基于采样密度不同土壤的水分特征参数的预测及误差分析

Prediction and Error Analysis of Moisture Characterisitvs Parameters of Soil at Different Sampling Density

目前土壤水分特征曲线（h-θ曲线）的研究普遍采用vanGenuchten模型（简称VG模型）。利用不同取样精度的土壤，将土壤质地（砂土、粘粒、粉粒含量）和容重作为输入值，探讨了使用基于土壤转换函数的BP神经网络模型来预测0～20cm表层土壤水分特征曲线参数，用甘肃称钩河流域小流域的土样进行了预测和误差分析。结果表明，使用线性回归能够减小预测误差与实测值差距；BP神经网络预测饱和体积含水量的准确性比预测剩余体积含水量和田间持水量要高。为了进一步提高预测精度，还应尽可能地包括土壤结构、有机质含量等信息。

Generally, van Genuchten model （VG model） is used to study soil water retention curve （h-θ curve）. Soil was sampled at different precision, and soil texture （sand, silt, clay content） and bulk density were used as inputs to predict soil water retention curve parameters of depths （0- 20 cm） in top-soil by bagging artificial neural networks, which is based on pedotransfer functions. Furthermore, soil samples collected from watershed of Chenggou in Gansu Province were used to predict soil water retention curve parameters and to analysis on their bi- as. The results indicated that linear regression could be used to reduce the bias of prediction parameters and tested parameters; the accuracy of predicting saturated volumetric capacity was better than the accuracy of predicting soil available water capacity and field moisture capacity by BP artificial neural networks.