Leveraging Artificial Intelligence and Machine-Learning Methods for Bias Correction of Reference Evapotranspiration Utilizing ERA5 Data

Abstract

This study addresses the challenge of mitigating biases in daily reference evapotranspiration (ETo) computations, utilizing the ECMWF Reanalysis version 5 reanalysis data set within coastal regions. Data from 14 weather stations across the coastal Caspian Sea basin covering the period of 2003–2023 were gathered to compute ETo as a benchmark for bias correction. Therefore, five distinct artificial intelligence models, namely, artificial neural networks, gene expression programming, adaptive network-based fuzzy inference systems, random forest (RF), and support vector machine (SVM), were used to mitigate bias in daily ETo. The results demonstrated that the RF model performed exceptionally well, achieving normalized root mean square error values below 10% and residual mean bias error values below 15%. It ranked as the top-performing model in eight out of the 14 stations during the testing phase. However, in the testing step, SVM has surpassed other methods in other stations. To provide a better perspective on the results, a map showcasing the top-performing methods based on minimizing bias for each station was presented. This research highlights the effectiveness of machine-learning methods for improving ETo estimates. It demonstrates the potential for successful implementation of the RF method to bias correction of the Food and Agriculture Organization (FAO)–Penman–Montith equation fed by ERA5 climate data in basin scale.