Deep Learning Reconstructs Air Pollution from Sparse Data in Paris
In Paris, scientists utilize deep learning techniques to create comprehensive air pollution maps from limited data collected by monitoring stations. The research centers on four main pollutants: NO2, O3, PM2.5, and PM10. Training occurs using simulated data, with validation against actual measurements from 9 to 28 stations. A diffusion-based generative framework is proposed and compared with deterministic models. Even with noise and variations in space, the models demonstrate significant structural similarity in simulated data and realistic spatial distributions in real-world observations. This research seeks to enhance assessments of pollution exposure and inform public health policies.
Key facts
- Deep learning used for urban air quality reconstruction from sparse observations.
- Focus on four pollutants: NO2, O3, PM2.5, PM10.
- Models trained on full-field simulation data.
- Evaluated on real-world data from 9 to 28 monitoring stations in Paris.
- Diffusion-based generative framework introduced.
- Benchmarked against deterministic deep learning models.
- High structural similarity on simulated validation data.
- Realistic spatial patterns produced on real-world observations.
Entities
Institutions
- arXiv
Locations
- Paris
- France