Abstract

Accurate estimation of reference evapotranspiration (ET0) remains essential for successful water resource management and agricultural development. This research compares four empirical prediction models, which include Hargreaves-Samani (HS), Modified HS, Ritchie and Makkink against three machine learning methods namely Support Vector Machines, Boosted Regression Trees, and Gaussian Process Regression (GPR), for the prediction of ET0 in Abuja, Nigeria. Meteorological data for 34 years from 1984 – 2022 including monthly minimum and maximum air temperatures, mean temperature, relative humidity, monthly average wind speed, monthly precipitation, surface pressure, specific humidity, and solar radiation, were used as inputs. The FAO-56 Penman-Monteith method, approved as the standard method for estimating ET0 was employed as a benchmark for comparison purposes. The results showed that for empirical models, HS led to better performance with root mean square error (RMSE) value of 0.1869. However, the machine learning models outperformed the empirical models across all modelling scenarios and generally, GPR produced the most superior performance with a determination coefficient (R²) value of 0.9979 in the validation phase. Research outcomes demonstrated that GPR and other applied machine learning models promise to predict ET0 effectively in Abuja, leading to better water resource management and agricultural decision systems

Key words: Meteorological Data, Machine Learning, Reference Evapotranspiration, FAO-56 Penman-Monteith, Abuja.