Alkanjo et al. (2026) Machine Learning as a Tool to Predict Reference Evapotranspiration

Identification

• Journal: Water Resources Management
• Year: 2026
• Date: 2026-02-01
• Authors: Safa Alkanjo, Kübra Kaya, Veysi Kartal, Veysel Süleyman Yavuz, Michael Nones
• DOI: 10.1007/s11269-025-04460-8

Research Groups

• Department of Civil Engineering, Faculty of Engineering, Siirt University, Siirt 56000, Türkiye
• Institute of Geophysics Polish Academy of Sciences, Warsaw, Poland

Short Summary

This study predicted monthly reference evapotranspiration (ET) in Siirt, Türkiye, using various machine learning and statistical regression models, along with a Design of Experiments (DoE) approach. The DoE model achieved the highest accuracy (R²=0.987), identifying average temperature as the most influential variable.

Objective

• To predict monthly reference evapotranspiration (ET) in Siirt province, Türkiye, by integrating various machine learning and statistical regression models, and a Design of Experiments (DoE) approach, using climate variables.

Study Configuration

• Spatial Scale: Siirt province, Türkiye, covering an area of 5,718 square kilometers.
• Temporal Scale: Monthly predictions based on long-term hydro-meteorological observations from 1980 to 2023.

Methodology and Data

• Models used:
  • Machine Learning: Support Vector Machine (SVM), Extreme Gradient Boosting (XgBoost), K-Nearest Neighbor (KNN), Random Forest (RF).
  • Statistical Regression: Quartile Regression, Partial Least Squares (PLS) Regression, LASSO Regression, Ridge Regression, Elastic Net Regression, NonParametric Regression, Linear Regression.
  • Design of Experiments (DoE).
• Data sources:
  • Observed meteorological data from the General Directorate of Meteorology (MGM) for the period 1980-2023.
  • Reference evapotranspiration calculated using the Penman-Monteith (FAO-56 PM) method.
  • Climate variables: Minimum temperature (Tmin), average temperature (Tav), maximum temperature (Tmax), minimum relative humidity (RHmin), average relative humidity (RHave), maximum relative humidity (RHmax), precipitation (P), solar hour (SH), wind speed (WS), average snow thickness (Stav), average pressure (PRav), number of cloudy days (Cdnu), daily average cloudiness (AC), and average solar radiation intensity (SRIav).

Main Results

• The Design of Experiments (DoE) model demonstrated the highest overall accuracy with a coefficient of determination (R²) of 0.987, a Mean Squared Error (MSE) of 926.465, and a Root Mean Squared Error (RMSE) of 30.438.
• Among machine learning models, Random Forest (RF) performed best (R²=0.948, MSE=1066.37, RMSE=32.666), while XgBoost was the lowest (R²=0.776, MSE=4608.27, RMSE=67.88).
• Among statistical regression methods, NonParametric regression performed best (R²=0.948, MSE=1061.627, RMSE=32.583), and Partial Least Squares (PLS) performed weakest (R²=0.901, MSE=2233.01, RMSE=47.255).
• Average temperature (Tav) was identified as the most influential variable affecting ET, contributing 55.303% to the DoE model's effectiveness, followed by wind speed (WS) at 34.800%.
• Correlation analysis showed strong positive correlations between ET and temperature variables (Tmin, Tav, Tmax, with Tmax at 0.83), solar hour (SH), and average solar radiation intensity (SRIav). Negative correlations were observed with relative humidity variables (RHmin, RHav, RHmax, with RHav at -0.90), precipitation, average snow thickness (Stav), average pressure (PRav), number of cloudy days (Cdnu), and daily average cloudiness (AC).
• Monthly ET ranged from 0.0145 meters to 0.5413 meters (mean 0.19048 meters).
• Average solar radiation intensity (SRIav) ranged from 2.91 MJ·m⁻² to 30.17 MJ·m⁻² (mean 15.15 MJ·m⁻²).
• Average snow thickness (Stav) ranged from -0.01 meters to 0.184 meters (mean 0.0091 meters).
• Precipitation (P) ranged from 0 meters to 0.2911 meters (mean 0.05712 meters).
• Wind speed (WS) ranged from 0 meters per second to 2.4 meters per second (mean 1.24 meters per second).
• Average pressure (PRav) ranged from 902.1 hectopascals to 921.3 hectopascals (mean 912.15 hectopascals).

Contributions

• Provides a comprehensive and original framework for monthly reference evapotranspiration (ET) prediction in the semi-arid Siirt province of Türkiye, integrating a wide array of machine learning and statistical models, including a Design of Experiments (DoE) approach.
• Achieves high prediction accuracy, with the DoE model demonstrating superior performance (R²=0.987), and identifies average temperature as the most critical climatic variable influencing ET.
• Advances hydrological understanding by revealing complex interactions between ET, atmospheric pressure, cloudiness, and solar radiation, offering new insights into hydrological dynamics under changing climate conditions.
• Delivers a rigorous, high-quality predictive structure with practical implications for improved agricultural planning, water resource management, and climate adaptation strategies in regions with similar climatic conditions.

Funding

• Scientific and Technological Research Council of Türkiye (TÜBİTAK)
• TÜBİTAK 2209-A University Students Research Projects Support Programme (Project number: 1919B012320270)
• Polish Ministry of Education and Science (subsidy for the Institute of Geophysics, Polish Academy of Sciences)

Citation

@article{Alkanjo2026Machine,
  author = {Alkanjo, Safa and Kaya, Kübra and Kartal, Veysi and Yavuz, Veysel Süleyman and Nones, Michael},
  title = {Machine Learning as a Tool to Predict Reference Evapotranspiration},
  journal = {Water Resources Management},
  year = {2026},
  doi = {10.1007/s11269-025-04460-8},
  url = {https://doi.org/10.1007/s11269-025-04460-8}
}