AQINet: A multimodal deep convolutional neural network to predict Air Quality Index via satellite imagery and meteorological data

Abstract

Air pollution is a pressing ecological issue with significant impacts on both public health and the environment. Poor air quality is a major contributor to respiratory diseases and is linked to millions of deaths annually, but many countries cannot afford air monitoring equipment. This lack of data makes it difficult to assess the health and environmental risks resulting from pollutant exposure. To address this problem, we present a multimodal model to inexpensively predict air quality levels in densely populated areas. Our research leverages both satellite imagery and meteorological data to create accurate air quality predictions. We sourced urban and suburban satellite imagery from the National Agriculture Imaging Program, meteorological data from Open-Meteo, and air quality data from OpenWeatherMap, to create a dataset named AQISet. AQISet is publicly available and free to download. The goal was for the model to implicitly learn spatial features in each image, such as roads, greenery, and bodies of water, and then combine this info with meteorological data to predict AQI. Using multiple computer vision techniques, the model was able to predict AQI with a mean absolute error of 16 AQI and a classification accuracy of 77% based on the EPA’s AQI standards categories. Our results establish a baseline for AQI prediction from satellite imagery and are a vast improvement over state-of-the-art pre-trained general computer vision models.