Machine Learning-Based Sedigraph Reconstruction for Enhanced Sediment Yield Estimation in the Upper Blue Nile Basin

Sediment-laden runoff in Ethiopia’s Upper Blue Nile Basin (UBNB) threatens the ecological balance of Lake Tana and the operational efficiency of the Grand Ethiopian Renaissance Dam (GERD), a critical hydropower infrastructure. High sediment loads, driven by intense monsoon rainfall and erodible soils, exacerbate erosion and sedimentation, affecting water quality and the longevity of infrastructure across the basin. Limited data availability and complex hydrological dynamics hinder accurate measurement and modeling of sediment yields. This study reconstructs 1990–2020 sedigraphs for Gilgel Abay (1,664 km²) and Gumara (1,394 km²) watersheds using machine learning (Gradient Boosting (GB), Random Forest (RF), Quantile Random Forest (QRF)) and traditional sediment rating curves (SRCs), utilizing intermittent Suspended Sediment Concentration (SSC) measurements and continuous predictors (discharge, rainfall, temperature, and evapotranspiration). QRF provided the highest accuracy (R2 = 0.73–0.75), capturing non-linear sediment dynamics and quantifying uncertainty via its 0.05–0.95 conditional quantile bounds. The average annual sediment yields were 30.7±7.2 t/ha/yr for the Gilgel Abay watershed and 27.5±10.74 t/ha/yr for the Gumara watershed, with 95–96% occurring during the wet season (June–October). Despite challenges from incomplete datasets due to instrument malfunctions and logistical constraints during extreme rainy events, QRF’s robust predictions inform the application of targeted erosion control strategies, such as reforestation and urban runoff mitigation, to enhance sustainable management of the UBNB. This approach bridges the reconstruction of sediment concentration measurements with data gaps while improving hydrological process understanding, offering a scalable framework for sediment management in monsoon-dominated basins.