A Physics-Informed Ensemble-Reconstructed Daily Streamflow Dataset for 816 Snow-Dominated Catchments Worldwide (1951–2023)

The dataset contains reconstructed daily streamflow records for 816 snowmelt-dominated catchments worldwide covering the period 1950–2023. Streamflow series were reconstructed using an observation-prioritized, performance-based weighted ensemble framework that integrates deep learning models and differentiable hybrid hydrological models. The reconstruction framework was developed to generate continuous daily streamflow records in catchments affected by incomplete observations and varying data availability.The dataset is distributed as a pickle file (ensemble_1950_2023_obs_prioritized.pickle). The file contains a dictionary in which each key corresponds to a unique basin identifier (e.g., GRDC_1), and each value is a pandas DataFrame containing daily streamflow records for that basin. The obs column contains the original observed streamflow records where available. The ensemble column provides continuous daily streamflow estimates for the entire study period, including both observed and missing intervals. The file gauge_information.csv contains geographic information and reconstruction statistics for the 816 snowmelt-dominated catchments included in the reconstructed streamflow dataset. The file includes the following variables: gauge_id: Unique identifier of the gauging station obtained from the Global Runoff Data Centre (GRDC). lat: Latitude of the gauging station (decimal degrees). long: Longitude of the gauging station (decimal degrees). reconstruction_rate: Percentage of reconstructed daily streamflow values relative to the total length of the streamflow record. This variable quantifies the proportion of missing observations that were filled using the ensemble reconstruction framework. Higher values indicate a greater degree of reconstruction, whereas a value of 0% indicates a fully observed streamflow record.