AI-generated satellite embeddings may give hydrologists a better tool for predicting river flow in places where no measurement stations exist.
Researchers tested AlphaEarth Foundation embeddings — representations learned from large satellite image datasets rather than designed by domain experts — against traditional basin attributes like soil type, terrain, and land cover. When used to train hydrological models on known basins and then predict flow in unmonitored ones, the embedding-based models achieved higher accuracy. The embeddings capture vegetation patterns, land surface properties, and long-term environmental dynamics in a single learned representation, compressing complexity that hand-crafted attributes routinely miss.
The ungauged basin problem is not new, but it is consequential: a significant share of the world's river basins lack flow records, making flood forecasting and water management harder in exactly the places that can least afford bad predictions. The finding that a foundation model trained on satellite imagery can substitute for — and outperform — decades of expert-designed hydrological descriptors is a meaningful shift in how that gap might be closed.
The study also found that basin selection matters: using the embeddings to identify environmentally similar donor basins improved predictions, while lumping in dissimilar basins degraded them — a reminder that better representations do not eliminate the need for careful experimental design.