Satellite Image Atmospheric Air Pollution Prediction through Meteorological Graph Convolutional Network with Deep Convolutional LSTM
Pratyush Muthukumar, Emmanuel Cocom, Kabir Nagrecha, Jeanne Holm, Dawn Comer, Anthony P. Lyons, Irene Burga, Chisato Fukuda Calvert, Mohammad Pourhomayoun
Abstract
Every five seconds, somebody around the world prematurely dies from the effects of air pollution. Air pollution is one of the world's leading risk factors for death. To mitigate the deadly effects of air pollution, it is imperative that we understand it, discover the patterns and sources, and predict it in advance. Air pollution prediction in real-time requires extremely powerful models that can solve this spatiotemporal problem in multiple dimensions. We used an advanced graph convolutional network coupled with a deep convolutional LSTM model to learn patterns over the spatial and temporal dimen-sion in real-time. Our model employs a graph convolutional network that models meteorological features and extracts high-level embeddings through unsupervised representation learning. We created a sequential encoder-decoder deep convolutional LSTM that allows for accurate and efficient satellite image based atmospheric Nitrogen Dioxide air pollution prediction over Los Angeles county 10 days into the future using data from 10 days in the past through the use of spatiotemporal satellite imagery and meteorological graph embedding inputs. Our results for predicting spatially continuous atmospheric Nitrogen Dioxide in Los Angeles over various time periods shows improvement in prediction over previous research done on this topic.