Climatic Drivers of Water Availability in the Atatürk Dam Basin: An Explainable Machine-Learning (XGBoost–SHAP) Framework Using ERA5-Land (1950–2025)

Abstract

Water availability in the Atatürk Dam basin, the largest reservoir of the Southeastern Anatolia Project (GAP) and a snow-fed Euphrates headwater, is sensitive to warming, yet its climatic drivers have not been ranked. Existing studies use trend tests, process models, or black-box machine learning. This study attributes and forecasts total runoff (a water-availability proxy) from basin-mean monthly ERA5-Land data (1950–2025), using XGBoost against linear and random-forest baselines under leakage-free cross-validation and exact TreeSHAP. The model reproduced the reanalysis runoff (R² ≈ 0.97, NSE ≈ 0.96), and SHAP ranked three-month accumulated snowmelt, deep-layer soil moisture, and water-year cumulative precipitation as the dominant drivers, with nonlinear thresholds; permutation and water-balance tests confirmed it is leakage-free. Antecedent snow-depth memory weakened over time while the snowmelt share was preserved, consistent with significant warming, snow decline, and rising evaporative demand (Mann–Kendall, Sen's slope); annual runoff showed no trend. From antecedent state, runoff was forecast about three months ahead (one-month R² ≈ 0.89). The framework can inform climate-resilient GAP reservoir operation; as it is reanalysis-based and unvalidated against gauges, future work should add gauge validation and gridded, scenario-based settings.