CRAN-PM 1 km PM2.5 predictions across Europe

Lead author · 2026 · under review

CRAN-PM

Cross-Resolution Attention Network for daily 1 km PM2.5 prediction across Europe, with zero-shot transfer to North America and India.

← All projects

CRAN-PM is a dual-branch Vision Transformer that predicts daily PM2.5 at 1 km resolution across all of Europe. A single forward pass produces a 29-million-pixel map in about 1.8 seconds on a single GPU.

The architecture combines a coarse global branch (25 km, captures continental dynamics) with a fine local branch (1 km, captures urban-scale detail), connected by cross-resolution attention and an elevation-aware bias. A wind-driven patch reordering module mirrors the design used in TopoFlow.

The most striking result is zero-shot transfer: trained only on European data (2017–2021), the model produces accurate PM2.5 predictions over the United States, Canada and India without any fine-tuning, including capturing the 2023 Canadian wildfire smoke plume drifting into the eastern US.

1 km

Resolution
across Europe

96M

Parameters
dual-branch ViT

1.8s

Inference
29M pixels / map

0-shot

Transfer
USA, Canada, India

Key ideas

  1. Two-branch cross-resolution attention. Coarse 25 km global branch captures synoptic atmospheric flow; fine 1 km local branch captures urban gradients; cross-attention fuses them.
  2. Elevation-aware attention. A learnable bias keyed on a digital elevation model lets the network respect topographic constraints on pollutant transport.
  3. Wind-driven patch reordering. Patches are re-sorted along the wind direction so attention flows downstream of sources.
  4. Zero-shot continental transfer. Trained on Europe (2017–2021), tested on Europe (2022), then evaluated unchanged on USA, Canada and India. The model captures wildfire smoke and Indo-Gangetic Plain hotspots out of the box.
  5. PixelShuffle decoder. Progressive upsampling from coarse latent to 1 km output, no blurry interpolation.

Visualisations

CRAN-PM dual-branch architecture Two-branch cross-resolution architecture Europe-wide ground truth vs prediction Europe-wide evaluation: ground truth, prediction, RMSE Spatial detail Po Valley, Paris Spatial detail: Po Valley, Paris and others Temporal evolution of PM2.5 Temporal evolution across six regions Zero-shot USA wildfire smoke Zero-shot USA & Canada (wildfire smoke) Zero-shot India Indo-Gangetic Plain Zero-shot India: Indo-Gangetic Plain

Resources

Project website arXiv preprint Code on GitHub

Credits

Lead author
Ammar Kheder
Status
Preprint on arXiv (2026), under review
Compute
LUMI supercomputer (AMD MI250X GPUs), distributed training
Framework
PyTorch with Distributed Data Parallel
Affiliation
LUT University, Atmospheric Modelling Centre (AMC-Lahti)

← Back to all projects