Predicting PM2.5 atmospheric air pollution using deep learning with meteorological data and ground-based observations and remote-sensing satellite big data

Air pollution is one of the world's leading factors for early deaths. Every 5 s, someone around the world dies from the adverse health effects of air pollution. In order to mitigate the effects of air pollution, we must first understand it, find its patterns and correlations, and predict it in advance. Air pollution prediction requires highly complex predictive models to solve this spatiotemporal problem. We use advanced deep learning models including the Graph Convolutional Network (GCN) and Convolutional Long Short-Term Memory (ConvLSTM) to learn patterns of particulate matter 2.5 (PM 2.5) over spatial and temporal correlations. We model meteorological features with a time-series set of multidimensional weighted directed graphs and interpolate dense meteorological graphs using the GCN architecture. We also use remote-sensing satellite imagery of various atmospheric pollutant matters. We utilize government maintained ground-based PM2.5 sensor data along with remote sensing satellite imagery using a ConvLSTM to predict PM2.5 over the greater Los Angeles county area roughly 10 days in the future using 10 days of data from the past in 46-h increments. Our error results on the PM2.5 predictions over time and along each sensor location show significant improvement over existing research in the field utilizing spatiotemporal deep predictive algorithms.