Leveraging Satellite Data and Machine Learning to Improve Urban Air Quality

by Marco Arezio

Air pollution is one of the main challenges faced by modern cities, with significant consequences for public health and quality of life. The complexity of the dynamics contributing to the formation and dispersion of pollutants makes accurate forecasting of pollution levels in urban areas particularly challenging.

This article explores an innovative approach based on machine learning models that combine data from satellite observations and weather forecasts to improve the accuracy of air pollution predictions. Specifically, we focus on the metropolitan area of Milan, one of the Italian cities most affected by air quality issues.

Integrating Satellite Data and Weather Forecasts

The accuracy of air quality forecasts depends on the availability of reliable data and the ability to integrate them into complex models. Satellite observations provide valuable information on pollutant concentrations such as nitrogen dioxide (NO2), ozone (O3), and fine particulate matter (PM2.5 and PM10), covering vast geographical areas and offering a comprehensive view of atmospheric conditions. These data are integrated with local weather forecasts, which include key parameters such as temperature, humidity, wind speed, and direction—all of which significantly influence pollutant dispersion in the air.

Satellite observations are conducted using advanced instruments, such as the TROPOMI sensor onboard the Sentinel-5P satellite of the European Space Agency (ESA), which can detect concentrations of various atmospheric gases with high resolution. The integration of these satellite data with weather forecasts from models like the European Centre for Medium-Range Weather Forecasts (ECMWF) allows for a more accurate and detailed understanding of air quality in dynamic urban environments such as Milan.

Machine Learning Models for Pollution Prediction

To enhance the ability to forecast air pollution, machine learning models are employed to combine heterogeneous data and process them to identify hidden patterns and nonlinear relationships among variables. This study utilized various machine learning algorithms, including artificial neural networks (ANN), boosted decision trees (such as XGBoost), and support vector regression (SVR).

One advantage of machine learning is its capacity to train models on large datasets, identifying correlations that traditional modeling techniques may overlook. For instance, models can recognize how specific combinations of meteorological factors, such as atmospheric pressure or wind direction, can promote the accumulation or dispersion of pollutants. This type of analysis is particularly useful in densely populated areas like Milan, where emission sources are numerous and variable.

Focus on the Metropolitan Area of Milan

Milan is characterized by a high population density, a large concentration of industrial activities, and a geography that limits air circulation, making the city particularly vulnerable to pollutant accumulation. The proposed model was applied to predict levels of PM2.5 and NO2, considering their significant implications for public health.

By integrating satellite data and local weather forecasts, the machine learning models significantly improved the accuracy of predictions compared to traditional methods. Neural networks, in particular, proved effective in modeling nonlinear interactions between weather conditions and pollutant concentrations, while XGBoost excelled at handling the temporal variability of emissions, such as traffic peaks or adverse weather conditions.

Results and Future Perspectives

Simulation results demonstrated that the integration of satellite data enhanced the model's ability to capture sudden pollution events, such as those caused by stagnant weather conditions or temporary emission spikes. The forecasts showed improved performance compared to models based solely on meteorological data, with a significant increase in the precision of PM2.5 and NO2 level estimates.

This machine-learning-based approach has the potential to provide more accurate and timely information on air quality, enabling local authorities to implement preventive measures to mitigate population exposure to pollutants. In the future, improved access to high-resolution satellite data and increased computational capacity could further enhance the reliability of these models, making them essential tools for sustainable urban environmental management.

Conclusions

The air pollution forecasting approach based on the integration of satellite observations and weather forecasts, using machine learning techniques, represents a significant step forward in understanding and managing urban air quality. In a city like Milan, such models offer the possibility of anticipating critical situations and adopting preventive measures to reduce their impact on public health. While challenges remain regarding the complexity of the atmospheric system and the availability of increasingly precise data, the potential of these technologies for pollution monitoring and forecasting is extremely promising.

© Reproduction Prohibited

Photo: Wikimedia