基于网格搜索和交叉验证支持向量机的地表土壤容重预测

Support Vector Machines for Surface Soil Density Prediction based on Grid Search and Cross Validation

为了改善支持向量机(SVM)对地表(0~2 cm)土壤容重预测的应用可行性及其效果,针对传统经验法选择SVM惩罚因子C和核函数参数g可能造成的较大误差的问题,提出了一种对SVM参数进行优化的方法—网格搜索与交叉验证相结合的方法。本文采用所提出的参数优化方法,以黄土高原区旱作农田土壤表层容重年度跟踪试验数据样本为依据,选取土壤(0~2 cm)粒径分布、有机质含量、体积含水率、累积接受水量和全盐量为输入变量,建立了地表土壤容重的SVM预测模型。结果表明:所建SVM模型预测值和试验实测值之间不存在显著性差异,利用SVM预测地表土壤容重是可行的;采用网格搜索与交叉验证相结合的方法对SVM参数进行优化,明显降低了模型的预测误差;在粒径分布、有机质含量、体积含水率为输入变量的基础上,增加全盐量为输入因子并不能显著提高模型的预测效果,而增加累积接受水量为输入因子的预测效果明显优于前几种情况,其训练样本和测试样本相对误差的平均值分别为6.23%和6.95%,都在可接受范围。研究成果可为土壤表层容重的实时预测提供有力支撑。

The traditional empirical method may cause large error when choosing SVM error penalty factor C and Kernel parameter g.In order to improve the application feasibility and the effect of support vector machine(SVM)on the surface soil(0~2 cm)bulk density forecast,a method combined grid search and cross validation was put forward to optimize SVM parameters.Based on the annual test data of surface soil(0~2 cm)bulk density in dry farmland of Loess Plateau,the contents of soil clay and silt,organic matter,volumetric moisture content,accumulated receiving water and total salt content were established as input variables of SVM prediction model for surface soil bulk density.The results showed that:There was no significant difference between measured values and the predicated values of the SVM model,which indicated that using SVM to predict soil bulk density was feasible;the SVM parameter optimization method of grid search and cross validation significantly reduced the prediction error of the model.Based on the input variables including particle size distribution,organic matter content,volumetric water content,the added input factor of total salt content did not significantly improve the prediction effect of the model,and the prediction effect was much better after increasing accumulated receiving water as input factor,the average value of relative error of the training samples and test samples was,respectively,6.23%and 6.95%,which were in the acceptable range.The research results can provide a strong support for the real-time prediction of surface soil bulk density.