Abstract:
Soil moisture constants such as saturated water content and field water capacity are important hydraulic parameters in soil science. However, they are frequently lacking in routine soil surveys because direct measurements are time-consuming and labor intensive. As such, it was necessary to find indirect, cheap and rapid ways of acquiring soil moisture constants from other more easily available soil physical and chemical properties. In this study, pedotransfer functions were established by using the methods of regression, BP neural network and Rosetta model to predict saturated soil water content, capillary water-holding capacity and field water capacity from soil organic matter, texture and bulk density in the Three-River Headwaters Region of Qinghai Province. The general performance of pedotransfer functions was evaluated based on mean error, root mean squared error and determination coefficient between observed and predicted values. Vital results obtained were grouped mainly into three aspects as follows: (1) Mean error and root mean squared error of pedotransfer functions derived from regression method for the three soil moisture constants were small with big determination coefficients, showing good prediction. Specifically, mean error and root mean squared error of field water capacity derived from regression method were below 3.397%, with determination coefficient above 0.868. (2) Mean error and root mean squared error of pedotransfer functions for the three soil moisture constants obtained by BP neural network method were below 4.685% with determination coefficients above 0.857. Moreover, observed and predicted values of the three soil moisture constants were close to the 1︰1 line. It therefore generally suggested that the estimation of the three soil moisture constants by the BP neural network method was good. (3) As for pedotransfer functions derived from Rosetta model, mean error and root mean squared error of the three soil moisture constants were larger and with lower determination coefficients than the other two methods. This showed that Rosetta model performed relatively poor in estimating soil moisture constants in the study area. In general, pedotransfer functions of saturated soil water content and capillary water-holding capacity established by BP neural network method were the best. Furthermore, regression method was best for estimating field water capacity and Rosetta model was worse than the other two methods. This research not only provided the scientific basis for studying soil hydraulic parameters in the Three-River Headwaters Region in Qinghai Province, but was also critical for estimating soil water at region scales.