Yetik (2025) Machine learning-based estimation of daily ETo under limited meteorological data

Identification

• Journal: Journal of Agricultural Faculty of Gaziosmanpasa University
• Year: 2025
• Date: 2025-12-30
• Authors: Ali Kaan Yetik
• DOI: 10.55507/gopzfd.1709027

Research Groups

• Niğde Ömer Halisdemir University, Faculty of Agricultural Science and Technology, Department of Biosystems Engineering, Niğde, Türkiye.

Short Summary

This study evaluated the performance of three machine learning models (ANN, LGBM, RFR) for estimating daily reference crop evapotranspiration (ETo) in Alanya, Turkey, under various limited meteorological data scenarios, finding that ANN and LGBM consistently outperformed RFR, with the best accuracy (R²=0.89) achieved using temperature, sunshine duration, and wind speed.

Objective

• To evaluate the performance of Artificial Neural Networks (ANN), Light Gradient Boosting Machines (LGBM), and Random Forest Regression (RFR) models for estimating daily reference crop evapotranspiration (ETo) in Alanya, Turkey, under various scenarios of limited meteorological data availability.
• To identify reliable machine learning-based modeling frameworks for precise irrigation planning and sustainable agricultural water management in data-limited Mediterranean climates.

Study Configuration

• Spatial Scale: Alanya, Turkey (36.55° N, 31.90° E), a coastal district in the southern Mediterranean region.
• Temporal Scale: Daily observations over 49 years (1975–2023).

Methodology and Data

• Models used: Artificial Neural Networks (ANN), Light Gradient Boosting Machines (LGBM), Random Forest Regression (RFR).
• Data sources:
  • Daily meteorological observations (average air temperature in degrees Celsius, sunshine duration in seconds, relative humidity in percent, and wind speed at 2 meters height in meters per second) from 1975 to 2023, obtained from the Turkish State Meteorological Service (TSMS).
  • Reference ETo (target variable) calculated using the FAO ETo Calculator (version 3.1), which implements the FAO-56 Penman–Monteith algorithm.
  • Ten input scenarios with different combinations of temperature (T), sunshine duration (S), relative humidity (H), and wind speed (W).
  • Data split: 70% (1975–2010) for training, 30% (2011–2023) for testing.

Main Results

• ANN and LGBM consistently outperformed RFR across most scenarios.
• Daily ETo values fluctuated notably between a minimum of approximately 5.79 x 10⁻⁹ meters per second and a maximum of around 8.56 x 10⁻⁸ meters per second.
• Single-variable scenarios:
  • Temperature (T) alone was the most reliable single predictor (R²=0.66, RMSE range: 8.68 x 10⁻⁹ to 1.08 x 10⁻⁸ meters per second; MAE: 8.68 x 10⁻⁹ meters per second).
  • Sunshine duration (S) alone (R²=0.52, RMSE range: 1.04 x 10⁻⁸ to 1.29 x 10⁻⁸ meters per second; MAE: 1.04 x 10⁻⁸ to 1.07 x 10⁻⁸ meters per second).
  • Relative humidity (RH) and wind speed (WS) alone showed poor predictive accuracy (R² < 0.06, RMSE > 1.57 x 10⁻⁸ meters per second; MAE: 1.57 x 10⁻⁸ to 1.62 x 10⁻⁸ meters per second).
• Two-variable scenarios:
  • Temperature + Sunshine duration (T+S) achieved the highest accuracy among two-variable scenarios (R²=0.85, RMSE between 5.67 x 10⁻⁹ and 5.90 x 10⁻⁹ meters per second; MAE: 5.67 x 10⁻⁹ to 5.90 x 10⁻⁹ meters per second).
• Three-variable scenarios:
  • Temperature + Sunshine duration + Wind speed (T+S+WS) yielded the best overall performance (R²=0.89, RMSE ranging from 5.21 x 10⁻⁹ to 6.25 x 10⁻⁹ meters per second; MAE: 5.09 x 10⁻⁹ to 5.21 x 10⁻⁹ meters per second).
  • Temperature + Sunshine duration + Relative humidity (T+S+RH) also performed well (R² between 0.86 and 0.87, RMSE between 6.94 x 10⁻⁹ and 8.10 x 10⁻⁹ meters per second; MAE: 5.44 x 10⁻⁹ to 5.67 x 10⁻⁹ meters per second).
• ANN showed slightly superior performance in complex input scenarios (e.g., RMSE of 5.21 x 10⁻⁹ meters per second in Scenario 9, nRMSE as low as 16.77%), with LGBM providing comparable accuracy.

Contributions

• Demonstrates the potential of machine learning models (especially ANN and LGBM) for accurate ETo estimation with limited meteorological data.
• Provides a systematic evaluation of different input variable combinations for ETo modeling in a Mediterranean climate.
• Offers reliable, machine learning-based modeling frameworks to support precise irrigation planning and sustainable water management in data-scarce regions.
• Highlights the dominant role of temperature and sunshine duration, with a beneficial contribution from wind speed, in ETo estimation in Mediterranean climates.

Funding

• Not specified. The author acknowledged the Turkish State Meteorological Service for providing meteorological data.

Citation

@article{Yetik2025Machine,
  author = {Yetik, Ali Kaan},
  title = {Machine learning-based estimation of daily ETo under limited meteorological data},
  journal = {Journal of Agricultural Faculty of Gaziosmanpasa University},
  year = {2025},
  doi = {10.55507/gopzfd.1709027},
  url = {https://doi.org/10.55507/gopzfd.1709027}
}