Water scarcity is becoming increasingly prevalent in semi-arid regions under climate change; however, the relationship between meteorological and hydrological drought remains poorly understood in highly regulated river systems. This study develops a hybrid machine learning (ML) framework to predict hydrological drought in the Al-Amara reach of the Tigris River Basin, Iraq. The framework integrates an LSTM-XGBoost ensemble for streamflow prediction, an XGBoost-Random Forest-Gradient Boosting ensemble for direct Streamflow Drought Index (SDI) prediction, and Shannon entropy-weighted fusion of SPI-3 and SPI-12 to derive the Composite Drought Index (CDI). Using 64 years of precipitation and streamflow data (1961–2024) and 39 input variables, the hybrid ensembles significantly outperform individual ML models, achieving R² = 0.920 and RMSE = 7.61 m³/s for streamflow prediction and R² = 0.9198 and RMSE = 0.3495 for SDI prediction. Historical analysis reveals very weak coupling between meteorological and hydrological drought (r = 0.093–0.145), potentially attributable to groundwater buffering, snowmelt dependency, and dam regulation. Projections for 2021–2100 under four IPCC AR6 SSP scenarios indicate weaker decoupling at higher emissions, with a potential tipping point between SSP2-4.5 and SSP3-7.0. Moderate drought frequency is projected to rise approximately 3.9-fold, while late-century (2081–2100) streamflow under SSP5-8.5 is expected to decline by approximately 57% relative to the historical mean in the study area.
Copyrights © 2026