Predicting streamflow with LSTM networks using global datasets

Streamflow predictions remain a challenge for poorly gauged and ungauged catchments. Recent research has shown that deep learning methods based on Long Short-Term Memory (LSTM) cells outperform process-based hydrological models rainfall-runoff modeling, opening new possibilities prediction in basins (PUB). These studies usually feature local datasets model development, while at global scale require training datasets. In this study, we develop LSTM over 500 catchments from the CAMELS-US data base using ERA5 meteorological forcing catchment characteristics retrieved with HydroMT tool. Comparison against an trained shows that, latter generally yields superior performances due to higher spatial resolution (overall median daily NSE 0.54 vs. 0.71), results most of Western North-Western US (median 0.83 0.78). No significant changes performance occur when substituting sources deriving characteristics. encourage further worldwide streamflow available Promising directions include different regions world quality forcing.