National ground-level NO2 predictions via satellite imagery driven convolutional neural networks

Abstract

Outdoor air pollution, specifically nitrogen dioxide (NO2), poses a global health risk. Land use regression (LUR) models are widely used to estimate ground-level NO2 concentrations by describing the satellite land use characteristics of a given location using buffer distance averages of variables. However, information may be leaked in this approach as averages ignore the variances within the averaged region. Therefore, in this study, we leverage a convolutional neural network (CNN) architecture to directly pass data grids of various satellite data for the prediction of U.S. national ground-level NO2. We designed CNN architectures of various complexity which inputs both satellite and meteorological reanalysis data, testing both high and low resolution data grids. Our resulting model accurately predicted NO2 concentrations at both daily (R2 = 0.892, RMSE = 2.259, MAE = 1.534) and annual (R2 = 0.952, RMSE = 0.988, MAE = 0.690) temporal scales, with coarse resolution imagery and simple CNN architectures displaying the best and most efficient performance. Furthermore, the CNN outperforms traditional buffer distance models, including random forest (RF), feedforward neural network (FNN), and multivariate linear regression (MLR) approaches, resulting in the MLR performing the poorest at daily (R2 = 0.625, RMSE = 4.281, MAE = 3.102) and annual (R2 = 0.758, RMSE = 2.218, MAE = 1.652) scales. With the success of the CNN in this approach, satellite land use variables continue to be useful for the prediction of NO2. Using this computationally inexpensive model, we encourage the globalization of advanced LUR models as a low-cost alternative to traditional NO2 monitoring.