Ou et al. (2026) Foundation‐Scale Satellite Embeddings Reframe Hydrological Generalization as a Representation Problem

⚠️ Warning: This summary was generated from the abstract only, as the full text was not available.

Identification

• Journal: Geophysical Research Letters
• Year: 2026
• Date: 2026-04-03
• Authors: Zhigang Ou, Yupeng Zheng
• DOI: 10.1029/2025gl121604

Research Groups

Google (AlphaEarth Foundations model development)

Short Summary

This study addresses the challenge of transferring hydrological predictive skill across time and distinct basins by introducing 64-dimensional satellite embeddings (SE) from Google's AlphaEarth Foundations model, demonstrating consistent and significant performance improvements over static-attribute baselines in Australian catchments.

Objective

• To enhance the transferability of hydrological predictive skill across time and among distinct basins, particularly under nonstationary conditions, by leveraging dynamic surface characteristics encoded in foundation-scale satellite embeddings.

Study Configuration

• Spatial Scale: 455 Australian catchments
• Temporal Scale: 2017–2022

Methodology and Data

• Models used: SE-enhanced hydrological models; AlphaEarth Foundations model (for generating embeddings).
• Data sources: Google's AlphaEarth Foundations model (providing 64-dimensional satellite embeddings that encode dynamic surface characteristics).

Main Results

• SE-enhanced models consistently outperform static-attribute baselines in hydrological prediction.
• Temporal gains are most pronounced in regions under intense human disturbance, showing a median relative error reduction (RER) of 11.5%.
• For the most isolated basins, SE reshapes the feature space into a continuous, globally connected representation, yielding a median RER of 32.1%.

Contributions

• Introduces foundation-scale satellite embeddings (from Google's AlphaEarth Foundations model) as a novel approach to encode dynamic surface characteristics for hydrological modeling.
• Demonstrates a fundamental shift in hydrological generalization from traditional parameter transfer towards representation learning.
• Offers a new pathway for robust hydrological prediction under nonstationary conditions and for basins that appear distinct based on static attributes.
• Provides a continuous, globally connected representation of the feature space, overcoming the fragmentation caused by discrete static attributes.

Funding

Not specified in the abstract.

Citation

@article{Ou2026FoundationScale,
  author = {Ou, Zhigang and Zheng, Yupeng},
  title = {Foundation‐Scale Satellite Embeddings Reframe Hydrological Generalization as a Representation Problem},
  journal = {Geophysical Research Letters},
  year = {2026},
  doi = {10.1029/2025gl121604},
  url = {https://doi.org/10.1029/2025gl121604}
}