Artificial neural networks versus gene expression programming for estimating reference evapotranspiration in arid climate

• Eight GEP models have been compared with eight ANN models to estimate reference evapotranspiration.
• Climatic data were collected from 19 Metrologic stations covering the period from 1980 to 2010 for Saudi Arabia.
• The results showed that the GEP models performed slightly worse than the ANN models, the GEP models used explicit equations.
• Wind speed and plant height have a great positive impact in increasing the accuracy of calculating ETref.

Artificial neural networks (ANNs) and gene expression programming (GEP) were compared to estimate daily reference evapotranspiration (ETref) under arid conditions. The daily climatic variables were collected by 13 meteorological stations from 1980 to 2010. The ANN and GEP models were trained on 65% of the climatic data and tested using the remaining 35%. The generalised Penman–Monteith (PMG) model was used as a reference target for evapotranspiration values, with hc varies from 5 to 105 cm with increment of a centimetre. The developed models were spatially validated using climatic data from 1980 to 2010 taken from another six meteorological stations. The results showed that the eight ETref models developed using the ANN technique were slightly more accurate than those developed using the GEP technique. The ANN models’ determination coefficients (R2) ranged from 67.6% to 99.8% and root mean square error (RMSE) values ranged from 0.20 to 2.95 mm d-1. The GEP models’ R2 values ranged from 64.4% to 95.5% and RMSE values ranged from 1.13 to 3.1 mm d-1. Although the GEP models performed slightly worse than the ANN models, the GEP models used explicit equations.