Abstract A persistent challenge in hydrological modeling is reliably transferring predictive skill across time and among basins that appear distinct based on conventional static attributes. Here we address this challenge by leveraging Google's newly released data set generated from the AlphaEarth Foundations model, providing 64‐dimensional satellite embeddings (SE) that encode dynamic surface characteristics. Evaluated over 455 Australian catchments (2017–2022), SE‐enhanced models consistently outperform static‐attribute baselines. Temporal gains are most pronounced in regions under intense human disturbance, where median relative error reduction (RER) reaches 11.5%, reflecting reduced mismatch between static representations and evolving land‐surface states. Overcoming the fragmentation caused by discrete static attributes, SE reshapes the feature space into a continuous, globally connected representation, yielding a median RER of 32.1% for the most isolated basins. These results indicate that foundation‐scale embeddings fundamentally shift hydrological generalization from parameter transfer toward representation learning, offering a new pathway for robust prediction under nonstationary conditions.
Ou et al. (Fri,) studied this question.