Uncertainty Quantification for Deep Learning Streamflow Reconstruction in Ungauged Basins

Reconstructing historical streamflow in ungauged basins remains a fundamental challenge in hydrology. This is especially true in data-sparse regions where infrastructure planning requires long-term discharge records that do not exist. Deep learning models trained on large-sample datasets can predict streamflow at locations excluded from training. However, a critical question persists: without observations, how can we assess reconstruction reliability? In this work, we develop and evaluate a framework for streamflow reconstruction in truly ungauged basins. We use two recurrent neural network architectures—Long Short-Term Memory (LSTM) and Mamba—trained on globally distributed catchments from the Caravan dataset. Training basins are selected using shape-based time-series clustering with Dynamic Time Warping. This ensures hydrological similarity to target regions. Models are driven by fused multi-source precipitation products (ERA5-Land, CHIRPS, MSWEP, CPC) alongside static catchment attributes. No local calibration is required. We propose ensemble disagreement—the spread among independently trained model instances from cross-validation—as a proxy for reconstruction quality. On a 100-basin holdout set, we demonstrate a negative correlation between ensemble disagreement and Nash-Sutcliffe Efficiency: basins where models agree tend to achieve higher reconstruction skill. This relationship provides practitioners with a principled basis for assigning confidence to streamflow reconstructions in ungauged basins, even in the absence of ground truth.