Performance evaluation of machine learning algorithms for estimating reference evapotranspiration based on NASA POWER weather data: a case study in Nigeria.

The Penman-Monteith (PM) method is recognized as the globally accepted approach for estimating reference evapotranspiration (ETo). However, its use is constrained in areas with limited or unavailable data. Predicting ETo using multiple support vector machine (SVM) kernels and decision tree (DT) ensembles with NASA POWER data is innovative, as previous SVM-based ETo prediction studies have relied primarily on linear kernels. This study aims to evaluate the performance of different machine learning (ML) models, specifically SVM and DT and their ensembles, using NASA Power data as input. For this purpose, ML models were trained using average values of the monthly climatic data (maximum and minimum air temperatures, relative humidity, and wind speed) from NASA POWER. ETo was used as the output variable and was calculated from ground-observed data using the PM method. The developed ML models underwent training and validation to determine ETo in areas with different weather conditions in Nigeria: Kano-dry weather, Onne-wet weather, and Ibadan-moderate weather. Thirty and 70 % of the data were used during training and validation, respectively. The SVMs used in this study include linear SVM, quadratic SVM, cubic SVM, fine Gaussian (FG) SVM, medium Gaussian (MG) SVM, and coarse Gaussian SVM. The decision trees include fine, medium, and coarse trees, along with their ensembles: bagged and boosted trees. The model performance was evaluated using various error metrics. The FG SVM model exhibited the most accurate and precise estimation of ETo, with root mean square error (RMSE) values of 0.38 and 0.599 mm during the training and testing phases, respectively. Additionally, the coefficient of determination (r