[Paper Review] Urban Air Pollution Forecasting: a Machine Learning Approach leveraging Satellite Observations and Meteorological Forecasts

Giacomo Blanco, Luca Barco+2 more|arXiv (Cornell University)|May 30, 2024

Air Quality Monitoring and Forecasting4 citations

TL;DR

This study proposes a machine learning framework that forecasts urban air pollution using Sentinel-5P satellite data, meteorological variables, and topographical features, achieving a mean absolute percentage error (MAPE) of approximately 30% for five major pollutants in Milan. The models are trained without reliance on ground monitoring stations, enabling deployment in data-scarce urban areas.

ABSTRACT

Air pollution poses a significant threat to public health and well-being, particularly in urban areas. This study introduces a series of machine-learning models that integrate data from the Sentinel-5P satellite, meteorological conditions, and topological characteristics to forecast future levels of five major pollutants. The investigation delineates the process of data collection, detailing the combination of diverse data sources utilized in the study. Through experiments conducted in the Milan metropolitan area, the models demonstrate their efficacy in predicting pollutant levels for the forthcoming day, achieving a percentage error of around 30%. The proposed models are advantageous as they are independent of monitoring stations, facilitating their use in areas without existing infrastructure. Additionally, we have released the collected dataset to the public, aiming to stimulate further research in this field. This research contributes to advancing our understanding of urban air quality dynamics and emphasizes the importance of amalgamating satellite, meteorological, and topographical data to develop robust pollution forecasting models.

Motivation & Objective

• Address the challenge of urban air pollution forecasting in cities lacking ground monitoring infrastructure.
• Improve predictive accuracy by integrating multi-source data: satellite observations (Sentinel-5P), meteorological forecasts, and topographical features.
• Develop a scalable, station-independent forecasting system applicable across urban areas using grid-based predictions.
• Enable proactive public health and policy interventions through reliable day-ahead pollution forecasts.
• Stimulate further research by publicly releasing a comprehensive urban air quality dataset.

Proposed method

• Train multiple machine learning models—XGBoost, SGD regression, and linear regression—on a fused dataset combining satellite-derived pollutant concentrations, meteorological variables, and topographical data.
• Use a temporal window approach with past data windows of 1, 7, and 14 days to capture short- and medium-term pollution trends.
• Apply annual cross-validation by withholding one year of data at a time for validation while training on the remaining years.
• Incorporate spatial gridding (500 m resolution) to extrapolate predictions across the entire urban area beyond monitoring station locations.
• Leverage multi-temporal satellite observations to capture dynamic changes in pollutant levels over time.
• Optimize model performance using standard regression metrics: MAE, MAPE, and RMSE, evaluated across all pollutants and window sizes.

Experimental results

Research questions

• RQ1 Can machine learning models effectively forecast urban air pollution using only satellite, meteorological, and topographical data without reliance on ground monitoring stations?
• RQ2 How does the inclusion of multi-temporal satellite data and meteorological variables improve prediction accuracy compared to traditional methods?
• RQ3 What is the optimal temporal window (1, 7, or 14 days) for past data in predicting next-day pollution levels across different pollutants?
• RQ4 How do different machine learning algorithms (XGBoost, SGD, linear regression) compare in performance across multiple pollutants and window settings?
• RQ5 To what extent can the model generalize across urban areas through spatial gridding, enabling predictions in regions without monitoring stations?

Key findings

• The XGBoost model achieved the best overall performance across all pollutants and evaluation metrics, consistently outperforming SGD and linear regression.
• The mean absolute percentage error (MAPE) for all five pollutants (PM2.5, PM10, NO2, O3, SO2) was approximately 30% across all temporal windows.
• For PM10, longer temporal windows (14 days) improved prediction accuracy, while for PM2.5 and NO2, a seven-day window yielded the best results.
• O3 and SO2 predictions were most accurate with a one-day temporal window, indicating distinct temporal dynamics for these pollutants.
• Despite higher MAE values, SGD regression achieved better MAPE for SO2 under one- and seven-day windows, suggesting improved relative error performance.
• The model demonstrated robustness in extrapolating predictions across the urban grid, enabling pollution mapping in areas without monitoring stations.