Performance Evaluation of ETo Prediction Methods: Dispersion Analysis and Accuracy Criteria Across Time Intervals
Year: 2025
Type: article
Abstract: Accurate forecasting of reference evapotranspiration (ET$_o$) is essential for sustainable water resource management and precision agriculture, yet systematic comparisons across extended forecasting horizons remain limited. This study evaluates three ET$_o$ prediction methods---Random Forest (RF), Cartesian Genetic Programming (CGP), and Convolutional Neural Network accelerated by Graphics Processing Unit (CNN-GPU)---across six time intervals ranging from 1 to 364 days, using data from two semi-arid stations in northwestern Iran. Model performance was assessed via dispersion analysis (scatter and violin plots) and four accuracy metrics (RMSE, MAE, R$^2$, SI). Results indicate that RF and CNN-GPU consistently outperform CGP, particularly at extended horizons. At the 364-day interval, CNN-GPU achieved the highest accuracy (RMSE: 0.678~mm/day, R$^2$: 0.874), while RF maintained robust and stable performance (RMSE: 0.683~mm/day, R$^2$: 0.872, SI: 0.244--0.278). CGP exhibited comparatively higher error at extended horizons (RMSE: 0.702~mm/day at 364~days) and greater prediction variability. Statistical validation via two-sample t-tests, F-tests, and ANOVA confirmed significant inter-method differences at lags of 22~days and beyond (p~$<$~0.05), with CNN-GPU frequently demonstrating superior accuracy at longer timescales. Time-series analyses further confirmed that RF and CNN-GPU effectively capture seasonal ET$_o$ patterns, whereas CGP exhibits increasing error dispersion with lag. All methods produced physically inconsistent negative lower-bound predictions, identifying a shared limitation requiring post-processing or architectural constraints in operational deployment.
Institution: University of Tabriz
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: hybrid
OpenAlex ID: W7091546859