Spatiotemporal drought dynamics in the Southeastern Anatolia, Türkiye: a CEEMDAN-LSTM hybrid deep learning approach with SHAP-based explainability and uncertainty quantification

Abstract

Drought has become one of the most critical hydroclimatic hazards under climate change with increasing temperature, enhanced atmospheric evaporative demand and increasing hydroclimatic variability. This research explores the spatiotemporal behavior and predictability of meteorological drought in the Southeastern Anatolia Region by means of the Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) at 3-, 6-, and 12-month timescales for the period 1965–2025. The variability of drought, long-term trends, spatial patterns and forecasting performance were analyzed using the monthly precipitation and temperature data of 15 meteorological stations. A hybrid Complete Ensemble Empirical Mode Decomposition with Adaptive Noise–Long Short-Term Memory (CEEMDAN-LSTM) framework was applied and compared with Persistence, Random Forest, XGBoost, and standalone LSTM models. Model interpretability was evaluated with SHapley Additive exPlanations (SHAP), and forecast uncertainty was measured in terms of quantile regression-based prediction intervals. The results reveal a significant regional drying trend after 2000, with the regional mean SPEI-12 shifting from near-normal conditions to persistent negative values. The drought episode of 2020–2025 is the most extreme and longest drought episode in the 60-year observational record. Compared with the SPI, the SPEI exhibited stronger and more spatially coherent drying trends, which revealed the amplification role of temperature-driven evapotranspiration under warming conditions. CEEMDAN decomposition successfully identified dominant interannual and multi-decadal oscillatory modes, while residual components showed a persistent long-term drying trend. The CEEMDAN-LSTM model displayed the best predictive performance for most index-timescale combinations across all forecasting approaches, with R² values of 0.990 for SPEI-6 and 0.981 for SPI-12. SHAP analysis revealed that seasonal evapotranspiration variability strongly controls SPEI predictability, while SPI forecasts are mainly driven by lagged precipitation persistence. Uncertainty analysis further revealed decreased prediction interval reliability during the unprecedented 2022–2025 drought period, highlighting the effects of climatic no stationarity and the limitations of historical training data under intensifying drought conditions. The results, generally, demonstrate the high potential of hybrid deep learning frameworks for drought forecasting and offer valuable scientific support for climate adaptation and sustainable water resource management in the Euphrates–Tigris Basin.