Suspended Sediment Fluxes and Decadal Trends in the Humid Tropics: Machine Learning Reconstruction and Coupled Modelling in Upper Blue Nile Tributaries

Sediment-laden runoff in Ethiopia’s Upper Blue Nile Basin (UBNB) threatens the ecological health of Lake Tana and the operational efficiency of the Grand Ethiopian Renaissance Dam (GERD). Limited event-based sediment sampling hinders the accurate estimation of fluxes and process-based modeling in this data-scarce region. This study reconstructs continuous daily sedigraphs (1990–2020) for the Gilgel Abay (1,664 km²) and Gumara (1,394 km²) watersheds using machine-Learning (ML) methods, including Gradient Boosting (GB), Random Forest (RF), and Quantile Regression Forests (QRF), along with traditional techniques, using discharge, rainfall, temperature, and evapotranspiration as predictors.

QRF achieved the highest validation accuracy at the daily scale (R² = 0.62–0.72), capturing non‑linear sediment dynamics and providing uncertainty‑quantified yields (90% CI: 17.15–54.37 t/ha/yr for Gilgel Abay; 21.15–40.61 t/ha/yr for Gumara). Mean annual sediment yields were 27.5 ± 7.2 t/ha/yr (Gilgel Abay) and 23.8 ± 10.7 t/ha/yr (Gumara), with 93–95% of transport occurring during the monsoon season (June–October), emphasizing strong rainfall control.

The reconstructed records enabled the first successful calibration and validation of the WASA-SED model for coupled daily streamflow and suspended-sediment dynamics in the Ethiopian Highlands. Monthly simulations showed strong performance (NSE 0.66–0.86; R² 0.72–0.87). Flow- and sediment-duration curves indicated excellent skill during high-flow events, which dominate sediment export, with underestimation in mid- and low-sediment ranges.

Decadal analyses revealed contrasting watershed trajectories. In Gilgel Abay, rainfall intensified (from 136.9 mm/month in the 1990s to 208 mm/month in the 2020s), streamflow increased by 78% (55 to 98 m³/s), and sediment loads peaked mid‑period before declining. In Gumara, rainfall remained stable, but streamflow rose 54% (35 to 54 m³/s), and sediment loads increased 61% (8.2 to 13.2 × 10³ t/day), influenced by wetland loss (−63%) and rapid urban expansion.

This integrated ML–process modelling framework bridges sediment data gaps, advances hydro-sediment process understanding, and supports targeted erosion mitigation for the sustainable management of the UBNB. The approach is transferable to other humid tropical basins facing similar data limitations.

 

Keywords: sediment reconstruction, QRF, WASA‑SED, decadal trends, Upper Blue Nile, humid tropics, data‑scarce modelling