HydroTransformer

A transformer framework that learns streamflow from distributed, gridded inputs and generalizes across space and time.

Code

Atmospheric river reaching the Pacific Northwest, GEOS-5 model, December 4, 2023. NASA Earth Observatory.

Most data-driven hydrologic models take basin-averaged (“lumped”) inputs, which throws away how rainfall, temperature, and land surface vary within a basin. HydroTransformer uses the distributed, gridded inputs directly and models them with a spatiotemporal transformer.

What it shows

The model is evaluated on how well it generalizes to basins and time periods not seen in training, which is the setting that matters for prediction in ungauged locations and under a changing climate.

Publications

EAAI Under review

From Aggregated to Distributed: A Transformer Framework for Generalizable Streamflow Modeling Across Space and Time

Muhammad Talha, A. Pouyan Nejadhashemi, and Xiaofeng Liu

Engineering Applications of Artificial Intelligence · 2026

Code

@unpublished{talha2026hydrotransformer,
  title = {From Aggregated to Distributed: A Transformer Framework for Generalizable Streamflow Modeling Across Space and Time},
  author = {Talha, Muhammad and Nejadhashemi, A. Pouyan and Liu, Xiaofeng},
  note = {Under review, Engineering Applications of Artificial Intelligence},
  year = {2026}
}