Dynamic Air Pollution Estimation
Project Overview
Accurately monitoring air pollution is a global challenge, currently limited by the sparse coverage of ground stations and the coarse resolution of satellite instruments. Building on pioneering work that fuses multi-source satellite data to estimate NO₂ at high resolution, this project aims to advance the field by introducing multi-pollutant modeling and enhanced temporal dynamics.
Our goal is to develop a deep learning framework capable of predictiong multiple pollutants and providing high-frequency, monthly estimates of air quality.
Methodology
The project extends existing deep learning architectures through two primary development tracks:
- High-Frequency Temporal Modeling: Moving beyond time-averaged data to model sequences, allowing for the prediction of pollutant dynamics and seasonal trends throughout the year.
- Multi-Target Prediction: Modifying the model architecture to simultaneously estimate concentrations of multiple key pollutants (such as PM, O₃, SO2) alongside NO₂, leveraging their co-emission patterns.
- Data Fusion: Integrating high-resolution optical imagery (Sentinel-2) with atmospheric measurements (Sentinel-5P) and meteorological datasets (e.g., wind and precipitation) to improve prediction accuracy.
Key Objectives
- Global Scalability: Training on data-dense regions and testing generalization on diverse geographic areas to ensure the model works globally.
- Monthly Resolution: producing dynamic air quality maps that reflect short-term pollution events rather than just long-term averages.
- Unified Framework: Merging temporal and multi-target components into a single, robust model validated against ground-truth networks like the EEA.
Expected Outcomes
This work aims to deliver a functional deep learning model capable of generating detailed, monthly air quality maps for multiple pollutants. The final results will include a scientific analysis of performance trade-offs in multi-task settings and a valid assessment of the model's ability to predict temporal trends across different environmental contexts.