Abstract:Pedotransfer function was established to predict soil water retention curve by using the feed-forward neural networks methods. The prediction performance and application uncertainty of pedotransfer function were analyzed according to the error statistics index and Hydrus-1D water dynamics model. The results showed that mean absolute error value of θ1000, θ10000, θ15000(soil water retention θ1000, θ10000, θ15000 at soil water suction equal to 1000cm, 10000cm and 15000cm, respectively) using pedotransfer functions with particle size distribution, bulk density, θ60 as predictor was 42.86%,23.87% and 26.15% lower than the value of PTF1 using particle size distribution as predictor. Mean absolute error of θ100,θ10000,θ15000 using pedotransfer function which added θ15000 as predictor was 8.67%,16.96% and 15.95% lower than pedotransfer function using θ60(soil water θ60 at soil water suction equal to 60cm) as predictor. The parameters of van Genuchten equation were predicted using pedotransfer function were used to simulate soil water, the MAE value of PTF using θ60 as predictor was 11.11% lower than PTF using particle size distribution as predictor. Therefore, adding additional θ15000 cannot reduce the uncertainty of pedotransfer function application comparing with the pedotransfer function using particle size, bulk density θ60 as predictor.