Long-Term Meteorology Data Analysis and Machine Learning-Driven Evapotranspiration Prediction for Optimized Irrigation Management | ||||
| Egyptian Journal of Soil Science | ||||
| Article 5, Volume 65, Issue 3, September 2025, Page 1313-1330 PDF (2.48 MB) | ||||
| Document Type: Review papers | ||||
| DOI: 10.21608/ejss.2025.378816.2137 | ||||
 View on SCiNiTO | ||||
| Authors | ||||
Yasser E. Arafa  1; ali A. Abd El-Aziz2; Abed Alataway3
| ||||
| 1Agricultural Engineerin gricultural Engineering Department, Faculty of Agriculture, Ain Shams University, Shoubra El-Kheima, Cairo, Egypt | ||||
| 2Soil Chemistry and Physics Department, Desert Research Center, 1 Mathaf El-Matareya Street, Cairo, Egypt | ||||
| 3Prince Soltan Institute for Environmental,Water and Desert Research, King Saud University, KSA | ||||
| Abstract | ||||
| Climate-related pressure on irrigation water demands is increasing worldwide, particularly for arid and semi-arid climates. Developing appropriate tools and approaches to efficiently estimate reference evapotranspiration (ETo) is critical. In this research paper, we investigated the potential of machine learning (ML) algorithms using long-term historical climatological data to estimate ETo, which could ultimately allow for improved irrigation water management in situations with limited information, such as those involving limited access to meteorological data. The empirical equations representing each ETo model included FAO-Penman-Monteith (PM), Hargreaves (HA), and Blaney-Criddle (BC), and the machine learning algorithms included six ML models: Linear Regression, K-Nearest Neighbors, Support Vector Regression, Decision Tree, Random Forest, and XGboost. Each model was assessed using historical meteorological datasets from local airports. The statistical evaluation methods chosen to assess each model were coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). Our statistical analysis and discussion of results indicated a general trend concerning increasing temperature and solar radiation reference values, with a significant increase of ETo indicated by the PM and BC methods, while HA showed a marginal decrease in reference values. All statistical algorithms provided a high predictive capability for ETo predictions, whereas the consistently performing statistical method across all comparisons was the Random Forest (RF) model. For FAO-PM, RF produced R² = 0.96, RMSE = 0.44 mm/day, and MAE = 0.32 mm/day for training, with R² = 0.94, RMSE = 0.51 mm/day, and MAE = 0.30 mm/day for testing. For HA, RF performance remained comparable. In testing, the BC equation achieved the highest accuracy, with RF scoring R² = 0.98 and RMSE = 0.12 mm/day. SVR did similarly to RF across all three equations, underscoring its strong performance. Combined with the BC equation, the RF or SVR performance represented the highest performance for predicting ETo in all experiments, especially with limited data available and in changing climates. Overall, these results support the reliability and consistency of ML algorithms to support informed, data-based irrigation decisions. | ||||
| Keywords | ||||
| Learning machine; Random forest; Reference evapotranspiration; K-nearest neighbor; Support vector regression | ||||
|
Statistics Article View: 311 PDF Download: 37 |
||||
