Abstract

Reliable assessment of ET<inf>0</inf>is significant in smart irrigation scheduling, cropping system planning and management, and determination of agricultural water demand, especially in data-scarce regions. In this research, we analyzed the comparative performance of the ensemble learning and robust regression model for estimating monthly ET<inf>0</inf>in the Tana basin of Ethiopia, using 12 years (2010-2022) of monthly data from four agroecological zones and weather station data. We employed four machine learning methods, based on the Extra Trees (ET), Gradient Boosting Regressor (GBR), Bayesian Ridge Regression (BRR), and Huber Regression (HR) algorithms, and four different input configurations that depend on the available information. The ensemble models (ET and GBR) did comparably better than linear ones, at high point accuracy (NSE > 0.95) with fewer predictors. The GBR model had shown adequate generalization capability across stations and scenarios. We evaluated these predictions against the WaPOR and ERA5-Land ET<inf>0</inf>in situ-based data and found that ML-based models are promising alternatives in regions with data deficiency. The research highlights the importance of AI-enabled technologies and open-source datasets as smart agricultural tools for sustainable irrigation scheduling over water resources management under climate uncertainty.