Akiner et al. (2026) Is runoff the key input of evapotranspiration?: AI-based hydro-climatic assessment in Southeastern Türkiye’s Dams Region

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-03-21
• Authors: Muhammed Ernur Akiner, Mehmet Ali Çelik
• DOI: 10.1007/s00704-026-06137-7

Research Groups

• Department of Environmental Protection Technologies, Vocational School of Technical Sciences, Akdeniz University, Antalya, Türkiye
• Geography Department, Faculty of Arts and Sciences, Igdir University, Igdir, 76000, Türkiye
• Department of Geography, Nakhchivan State University, AZ, 7012, Nakhchivan, Azerbaijan

Short Summary

This study assesses hydro-climatic variability and evapotranspiration dynamics in Southeastern Türkiye's Dams Region using AI models and long-term reanalysis data. It identifies runoff as the most critical, non-linearly influential factor in evapotranspiration, highlighting its importance for water resource management in semi-arid, dam-regulated environments.

Objective

• To evaluate the hydro-climatic variability and evapotranspiration (ET₀) dynamics in Southeastern Türkiye's Dams Region, specifically investigating the influence of runoff as a key input.
• To develop and compare various computational intelligence models, including a novel Hybrid RF–ANN–Bi-LSTM architecture, for predicting daily ET₀.
• To apply SHAP-based sensitivity analysis to interpret the nonlinear and threshold-like effects of hydro-climatic drivers on ET₀, enhancing the physical interpretability of AI models in hydrology.

Study Configuration

• Spatial Scale: Southeastern Türkiye’s Dams Region, encompassing the Keban, Karakaya, and Atatürk reservoirs, and six provinces (Malatya, Adıyaman, Şanlıurfa, Diyarbakır, Elazığ, and Tunceli). The ERA5-Land data has an approximate 9 km grid spacing.
• Temporal Scale: Daily data from 1960 to 2023, spanning 63 years.

Methodology and Data

• Models used: Multiple Linear Regression (MLR), Random Forest (RF), Support Vector Regression (SVR), Artificial Neural Networks (ANN, specifically a Multilayer Perceptron (MLP) architecture), Bidirectional Long Short-Term Memory (Bi-LSTM), and a Triple Hybrid (RF–ANN–Bi-LSTM) approach. SHAP (SHapley Additive exPlanations) was used for sensitivity analysis, and K-means clustering for grouping provinces.
• Data sources: ERA5-Land reanalysis dataset from the European Centre for Medium-Range Weather Forecasts (ECMWF). Variables included: maximum, minimum, and mean air temperature (°C), dew point temperature (°C), precipitation (mm), snowfall (mm), U- and V-components of wind (m/s), surface pressure (hPa), runoff (mm), and evaporation (mm). Daily reference evapotranspiration (ET₀) in mm, calculated using the FAO Penman–Monteith equation, served as the dependent variable.

Main Results

• Runoff was consistently identified as the most important meteorological factor influencing ET₀ dynamics across all models and regions, often surpassing traditional drivers like temperature and precipitation.
• The Hybrid RF–ANN–Bi-LSTM model demonstrated the highest predictive accuracy (e.g., R² = 0.9133 in Diyarbakır) across all provinces, outperforming individual machine learning and deep learning models.
• Spatial heterogeneity in model performance was observed and systematically grouped into three clusters based on predictive skill and runoff regime characteristics. Provinces with stable, high runoff (e.g., Tunceli, Diyarbakır, Elazığ) showed higher model accuracy, while those with water-limited, unstable runoff (e.g., Şanlıurfa) exhibited lower accuracy.
• SHAP analysis revealed a strong, nonlinear, and threshold-like influence of runoff on ET₀ predictions, acting as a major limiting factor under low-flow conditions and high temperatures.
• Temperature variables (Tmax, Tmean) also exhibited nonlinear, threshold, and plateau effects on ET₀, with warming having a more pronounced impact during dry periods.
• While all models showed a significant decrease in predictive ability for extreme high-ET₀ events (Q90), the Hybrid model consistently recorded the lowest error metrics (MAE and RMSE) under these challenging conditions.

Contributions

• Introduces a novel approach by explicitly integrating runoff as a primary predictor for daily reference evapotranspiration (ET₀) within a multi-model AI ensemble, including a new Hybrid RF–ANN–Bi-LSTM architecture.
• Utilizes SHAP-based sensitivity analysis to provide physically interpretable insights into the nonlinear, threshold, and interaction effects of hydro-climatic drivers on ET₀, addressing the need for explainable AI in hydrology.
• Offers a comprehensive hydro-climatic assessment of a water-scarce, dam-regulated region in Southeastern Türkiye, linking water availability to atmospheric demand and highlighting regional vulnerabilities.
• Provides a robust and scalable methodology for improving hydro-climatic modeling accuracy and decision support systems for water resource management and climate-risk mitigation in semi-arid transboundary river basins globally.

Funding

The authors received no financial support from public, commercial, or not-for-profit funding agencies to conduct this research.

Citation

@article{Akiner2026Is,
  author = {Akiner, Muhammed Ernur and Çelik, Mehmet Ali},
  title = {Is runoff the key input of evapotranspiration?: AI-based hydro-climatic assessment in Southeastern Türkiye’s Dams Region},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06137-7},
  url = {https://doi.org/10.1007/s00704-026-06137-7}
}