Watershed-Agnostic Forecasting

An explainable, distributed CNN–LSTM hydrologic model that is not tied to a single watershed.

Paper Code

Michigan's Lower Peninsula, VIIRS on NOAA-20, October 19, 2024. NASA Earth Observatory.

A CNN–LSTM that takes distributed (gridded) inputs and predicts streamflow without being trained for one specific watershed. Convolutional layers capture spatial patterns within a basin, and the LSTM captures how they evolve over time.

Explainability

Attribution analysis shows which regions and which inputs drive each prediction, so the model can be checked against hydrologic understanding rather than trusted as a black box. Published in Environmental Modelling & Software.

Publications

EMS Journal

Watershed-agnostic modeling: An explainable, distributed, data-driven hydrologic model

Muhammad Talha and A. Pouyan Nejadhashemi

Environmental Modelling & Software, 205, 107149 · 2026

DOI Code

@article{talha2026watershed,
  title = {Watershed-agnostic modeling: An explainable, distributed, data-driven hydrologic model},
  author = {Talha, Muhammad and Nejadhashemi, A. Pouyan},
  journal = {Environmental Modelling \& Software},
  volume = {205},
  pages = {107149},
  year = {2026},
  publisher = {Elsevier},
  doi = {10.1016/j.envsoft.2026.107149}
}