Multi-Pollutant Ground-level Air Pollution Prediction through Deep MeteoGCN-ConvLSTM

Air pollution is the fourth-largest threat to human health. The harmful effects of air pollutants have costed the global economy nearly $3 trillion. It is imperative that a solution for mitigating the harmful effects of the most pervasive ground-level air pollutants - Carbon Monoxide
$(CO)$
, Nitric Oxide
$(NO)$
, Nitrogen Dioxide
$(NO_{2})$
, Ozone
$(O_{3})$
, and particulate matter 2.5
$(PM2.5)$
- is implemented, especially in urban areas. Recent advances in deep learning such as the Convolutional Long Short Term Memory (ConvLSTM) architecture are capable of learning complex spatiotemporal patterns with multisource data. We propose a novel sequential encoder-decoder ConvLSTM architecture capable of predicting hourly
$CO, NO, NO_{2}, O_{3}$
, and
$PM2.5$
spatially continuously over Los Angeles. Our model utilizes multisource satellite imagery collected from the ESA Tropospheric Monitoring Instrument (TROPOMI), remote-sensing data collected by the NASA Moderate Resolution Imaging Spectroradiometer (MODIS) instrument onboard the NASA Terra+Aqua satellites, and site-monitoring sensor observations of atmospheric and ground-level air pollution, meteorological data, and wildfire data. Our results show that our MeteoGCN-ConvLSTM model is competitive with state-of-the-art approaches across all predicted air pollutants. Moreover, our results show the versatility of our model when provided with solely atmospheric satellite imagery and remote-sensing data as input.