基于BP神经网络的土壤水力学参数预测

Prediction of Soil Hydraulic Characteristic Parameters Based on BP Neural Network

为了获取区域土壤水分和溶质运移模拟所需的土壤水力学参数,利用黄淮海平原曲周县的试验资料建立基于BP神经网络的土壤转换函数模型。本文采用土壤粒径分布、容重、有机质含量等土壤基本理化性质,来预测土壤饱和导水率Ks、饱和含水量sθ、残余含水量θr、以及van Genuchten公式参数α、n的对数形式ln(α)和ln(n),并与多元线性逐步回归方法进行比较。t检验结果表明,BP神经网络训练和预测得到的模拟值与实测值之间吻合很好,该方法具有较高的预测精度。通过对平均相对误差的比较,得出在粒径分布的基础上增加容重、有机质含量等输入项目,可以提高部分土壤水力学参数的预测精度,而有些参数的预测精度反而降低。以误差平方和为标准的比较结果表明,BP神经网络模型的预测效果总的来看要优于多元线性回归法。

In order to acquire the soil hydraulic characteristic parameters needed in the simulation of regional soil moisture and solute transport, Pedo - Transfer Functions （PIFs） models using Back - Propagation （BP） neural network were established based on the soil survey data in QuZhou County of the North China Plain. The models predicted saturated hydraulic conductivity Ks, saturated water content θs, residual water content θr, the logistic form of van Genuchten model parameters ln（α） and ln（n）. The input variables included soil texture, bulk density and soil organic matter content. The t-test result showed that the prediction value agreed with the observation data fairly well and the BP neural network model had rather high prediction precision. Based on the BP neural network model with input of soil texture, it could be concluded that addition of bulk density and organic matter content into the model input could improve the prediction precision for some soil hydraulic parameters, while in the meantime reduce the precision for other parameters by comparing the mean relative error. It indicated that BP neural network method performed better than multivariate stepwise regression models, as the sum of quadratic error of prediction by BP neural network model was less than by the regression method.