Infilling Missing Daily Evapotranspiration Data Using Neural Networks

This study used artificial neural networks (ANNs) computing technique for infilling missing daily saltcedar evapotranspiration (ET) as measured by the eddy-covariance method. The study site was at Bosque del Apache National Wildlife Refuge in the Middle Rio Grande Valley, New Mexico. Data was collected from 2001 to 2003. Several ANN models were evaluated for infilling of different combinations of missing data percentages and different gap sizes. The ANN model using daily maximum and minimum temperature, daily solar radiation, day of the year, and the calendar year as inputs showed the best estimation performance. Results showed coefficient of determination (R-2) of 0.96, root-mean-square error (RMSE) of 0.4 mm/day for 10% missing data and a maximum of half-month gap size data set. Missing data greater than 30% and maximum data gap size greater than 3 months resulted in R-2 less than 0.90 and RMSE greater than 0.6 mm/day. The results from this study suggest that infilling of daily saltcedar ET using ANN and readily available weather data where the ET observations exist before and after the gap is a reliable and convenient method. It could be used to obtain continuous ET data for modeling and water management practices.