Work

Earth System Foundation Models for Predicting Extreme Climate Events Through Integrated Satellite, Oceanic, and Land-Surface Intelligence

Year: 2026
Type: article
Abstract: : This study investigates the scientific and operational potential of Earth system foundation models for anticipating extreme climate events through the joint assimilation and representation of atmospheric, oceanic, and land-surface information. The manuscript develops a rigorous synthesis framework built on real observational and reanalysis data sources, including ERA5, SMAP, NOAA OISST v2.1, GPM IMERG, MODIS surface products, and the NOAA IBTrACS archive. Rather than treating weather, ocean, and land processes as isolated forecasting domains, the proposed framework formalizes extreme-event prediction as a multiscale sequence-learning problem in which precursor signals propagate across coupled geophysical subsystems. Particular attention is given to rapidly intensifying tropical cyclones, compound flood events, and heat-related marine and terrestrial extremes, because these hazards expose the limits of conventional single-domain models and the benefits of cross-sphere intelligence. The methodological contribution consists of Earth system intelligence pipeline that integrates heterogeneous spatial resolutions, temporal frequencies, and physical variables into a unified tokenized representation suitable for transformer-based foundation modeling. The paper specifies real data collection procedures, harmonization rules, loss functions, and event-centric evaluation protocols using track error, anomaly reconstruction, rare-event likelihood, and warning-skill diagnostics. To maintain realism, the empirical section relies on documented public data and published benchmark outcomes from peer-reviewed or official sources rather than synthetic examples. Comparative evidence from Aurora, Pangu-Weather, GraphCast, and related systems is interpreted alongside real extreme-event indicators from Hurricane Otis (2023), the Pakistan floods (2022), and North Atlantic marine heat anomalies (2023). The results indicate that Earth system foundation models achieve meaningful gains when the representation space is informed by coupled oceanic and land-surface precursors rather than atmospheric state alone. Published evidence shows that Aurora exceeded operational baselines on 92% of high-resolution weather targets, 86% of ocean-wave targets, 74% of air-quality targets, and 100% of evaluated tropical-cyclone-track targets, while Pangu-Weather reduced 3-day and 5-day tropical-cyclone direct-position errors relative to ECMWF-HRES. Case evidence further shows that extreme events consistently emerge from interacting anomalies in sea-surface temperature, moisture availability, antecedent soil state, and circulation structure. Accordingly, the paper concludes that Earth system foundation models offer a credible pathway toward next-generation early-warning systems, provided that uncertainty quantification, interpretability, and equitable operational deployment remain central to future research. Keywords: Earth system foundation models; extreme climate events; satellite intelligence; ocean-atmosphere coupling; land-surface intelligence; tropical cyclone prediction; flood forecasting; climate AI
Language: English
Open Access status: bronze
OpenAlex ID: W7147638416
FWCI: 16.28
Cites: 1
Cited by: 10
Related to: 10