Graph attention networks for daily PM10 prediction at unmonitored European monitoring sites: operational scope, diagnostics, and outlier-robust training
Abstract
Air quality networks measure pollutant concentrations only at instrumented sites, leaving most locations unobserved. Graph-based virtual-node methods increasingly fill this gap, but are usually summarised by a single leave-one-station-out (LOO) R2 that leaves unstated where the model can be trusted and why it fails. Here, the operating envelope is the object of study. A graph attention network (GAT) combining multi-pollutant lags, ERA5 meteorology, MODIS aerosol optical depth, and land-use composition predicts daily-mean PM10 at unmonitored locations on the 18-station Styrian (Austria) network for 2020–2023. Under LOO cross-validation with disjoint years, the GAT reaches a median R2 of 0.52, seed-stable and ahead of Lagged-IDW (0.26), Lagged-Kriging (0.22), and Ridge-AR (0.45). At two stations whose PM10 distributions lie outside the training support, performance degrades sharply. Ablation and attribution trace the failure to one component: inverse-distanceweighted imputation of held-out lag features interacting with the persistence-residual head; graph topology is immaterial. An outlier-gated imputation policy raises pooled mean R2 from 0.31 to 0.38, node-masking training cuts per-station variance fourfold, and zeroing all auxiliary inputs lifts both outlier stations to positive R2: the partition is induced by the feature pipeline, not the architecture. The partition reproduces on NO2, and the model transfers to the 33-station Bavarian network (median R2 0.39, no negative folds). The outof-scope stations are clean-air sites under-represented in an urban-industrial network, and the worst is flagged in advance by a static-feature outlier detector: a deployable signal of where virtual-node estimates can be trusted and where the network must grow in regime diversity.
Related
- Hybrid stacked bidirectional deep learning with attention based air pollution prediction for environmental sustainability
- Long-Term Monitoring of Greenhouse Gas Emissions and Interpretable Machine Learning Models for Methane Prediction in Dairy Cattle Slurry Storage.
- Operational chemical weather forecasting with the ECCC online Regional Air Quality Deterministic Prediction System version 023 (RAQDPS023) – Part 1: system description
- WarpFlow v1.0: A Single-Source CPU/GPU Shallow-Water Solver for Operational City-Scale Urban Flood Modeling
- A machine learning model for early season wildfire risk in Ghana’s Guinea savannah assessing discrimination, operational prevalence, spatial transferability, and CMIP6 climate scenarios
BibTeX
@misc{strbad2026graph,
title = {{Graph attention networks for daily PM10 prediction at unmonitored European monitoring sites: operational scope, diagnostics, and outlier-robust training}},
author = {Dejan Strbad and Valentino Petrić and Nikolina Račić and Francesco Mureddu and Mario Lovrić and Mario Lovric},
howpublished = {ChemRxiv},
year = {2026},
month = aug,
doi = {10.26434/chemrxiv.15008048/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008048/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Graph attention networks for daily PM10 prediction at unmonitored European monitoring sites: operational scope, diagnostics, and outlier-robust training
AU - Dejan Strbad
AU - Valentino Petrić
AU - Nikolina Račić
AU - Francesco Mureddu
AU - Mario Lovrić
AU - Mario Lovric
PY - 2026
DA - 2026/08/30
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008048/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008048/v1
AB - Air quality networks measure pollutant concentrations only at instrumented sites, leaving most locations unobserved. Graph-based virtual-node methods increasingly fill this gap, but are usually summarised by a single leave-one-station-out (LOO) R2 that leaves unstated where the model can be trusted and why it fails. Here, the operating envelope is the object of study. A graph attention network (GAT) combining multi-pollutant lags, ERA5 meteorology, MODIS aerosol optical depth, and land-use composition predicts daily-mean PM10 at unmonitored locations on the 18-station Styrian (Austria) network for 2020–2023. Under LOO cross-validation with disjoint years, the GAT reaches a median R2 of 0.52, seed-stable and ahead of Lagged-IDW (0.26), Lagged-Kriging (0.22), and Ridge-AR (0.45). At two stations whose PM10 distributions lie outside the training support, performance degrades sharply. Ablation and attribution trace the failure to one component: inverse-distanceweighted imputation of held-out lag features interacting with the persistence-residual head; graph topology is immaterial. An outlier-gated imputation policy raises pooled mean R2 from 0.31 to 0.38, node-masking training cuts per-station variance fourfold, and zeroing all auxiliary inputs lifts both outlier stations to positive R2: the partition is induced by the feature pipeline, not the architecture. The partition reproduces on NO2, and the model transfers to the 33-station Bavarian network (median R2 0.39, no negative folds). The outof-scope stations are clean-air sites under-represented in an urban-industrial network, and the worst is flagged in advance by a static-feature outlier detector: a deployable signal of where virtual-node estimates can be trusted and where the network must grow in regime diversity.
ER -