Estimating PM2.5 concentration of the conterminous United States via interpretable convolutional neural networks

We apply convolutional neural network (CNN) model for estimating daily 24-h averaged ground-level PM_2.5 of the conterminous United States in 2011 by incorporating aerosol optical depth (AOD) data, meteorological fields, and land-use data. Unlike some of the recent supervised learning-based approaches, which only utilized the predictors from the location of which PM_2.5 value is estimated, we naturally aggregate predictors from nearby locations such that the spatial correlation among the predictors can be exploited. We carefully evaluate the performance of our method via overall, temporally-separated, and spatially-separated cross-validations (CV) and show that our CNN achieves competitive estimation accuracy compared to the recently developed baselines. Furthermore, we develop a novel predictor importance metric for our CNN based on the recent neural network interpretation method, Layerwise Relevance Propagation (LRP), and identify several informative predictors for PM_2.5 estimation.