Machine Learning for Predicting Air Quality Using Land Images

Urban air quality profoundly impacts our daily lives and overall health. Unfortunately, air pollution levels have been escalating in many cities. To monitor, and eventually predict an dreduce pollution levels , a multitude of monitoring stations and low-cost sensors have been deployed to measure air pollution across urban areas over the past decade. This study focuses on the prediction of PM2.5 concentrations across a city using advanced machine learning techniques. One innovative aspect of our approach involves harnessing convolutional neural networks (CNNs) to extract urban features from satellite imagery. These urban characteristics, in conjunction with pollution data from representative monitoring stations and meteorological patterns, form the basis for our PM2.5 prediction model. Our research explores the spatial autocorrelation of PM2.5 data, revealing a significant correlation decrease between sensors as the distance between them increases. This observation suggests an opportunity to enhance the existing sensor network. We aim to develop improved methods for selecting a subset of representative sensors while preserving critical spatial variations in PM2.5 levels. By combining machine learning, satellite imagery, and spatial analysis, our study not only helps restoring PM2.5 data completeness but also offers insights into optimizing sensor deployment strategies for more effective urban air quality monitoring.