times denser than existing AQP datasets. Further, we present a novel method to handle the data effectively. Specifically, we first convert the observed data from irregularly distributed monitoring stations into a regular image-like pollution map, which can then be processed with advanced deep convolutional networks. The experimental results show that the proposed approach can simultaneously model the temporal and spatial relations in our large-scale densely-observed dataset, leading to significantly improved AQP results.