Air pollution has become an increasingly salient public health concern, given its association with lung cancer, cardiovascular disease, respiratory disease, and metabolic disease. Globally, the distribution of PM2.5 monitoring technologies are not equally distributed, contributing to PM2.5 data scarcity. Insufficient data can substantially limit government capacities for mitigating PM2.5 related mortality and public health risks. As a solution to this, we incorporate simple, globally accessible, ground level data such as temperature, geographical parameters such as latitude and longitude, and satellite observations into a deep learning architecture that predicts out-of-sample, ground level PM2.5 concentrations. Predictions are made using a deep learning model for three cities within the continental United States with varying efficacy. The learning algorithm works best for areas with spatially homogeneous air quality and low climactic variability.