Reconstruction of global long-term daily streamflow dataset using machine learning models for revealing streamflow changes

Study regions: 314 major global watersheds. Study focus: In the context of widespread streamflow observation data gaps and significant basin heterogeneity worldwide, this study aims to construct a high-precision, long-term global daily-scale streamflow reconstruction dataset. Focusing on 314 major river basins globally (1980–2020), we systematically evaluate the performance of four machine learning models—Random Forest (RF), Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—in reconstructing streamflow sequences and identify optimal modeling strategies suitable for different basin conditions. To reduce input data uncertainty, the study concentrates on reconstructing streamflow at basin outlets, leveraging their larger catchment areas and relatively reliable meteorological forcing information. The resulting dataset provides a high-quality resource for analyzing global streamflow variability and its climatic drivers. New hydrological insights for the region: Based on the reconstructed dataset, a study of long-term streamflow patterns in global river basins, with a focus on the changing characteristics of extreme flows (high and low flows) and their climatic drivers, reveals the following: African river basins show the highest proportion of significant increasing trends in both long-term and extreme streamflow; South America and Australia have a relatively large number of river basins (approximately 58 % and 59 %, respectively) where long-term streamflow shows a significant decrease. Globally, the proportion of river basins with significantly increasing mid- to long-term streamflow is generally below 10 %, while a higher proportion of basins show rising trends in low flows and mean flows, reflecting a possible trend toward wetter conditions in most regions. Additionally, ENSO plays an important regulatory role in streamflow variability, particularly in tropical regions, where El Niño and La Niña events correspond to significant alternating dry and wet anomalies in streamflow responses.