Weather Foundation Models as Earth-Observation Backbones: Forecasting Satellite Soil Moisture from a Frozen Latent

Data-driven weather foundation models encode a spatially complete, physically coupled, time-evolving state of the atmosphere. We ask whether such a frozen representation can be decoded into a land-surface variable it never ingests, supervised by satellite observations. ForeSM (FOundation-model REpresentations for Soil Moisture), a <1M-parameter per-patch head on the frozen latent of the Aurora weather model, is trained against soil-moisture observations of the NASA Cyclone Global Navigation Satellite System (CYGNSS) constellation with a masked loss over sparse data points; soil moisture and precipitation appear nowhere in Aurora's inputs or outputs. The head yields gap-free 6-hourly soil moisture that generalizes to two held-out years (r=0.946, RMSE 0.040 m³ m⁻³). Autoregressive rollouts provide 1–14-day forecasts that beat persistence at every verifiable lead, by a margin widening from +0.008 to +0.026 in r, and are competitive with operational NWP soil-moisture predictions and outperform ECMWF IFS forecasts. An input ablation isolates the contribution of the representation: the identical, capacity-matched head trained on Aurora's raw meteorological inputs reaches a within-cell anomaly correlation of 0.39, compared with 0.50 from the frozen latent representation and 0.27 from a weather-blind baseline using only static fields and calendar features. The gain therefore comes not merely from access to weather inputs, but from the information encoded in Aurora's learned representation.