Review on Machine Learning Techniques for Prediction of Reference Evapotranspiration

Abstract

Reference evapotranspiration (ET0) is a fundamental parameter in hydrological modeling, irrigation scheduling, and sustainable water resource management (WRM). Traditional physically based models often face challenges of non-linearity and data scarcity, making Soft Computing (SC) techniques increasingly valuable. This systematic review critically examines the architectural design, operational mechanisms, and comparative predictive performance of three leading SC approaches: Artificial Neural Networks (ANNs), Adaptive Neuro-Fuzzy Inference Systems (ANFIS), and Wavelet-coupled ANN (WANN) models for ET0 prediction. ANNs function as powerful nonlinear function approximators, utilizing training algorithms (such as the Levenberg-Marquardt backpropagation and methods incorporating momentum) to optimize internal parameters and minimize the Mean Squared Error (MSE). ANFIS combines the adaptive learning capability of neural networks with the interpretability of fuzzy logic, making it well-suited for systems handling uncertain data. However, empirical insight across diverse climatic conditions indicates the superior performance of hybrid models. Wavelet Transform (WT) enhances ET0 prediction by decomposing complex, non-stationary time series, enabling WANN models to capture both large-scale and fine-scale dynamic. These hybrid models demonstrate significantly improved predictive accuracy, with performance metrics reaching R=0.96 and RMSE=0.632 mm/day in arid regions, surpassing ANN and ANFIS models. Sensitivity analyses further underscore the importance of data quality, identifying solar radiation (Rs) as the most influential meteorological input.