Estimating daily pan evaporation with artificial neural networks

Daily pan evaporation has been shown to be an important variable in making crop management decisions and in modeling crop response to weather conditions. However, daily pan evaporation is difficult to measure accurately and consistently over longer time periods. The objective of this research was to develop artificial neural network (ANN) models to estimate daily pan evaporation using measured weather variables as inputs. Weather data from Rome, Plains, and Watkinsville, Georgia, consisting of 2044 daily records from 1992 to 1996 were used to develop the models of daily pan evaporation. Additional weather data from these locations, which included 720 daily records from 1997 and 1998, served as an independent evaluation data set for the models. The measured variables included daily observations of rainfall, temperature, relative humidity, solar radiation, and wind speed. Daily pan evaporation was also estimated using multiple linear regression and the Priestley-Taylor method and was compared to the results of the ANN models. The ANN model of daily pan evaporation with all available variables as inputs was the most accurate model delivering an r2 of 0.717 and a root mean square error of 1.11 mm for the independent evaluation data set. ANN models were developed with some of the observed variables eliminated to correspond to different levels of data collection as well as for minimal data sets. The accuracy of the models was reduced considerably when variables were eliminated to correspond with National Weather Service cooperative weather stations. Pan evaporation estimated with ANN models was slightly more accurate than pan evaporation estimated with a multiple linear regression model or the Priestley-Taylor equation. Future efforts will focus on the inclusion of the ANN model as part of the quality control procedure to estimate missing pan evaporation data of the automated weather station network.