.png){kind=icon}
.png){kind=icon}
Return
Data-Driven Surface Ozone Forecasting Using Node-Based Gated Spatiotemporal Fusion
.png){kind=icon}
PRE
AI
DOI：10.1109/tgrs.2026.3725283.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reliable ground-level ozone ( $\mathrm{O_{3}}$ ) forecasting is critical for mitigating impacts on public health and environmental sustainability and for supporting geoscience- and remote sensing-driven air quality applications. Traditional chemical transport models (CTMs), such as GEOS-Chem, remain widely used; however, their high computational cost and comparatively coarse spatial resolution limit timely, fine-grained guidance. This article introduces spatiotemporal node adaptive fusion (ST-NAF), a framework that integrates a graph attention network (GAT) with Transformer encoders to model intersite spatial relationships and temporal evolution in a unified architecture. Daily $\mathrm{O_{3}}$ observations from the Tropospheric Ozone Assessment Report (TOAR) dataset for 1999–2017 are used to train an ST-NAF-based model for forecasting. Relative to GEOS-Chem, the ST-NAF-based model reduces prediction bias by 55.61% in rural areas, 55.99% in suburban areas, and 55.90% in urban areas, while improving computational efficiency for large-scale forecasting. Using the United States subset of the global TOAR dataset of daily $\mathrm{O_{3}}$ ppb (parts per billion) from 1999 to 2017, the model ingests 30-day site histories, constructs distance-informed graph edges, and forecasts site-level $\mathrm{O_{3}}$ concentrations, achieving the above bias reductions versus GEOS-Chem with lower compute. The compact design enables deployment in low-resource environments and provides a practical, data-driven complement to remote sensing-oriented ozone assessment and decision support.
Keywords:
Atmospheric pollution
deep learning (DL)
forecasting
ground-level ozone
spatiotemporal modeling
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W