Cascade Spatial Autoregression for Air Pollution Prediction

Recent years have witnessed a growing interest in air quality prediction and a variety of predictions models have been applied for this task. However, all of these models only use local attributes of each site for prediction and neglect the spatial context. Indeed, the concentrations of air pollutants follow the first law of geography: everything is related to everything else, but nearby things are more related than distinct things. To that end, in this paper, we apply the spatial autoregression model (SAR) to air pollution prediction, which considers both local attributes and predictions from the neighborhoods. Specifically, as SAR can only handle a snapshot of spatial data but our input data are time series, we develop the cascade SAR, which is able to take care of both spatial and temporal dimensions without incurring extra computation. Finally, the effectiveness of the cascade SAR is validated on the dataset of the London Air Quality Network.