In brief
Global atmospheric reanalyses like ERA5 exhibit persistent regional biases over complex terrain that can distort downstream hydrological, energy, and risk models. Combining local weather station networks (AVAMET) with spatial neighborhood graphs via GraphSAGE and Jumping Knowledge enables simultaneous correction across multiple meteorological variables. Rigorous chronological validation shows up to an 81% error reduction in pressure, humidity, temperature, and wind compared to classical baselines, while revealing clear operational boundaries for localized extreme precipitation.
Contribution
A multivariable graph neural network post-processing framework that captures spatial-neighborhood context and inter-variable dependencies to correct ERA5 biases, evaluated under leakage-free chronological splits, uncertainty bounds, and seasonal robustness tests.
Key finding
GraphSAGE with Jumping Knowledge reduces RMSE and MAE across six non-precipitation variables (surface pressure, relative humidity, 2-m temperature, wind speed, wind direction, and wind gust) by up to 81% over classical and tabular baselines, while extreme precipitation remains an open challenge for spatial graph post-processing.
Models and data
Distinctive research elements
Models and methods
- GraphSAGE + Jumping Knowledge
- Random Forest reference
- Oracle classical reference
Data and evaluation
- ERA5 reanalysis (ECMWF)
- AVAMET weather station network (Valencian Community)
Author-written overview
Research overview
Global meteorological reanalyses such as ERA5 provide spatially coherent atmospheric fields, yet sub-grid orography and complex coastal dynamics induce systematic local biases that limit their direct utility in regional hydrology, wildfire risk assessment, renewable energy, and agriculture. While traditional bias correction approaches operate variable-by-variable and tabular machine-learning baselines neglect spatial topology, this study introduces a multivariable graph-based post-processing framework designed to model both spatial-neighborhood structure and inter-variable physical dependencies. Using ground observations from the dense AVAMET weather station network across the Valencian Community (a Western Mediterranean region with complex terrain prone to severe convective instability), station records are projected onto the ERA5 grid to train a GraphSAGE architecture with Jumping Knowledge under strict, leakage-free chronological validation splits. The graph model is benchmarked against raw ERA5, an oracle classical reference, and a fixed Random Forest baseline using block-bootstrap uncertainty estimation, fixed-window seasonal robustness diagnostics, and spatial and cross-variable consistency evaluations. On the temporal test partition, GraphSAGE+JK achieves substantial error reductions across six non-precipitation variables—cutting RMSE and MAE for surface pressure, relative humidity, 2-m temperature, wind speed, wind direction, and maximum wind gust by up to 81% relative to classical models. Conversely, precipitation remains an unresolved challenge: the graph model does not improve aggregate pointwise accuracy or high-threshold extreme occurrence detection. These findings confirm the effectiveness of graph neural networks for local predictive bias correction of continuous meteorological variables while highlighting localized precipitation extremes as a key open frontier.
Cite this work
BibTeX
@article{semper2026graph,
title = {Graph-based multivariable bias correction of meteorological variables under strict temporal validation in a western mediterranean region prone to atmospheric instability},
author = {Marc Semper and Manuel Curado and Jose F. Vicent},
journal = {Journal of Computational Science},
year = {2026},
eid = {102988},
doi = {10.1016/j.jocs.2026.102988}
}Always check the publisher record for final volume, issue and page information before citing.