DeepAir: Deep learning and satellite imagery to estimate high-resolution PM2.5 at scale

Air pollution, specifically PM2.5, has become a significant global concern owing to its detrimental impacts on public health. Even so, the high-resolution monitoring of air pollution is still a challenge on a global scale. To cope with this, machine learning techniques have been utilized to infer the concentration of air pollutants at a fine scale. In this study, we propose DeepAir, a learning framework for estimating PM2.5 concentrations at a fine scale with sparsely distributed observations. DeepAir integrates a pre-trained convolutional neural network with the LightGBM method. This framework estimates the PM2.5 concentration of a given patch, utilizing a synergy of geographical information, meteorological conditions, and satellite observations. We select California as the focal region and train the model with data from 2014 to 2017 provided by 130 PM2.5 observation stations in the state. Upon training, the model can be applied to estimate the daily PM2.5 concentrations at 1km resolution across California. Our methodology meticulously incorporates meteorological variables, with a particular emphasis on wildfire propagation, and contemplates the complex interplay of various features. To ascertain the efficacy of our model, we employ the 10-fold cross-validation technique, which confirms that our model surpasses traditional machine learning and standalone deep learning methods.