AI4AirQuality: High-Resolution Air Pollution Downscaling
The project focuses on developing and benchmarking three advanced architectures: a convolutional baseline (U-Net), the Swin v2 Visual Transformer, and a novel spatial adaptation of the Modulated Adaptive Fourier Neural Operator (ModAFNO). Causal inference guides the selection of physically meaningful input features, enhancing model accuracy and interpretability. Trained on CAMS data and evaluated with FAIRMODE metrics, this work paves the way for scalable, high-resolution air quality insights to support informed, data-driven decisions.
Follow the developments on Github.
Mentors
- Martin Ramacher
- Johannes Bieser
- Johannes Flemming
- Miha Razinger
- Paula Harder
Participants
Kevin Monsalvez-Pozo
Marcos Martinez-Roig
Nuria P. Plaza-Martín
Víctor Galván Fraile
Francisco Granell-Haro