A Machine Learning-driven multi-sensor framework for detecting downstream hydrological impacts of the Grand Ethiopian Renaissance Dam

Abstract The Grand Ethiopian Renaissance Dam (GERD), Africa's largest hydroelectric facility, has been filling the Blue Nile since 2020 amid an unresolved dispute among Ethiopia, Sudan, and Egypt. Quantifying its downstream footprint against natural climate variability has remained methodologically unresolved without independent basin-scale evidence. We present a machine learning framework for multi-sensor satellite detection and attribution of GERD operational impacts, integrating Sentinel-2 derived surface water volumes with terrestrial water storage, precipitation, soil moisture, and climate teleconnection indices across the Blue Nile basin at 5 km resolution (271,887 cells) over 2015–2025. An Isolation Forest and Local Outlier Factor ensemble trained on pre-GERD residuals flags post-impoundment anomalies. Five independent causal attribution tests are applied: paired-difference permutation testing, difference-in-differences using the unregulated Dinder River as a climate control, filling-phase monotonicity, ENSO phase stratification, and a gradient-boosted distribution-shift diagnostic. Basin-wide post-GERD anomaly rates rose from 1.47% to 9.63% (+ 8.16 pp), peaking at the GERD nearfield (+ 13.61 pp) and remaining detectable 700 km downstream at Khartoum (+ 8.43 pp). Anomaly rates increased monotonically through the first three filling phases (1.18% → 7.33% → 9.35% → 11.95%). Climate alternatives are ruled out: ENSO phase produced only a 0.57 pp range, and pre-GERD predictive skill collapsed from R² = 0.683 to 0.391 (Highland Floodplain: 0.877 → 0.263). The framework is reproducible from open-access remote sensing and transferable to other contested transboundary systems.