Real-Time Spatiotemporal NO2 Air Pollution Prediction with Deep Convolutional LSTM through Satellite Image Analytics

Abstract:

Nearly one in ten deaths on Earth is caused by ambient air pollution, and it makes up nearly 15% of all deaths in developing nations and COVID-19 fatality rates are being assessed for a link to air pollution. Air pollution costs the United States over $790 billion each year or roughly 5% of the U.S. total GDP. Nitrogen dioxide (NO2) is an air pollutant most commonly emitted by road traffic via the burning of fuel. In the greater Los Angeles area, there are around 27 million tons of NO2 in the atmosphere, nearly doubling the next leading U.S. metropolitan area. In fact, Los Angeles county residents are exposed to 60% more vehicular pollution than the state average. To combat the dangers of air pollution, we must utilize collected data from various sources and forecast day-to-day concentration levels. Concentration changes in NO2 are dependent on location, time, meteorological conditions, and a plethora of spatiotemporal features. To forecast NO2 in Los Angeles county, we collected 5 years of satellite data imagery from the ESA Sentinel-2 satellite equipped with a 13-band spectral imagery instrument. We also utilized image-based meteorological data of wind vectors, relative humidity, precipitation, and surface pressure. Through a complex deep learning model, we predicted NO2 in Los Angeles County 10 days in advance using the data of 10 days in the past. First, we interpolated the effects of meteorological data on satellite image data. We then used a sequential encoder-decoder ConvLSTM structure that combined spatial predictive models (deep Convolutional Neural Networks) and temporal predictive models (deep long short-term memory). Predicting the air pollution in Los Angeles County over time and space allows us to study potential health risks that the pollutants may have on residents, how governmental policies and laws affect pollution, and what interventions work best. This provides us with the opportunity to address this situation in advance.