Comparison of three methods to develop pedotransfer functions for the saturated water content and field water capacity in permafrost region

In this study, pedotransfer functions (PTFs) for predicting the soil saturated water content (SWC) and field water capacity (FWC) from basic soil properties were developed by using multiple-linear regression (MLR), artificial neural network (ANN) and Rosetta method. A soil data set (N = 488 samples) in the Three-River Headwaters Region (Qinghai Province in China), was randomly divided into a training data set (N1 = 400 samples) for the prediction, and a testing data set (N2 = 88 samples) for the validation. The general performance of PTFs was evaluated based on the coefficient of determination (R2), root mean square error (RMSE) and mean error (ME) between the observed and predicted values. Some important conclusions were obtained from this research, which mainly contained three aspects as follows. (1) The general prediction effect of the MLR method was good. The absolute value of ME and RMSE for the SWC was below 0.0509, and the R2 was 0.9031. However, the absolute value of ME and RMSE for the FWC were bigger, and the R2 was lower than the ANN and Rosetta method respectively. (2) The performance of ANN was the best in three methods. The absolute value of ME and RMSE for the SWC and FWC was all below the 0.0386, and their R2 were above 0.8593. (3) The absolute value of ME and RMSE of the Rosetta method for the SWC were larger than other two methods, and the R2 was lower than the ANN but higher than MLR. The prediction effect for the FWC was fairly good for its relatively high R2 and low ME, RMSE. This research could provide the scientific basis for the study of soil hydraulic properties in the Three-River Headwaters Region of Qinghai Province and be helpful for the estimation of soil water retention in regional scale.