PM2.5 Air Pollution Prediction through Deep Learning Using Multisource Meteorological, Wildfire, and Heat Data

Abstract

Air pollution is a lethal global threat. To mitigate the effects of air pollution, we must first understand it, find its patterns and correlations, and predict it in advance. Air pollution is highly dependent on spatial and temporal correlations of prior meteorological, wildfire, and pollution structures. We use the advanced deep predictive Convolutional LSTM (ConvLSTM) model paired with the cutting-edge Graph Convolutional Network (GCN) architecture to predict spatiotemporal hourly PM2.5 across the Los Angeles area over time. Our deep-learning model does not use atmospheric physics or chemical mechanism data, but rather multisource imagery and sensor data. We use high-resolution remote-sensing satellite imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument onboard the NASA Terra+Aqua satellites and remote-sensing data from the Tropospheric Monitoring Instrument (TROPOMI), a multispectral imaging spectrometer onboard the Sentinel-5P satellite. We use the highly correlated Fire Radiative Power data product from the MODIS instrument which provides valuable information about the radiant heat output and effects of wildfires on atmospheric air pollutants. The input data we use in our deep-learning model is representative of the major sources of ground-level PM2.5 and thus we can predict hourly PM2.5 at unparalleled accuracies. Our RMSE and NRMSE scores over various site locations and predictive time frames show significant improvement over existing research in predicting PM2.5 using spatiotemporal deep predictive algorithms.

Keywords

• air pollution prediction
• spatiotemporal forecasting
• deep convolutional LSTM
• remote-sensing satellite imagery
• wildfire heat data
• meteorological data