基于BP网络和回归分析的红壤可蚀性预测

Prediction of red soil erodibility based on BP neural network and regression analysis

基于野外人工模拟降雨试验及室内土壤理化指标测定,研究广东省红壤可蚀性并构建其预测模型。可蚀性因子K与土壤黏粒、砂粒和有机质质量分数呈极显著相关关系(P<0.01),皮尔逊相关系数分别为-0.920、0.925和-0.642。采用逐步回归分析法筛选对K值有显著影响作用的11个单因子及多因子交互作用项,其中粉粒与有机质质量分数交互作用对K值影响最大。K值逐步回归预测模型效果较好,相对误差均小于25%。这11个单因子及多因子交互作用项为输入变量,K值为输出变量,建立K值BP神经网络预测模型。利用灰色关联度分析法确定了BP神经网络最优结构为11-11-1,网络训练算法为Levenberg-Marquardt。K值BP神经网络预测模型90%数据点相对误差小于10%,其精度明显优于逐步回归分析模型,能更为准确地反映影响因子与K值间的内在规律。

We investigated the soil erodibility（ K） of red soil in Guangdong Province and established its predication model based on field artificial rainfall events and indoor measurement of soil physical and chemical indexes. There was a significant correlation between K and clay,sand and organic matter content（ P〈 0. 01）,with Pearson＇s correlation coefficient of- 0. 920 4,0. 925 9 and- 0. 642 4,respectively. The stepwise regression analysis（ SRA） filtrated the single factors and interacted terms,which significantly affected K. The interacted influence between the silt and organic content was the largest. The SRA prediction model for K had a good estimation,and the relative error between predicted and observed values was less than 25%. Using the single factors and interacted terms filtrated by SRA as the input parameter,and K data as the output parameter,we established the BP predication model for K.The grey relational analysis indicated that the optimal network structure was 11-11-1 and the best training algorithm was the Levenberg-Marquardt. The relative error of over 90% data was less than 10% by the BP prediction model. The precision of BP model was obviously better than that of SRA model,and the former can more accurately reflect the inherent law between K and factors.