AQ-Net PM2.5 reanalysis map over Northern China

Lead author · SCIA 2025 · Springer LNCS

AQ-Net

Predicting pollution where no sensor dares to go. A deep spatio-temporal network for air-quality reanalysis.

← All projects

Most cities in the world are partially monitored: a handful of stations cover a few neighbourhoods, the rest is silent. AQ-Net bridges that gap, learning to reanalyse air quality across an entire region using only the sparse stations that exist, plus reanalysis weather and time encodings.

The model combines three building blocks: an LSTM with multi-head attention for long-range temporal dependencies, a cyclic encoding that lifts discrete time (day-of-year, hour-of-day) into continuous space, and a learnable neural kNN module that interpolates from sensed locations to unsensed ones in feature space rather than physical space.

AQ-Net was accepted at SCIA 2025 (Scandinavian Conference on Image Analysis) and published in Springer LNCS. I presented it as a speed talk and poster in Reykjavík.

SCIA

Venue
2025, Reykjavík

LSTM

Temporal
+ multi-head attention

kNN

Spatial
learnable neural

5y

Training data
2013–2017, N. China

Core ideas

  1. Time-aware encoding. LSTM stacked with multi-head attention captures long-term dependencies in the pollutant signal. A cyclic encoding projects discrete calendar time into continuous space so the model never sees a hard discontinuity at midnight or new year.
  2. Spatial interpolation by neural kNN. Instead of inverse-distance weighting in physical space, AQ-Net learns a feature-space embedding where nearest neighbours mean similar pollution dynamics. This handles topography and meteorology implicitly.
  3. Reanalysis-style outputs. The model can fill gaps (where stations are missing) and run for short-term (6–24 h) and long-term (2–7 days) horizons.
  4. Real-world data. Trained and evaluated on Northern China stations from 2013–2017, covering the post-2013 air quality crisis era when data quality and density became reliable.

Poster & visualisations

AQ-Net SCIA 2025 poster SCIA 2025 poster (click for PDF) AQ-Net architecture Architecture: LSTM + attention + neural kNN AQ-Net pollution map over China Reanalysed PM2.5 across Northern China Animated PM2.5 reanalysis Animated reanalysis (PM2.5) Cyclic time encoding visualisation Cyclic encoding of calendar time Upscaled PM2.5 prediction map Upscaled PM2.5 prediction map

Resources

Springer LNCS arXiv Code on GitHub Poster (PDF)

Credits

Authors
A. Kheder, B. Foreback, L. Wang, Z.-S. Liu, M. Boy
Venue
SCIA 2025, Scandinavian Conference on Image Analysis (Reykjavík), Springer LNCS
Data
Northern China air-quality stations, 2013–2017
Affiliation
LUT University, Atmospheric Modelling Centre (AMC-Lahti)

← Back to all projects