Self-Supervised Framework Reconstructs Subhourly Weather Fields from Coarse Archives

Researchers have developed a token-guided self-supervised framework capable of reconstructing subhourly atmospheric fields directly from coarse data archives without requiring subhourly supervision labels. Reported in Communications Earth & Environment, the framework addresses a key operational limitation in atmospheric science, where most reanalyses provide data every 1 to 3 hours even though fast-evolving weather events develop on 5- to 15-minute timescales and influence planning and human activities.

Standard machine-learning approaches to temporal downscaling generally require high-frequency supervision labels, which restricts their utility when fine-scale training data are unavailable. To bypass this requirement, the presented architecture learns temporal relationships directly from coarse historical archives. The system encodes temporal context and the requested output time as discrete tokens, incorporating them via cross-attention to generate intermediate states consistent with the surrounding coarse fields.

Across all evaluated settings, the token-guided framework achieved lower reconstruction errors than both recurrent and attention-based baselines.

When applied to the 1806–1900 subset of the NOAA Twentieth Century Reanalysis version 3, the framework successfully produced 15-minute ensembles of near-surface relative humidity and 2-meter temperature. These reconstructions facilitate detailed analyses of short-term variability within historical meteorological records and extend event-scale weather analysis to earlier eras characterized by sparse observational coverage.

Subject of Research: Subhourly temporal reconstruction of atmospheric reanalysis fields

Article Title: Self-supervised temporal reconstruction of reanalysis fields at subhourly scales

Article References: Wang, L., Li, Q., Zhou, Z., Pan, B., Li, X., Zhao, L., Li, S., Peng, X., & Wang, Y. (2026). Self-supervised temporal reconstruction of reanalysis fields at subhourly scales. Communications Earth & Environment. https://doi.org/10.1038/s43247-026-04211-6

Image Credits: AI Generated

DOI: 10.1038/s43247-026-04211-6

Keywords: atmospheric reanalysis, temporal reconstruction, self-supervised learning, weather modeling, machine learning, NOAA Twentieth Century Reanalysis, relative humidity, surface temperature, Self-supervised, temporal, reconstruction, reanalysis