In brief
The work models how atmospheric aerosols evolve across the planet by combining a global graph of observations with temporal patterns at several scales.
Contribution
MultiscaleTCNGraphSAGE combines multi-scale temporal convolutions, GraphSAGE spatial learning, CAMS aerosol observations and ERA5 meteorology.
Key finding
The proposed model improved all reported headline metrics over the strongest benchmark and better captured extreme aerosol events.
Models and data
Distinctive research elements
Models and methods
- MultiscaleTCNGraphSAGE
- Transformer+GCN
Data and evaluation
- CAMS AOD550
- ERA5 meteorology
Author-written overview
Research overview
This study forecasts global aerosol optical depth by combining monthly CAMS aerosol observations with meteorological variables from ERA5. The proposed MultiscaleTCNGraphSAGE architecture uses temporal convolutions with several kernel sizes and dilation rates to represent short- and long-range dynamics, while GraphSAGE layers learn spatial dependence across the global observation graph. Against the strongest evaluated benchmark, Transformer+GCN, the model reduces RMSE by 5.4% and MAE by 6.4%, while increasing the coefficient of determination by 2.7% and Pearson correlation by 1.3%. The results also indicate improved representation of extreme aerosol events.
Cite this work
BibTeX
@article{semper2026global,
title = {Global forecasting of aerosol optical depth through a deep learning spatiotemporal modeling},
author = {Marc Semper and Manuel Curado and Jose F. Vicent},
journal = {International Journal of Environmental Science and Technology},
year = {2026},
volume = {23},
number = {1},
eid = {69},
doi = {10.1007/s13762-025-06905-4}
}Always check the publisher record for final volume, issue and page information before citing.