A comparative study on the estimation of evapotranspiration using backpropagation neural network: Penman–Monteith method versus pan evaporation method

The study compares the prediction performances of evapotranspiration by the FAO56 Penman–Monteith method and the pan evaporation method using the artificial neural network. A backpropagation neural network was adopted to determine the relationship between meteorological factors and evapotranspiration or evaporation. The evapotranspiration in the ChiaNan irrigated area of Tainan was considered. Weather data compiled by Irrigation Experiment Station of ChiaNan Irrigation Association were the input layer variables, including (1) the highest temperature, (2) the lowest temperature, (3) the average temperature, (4) the relative humidity, (5) the wind speed, (6) hours of sunlight, (7) amount of solar radiation, (8) the dew point, (9) morning ground temperature and (10) afternoon ground temperature. The importance of the ten weather factors was ranked by the general influence (GI) factor. Results show that the correlation coefficient between the evapotranspiration in 2004 calculated by FAO56 Penman–Monteith method and the one predicted by the neural network model with a hidden layer of ten nodes is 0.993. The actual evapotranspiration is 911.6 cm, and value prediction by the neural network is 896.4 cm, between which two values the error is 1.67%. The results reveal that the backpropagation neural network based on the FAO56 Penman–Monteith method can accurately predict evapotranspiration. However, the correlation coefficient between the actual evaporation in 2004 and the value prediction by the neural network with a hidden layer of ten nodes and an output layer with the pan evaporation as its target output is 0.708. The pan evaporation is 1,673.1 cm, while the value predicted by the backpropagation neural network is 1,451.7 cm, between which values the error is 13.23%. The backpropagation neural networks with pan evaporation as target outputs predict the evaporation with large errors. Moreover, the use of four agricultural weather factors (determined by the GI) including wind speed, average temperature, dew point and maximum temperature as input variables, and a hidden layer of three nodes in the backpropagation neural network model can successfully predict evapotranspiration based on the FAO56 Penman–Monteith method (R = 0.98, error = 1.35%).