Reference Evapotranspiration Modeling Using New Heuristic Methods.

Area of Science:

• Environmental science
• Agricultural engineering
• Data science

Background:

• Accurate estimation of reference evapotranspiration (ETo) is crucial for water resource management and agricultural planning.
• Limited data availability poses a challenge for traditional ETo modeling techniques.
• Machine learning offers promising alternatives for ETo prediction.

Purpose of the Study:

• To evaluate the performance of two novel machine learning methods, LSSVR-GSA and DENFIS, for modeling monthly reference evapotranspiration (ETo).
• To compare the efficacy of these new methods against the M5 model tree (M5RT) approach.
• To identify the optimal input variables for enhancing ETo prediction accuracy.

Main Methods:

• Least-square support vector regression with a gravitational search algorithm (LSSVR-GSA).
• Dynamic evolving neural-fuzzy inference system (DENFIS).
• M5 model tree (M5RT).
• Input variables included temperature and extraterrestrial radiation data from three stations in China.

Main Results:

• LSSVR-GSA models demonstrated superior performance in estimating monthly ETo compared to DENFIS and M5RT.
• Models utilizing only extraterrestrial radiation data showed improved prediction accuracy across all tested methods.
• Incorporating periodicity information or combining air temperature with extraterrestrial radiation did not consistently enhance prediction accuracy.

Conclusions:

• LSSVR-GSA is a highly effective method for modeling reference evapotranspiration, particularly with limited data.
• Extraterrestrial radiation is a key predictor for accurate monthly ETo estimation.
• Further research may explore hybrid models or advanced feature engineering for even greater accuracy.