Abstract
Reliable daily streamflow forecasts are essential for flood-risk management and water-resource planning. We evaluate Long Short-Term Memory models driven by spatially distributed gridded versus point-based ERA5-Land meteorological inputs and additionally benchmark lumped and Conv1D-encoder baselines across four catchments that span snow-dominated, rainfall-driven, and mixed hydro-climatic regimes. For each basin, we train LSTM models with gridded, point-based, lumped, and Conv1D-encoder inputs under identical architectures. Gridded inputs consistently improve performance in comparison with point-based inputs: median Nash-Sutcliffe Efficiency rises from 0.84 to 0.93 in the mountainous Middle Fork Flathead River and from 0.78 to 0.90 in the Uba River, while the rainfall-fed Queets basin shows a moderate gain from 0.62 to 0.69. In the mixed South Santiam basin, performance increases from 0.54 to 0.58, and snow-water equivalent contributes only when supplied in a gridded form. Across all basins, the gridded configuration also outperforms the lumped and Conv1D baselines. These results demonstrate that spatially distributed predictors capture topographic and climatic heterogeneity, producing more robust forecasts across diverse settings. The findings support operational adoption of gridded-input LSTMs, especially in regions with pronounced spatial variability, while highlighting the diminishing performance in more homogeneous low-relief basins.