Assessing temporal and spatial generalization of LSTMs for streamflow modeling in French watersheds with and without European training data

French watersheds, characterized by their diverse climatic, geological, topographical, and spatial characteristics. Long short-term memory (LSTM) networks are emerging as a powerful solution for streamflow simulation in temporal (TI), spatial (SI), and spatio-temporal induction (STI). However, their application to French watersheds remains underexplored. This study evaluates LSTM performance in the three tasks across 310 French watersheds over 1987–2012 and investigates whether expanding training data to include 501 European basins improves streamflow simulations. Six experiments were conducted with LSTM trained on French basins (LSTM-TI-FR, LSTM-SI-FR, LSTM-STI-FR) and with LSTM trained on both French and European basins (LSTM-TI-EU, LSTM-SI-EU, LSTM-STI-EU). Results show that LSTM performs best in the TI task (median Kling-Gupta efficiency (KGE) = 0.78), with 90 % of watersheds achieving KGE ≥ 0.65. LSTM-SI-FR has satisfactory but lower median performance (KGE = 0.68). Results vary across watersheds, where 22 %, particularly those with low mean flow and variability, display poor accuracy (KGE < 0.48). LSTM-STI-FR provides the lowest performance. Larger basins (>1400 km²) with higher mean flows (>12 m³/s) perform better across all tasks, while low-flow basins are consistently less accurate. Unexpectedly, expanding training data with European basins decreased model performance slightly (−0.02 in median KGE). Overall, this study demonstrates LSTMs perform well for TI in gauged basins, but more research is needed for SI/STI use. • LSTM performs best for temporal, then spatial, followed by spatio-temporal induction • Large French basins with high mean flow are better represented • Low-flow basins with lower variability presents greater challenge • More than doubling training dataset size did not improve LSTM performance