Abstract: Evaporation is an important factor affecting thermal balance and water budget over the earth surface. A long-term observation of soil evaporation over semifixed dune was carried out with micro-lysimeters (MLS) in the high-frigid regions in the Qinghai-Tibetan Plateau of China during the period of 2006-2009, the dataset was consisted of the collected daily soil evaporation as the output and the corresponding meteorological observation data including relative air humidity, air temperature, wind speed and soil moisture content as the input. A desert soil evaporation model was developed to research soil evaporation over semi-fixed dune based on the radial basis function (RBF) neural network, and the multiple linear regression (MLR) was used to validate the model. The results show that the values calculated with RBF network output were consistent with the observed values, and the root mean squared error was 0.14 mm. Both the average absolute percent error and the root mean squared error for the RBF neural network were lower than those for the MLR model. The RBF neural network model is good for calculating desert soil evaporation other than the traditional mathematica1 evaporation model, and it is characterized by the simple development, high accuracy and strong adaptability.

Key words: neural network, soil evaporation, micro-lysimeter, Qinghai-Tibetan Plateau