Dynamic Air Pollution Estimation

Project Overview

This semester project, completed as part of a collaborative team effort, explores how artificial intelligence can help monitor air pollution more effectively across the globe.

Air quality monitoring today faces a major limitation: ground-based measurement stations are few and far between, while satellite data often lacks the detail needed for local-scale analysis. Our Project builds on the work of Scheibenreif et al. (2020) and added the capability for multiple pollutant predictions and temporal context.

What We Did

Working as a group, we developed machine learning models that:

• Combine multispectral satellite imagery with atmospheric data from European Space Agency missions (Sentinel-2 and Sentinel-5P) to predict pollution levels at a much finer resolution than satellites alone can provide
• Track pollution over time by creating monthly estimates instead of just long-term averages, helping to capture seasonal changes and pollution events
• Predict multiple pollutants including nitrogen dioxide (NO₂), particulate matter, and ozone from a single unified model

Key Insights

Through our collaborative work, we found that:

• Deep learning can fill the gap between sparse ground measurements and coarse satellite observations
• Adding temporal modeling helps capture dynamic pollution patterns throughout the year
• Multi-pollutant prediction is feasible but requires careful balancing of model complexity and performance
• Models trained on data-rich regions work in areas with sparse ground monitoring only with limited accuracy

Impact

This project demonstrates how AI and remote sensing can work together to provide better air quality information, especially in regions where traditional monitoring infrastructure is limited. The techniques developed could support public health initiatives, environmental policy, and scientific research on atmospheric pollution.

Demonstrator

Our project includes an interactive web application that allows users to visualize air pollution estimates from one specific area. It can be accessed here.