Real-Time Spatiotemporal Air Pollution Prediction with Deep Convolutional LSTM Through Satellite Image Analysis

Abstract

Air pollution is responsible for the early deaths of seven million people every year in the world. The first and the most important step in mitigating the air pollution risks is to understand it, discover the patterns and sources, and predict it in advance. Real-time air pollution prediction requires a highly complex model that can solve this spatiotemporal problem in multiple dimensions. Using a combination of spatial predictive models (deep convolutional neural networks) and temporal predictive models (deep long short-term memory), we utilized the convolutional LSTM structure that learns correlations between various points of location and time. We created a sequential encoder-decoder network that allows for accurate air pollution prediction 10 days in advance using data of 10 days in the past in the county of Los Angeles on a nitrogen dioxide metric. Through a 5D tensor reformatting of air quality satellite image data, we provide a prediction for nitrogen dioxide in various areas of Los Angeles over various time periods.