基于思维进化算法径向基函数神经网络的土壤有机质空间异质性研究

Spatial heterogeneity of soil organic matter based on mind evolutionary computation radial basis function neural network

提出一种基于思维进化算法径向基函数神经网络的土壤有机质空间异质性研究方法(MECRBF)。以江西省万年县为案例区,在全县范围内采集耕地表层(0~20 cm)土壤样品954个,分别采用该方法和以邻近信息和地理坐标为输入的径向基函数神经网络方法(RBF-Near),以及普通克里格法来模拟土壤有机质空间分布。以验证样本实测值和预测值的决定系数与逼近误差作为评判标准,对比各方法在县域尺度上土壤有机质空间变异和空间插值方面的效果。对763个采样点进行建模、191个验证样点进行独立验证的预测结果表明,在验证点预测中,MECRBF的均方根误差、平均绝对误差、平均相对误差较RBF-Near分别降低了0.50 g·kg^(-1)、0.39 g·kg^(-1)、1.40百分点,差异显著(P<0.05),较普通克里格法分别降低2.59 g·kg^(-1)、1.89 g·kg^(-1)、7.76百分点,差异显著(P<0.05)。从模拟效果来看,MECRBF的决定系数最高,逼近误差最小;从空间分布模拟图来看,MECRBF能更好地表达土壤有机质空间异质性。提出的MECRBF可为县域尺度下土壤性质空间异质性研究提供方法参考。

In the present study,a method named MECRBF was proposed based on mind evolutionary computation radial basis function neural network.Its ability to reveal spatial heterogeneity of soil organic matter was compared with radial basis function neural network(RBF-Near)based on spatial coordinates and neighbor information,and ordinary Kriging method with Wannian County,Jiangxi Province as study area.To establish and validate,954 soil samples were collected and randomly divided into 2 groups,i.e.modeling points(763)and validation points(191).Spatial distribution prediction capacities and prediction map of these methods were compared.It was shown that the root mean square errors(RMSE),mean absolute errors(MAE)and mean relative errors(MRE)of MECRBF in validation points was 0.50 g·kg-1,0.39 g·kg-1,and 1.40 percent smaller than those of RBF-Near(P<0.05),respectively,and was 2.59 g·kg-1,1.89 g·kg-1,and 7.76 percent smaller than those of ordinary Kriging(P<0.05),respectively.The prediction map obtained by MECRBF was more consistent with the actual geographical information than the others.Moreover,MECRBF method reduced the prediction errors.The proposed MECRBF could provide guidance to predict soil nutrients at county scale.