基于RBF神经网络的土壤重金属空间变异研究

本文采用径向基函数神经网络(Radial Basis Function Neural Network,RBFNN)预测太原市晋源区表层土壤中重金属Cr、Cd、Hg的空间变异,并与普通克里格(Ordinary Kriging)插值结果进行对比分析,以选择更合适的土壤重金属空间插值方法。研究结果表明:1)在拟合RBFNN模型过程中,选择合适的spread散布常数可以使模型达到最优,研究区域土壤Cr的最优散布常数为0.08,Cd的最优散布常数为0.10,Hg的最优散布常数为0.14,这组散布常数对于局部区域农田土壤重金属插值模拟有一定的参考意义。2)RBFNN方法与Ordinary Kriging方法对区域重金属浓度分布的预测趋势一致,土壤Cd含量在区域中部较高,尤其是从东北方向到西南方向的轴线上较高,向两侧形成扩散递减趋势;土壤中Cr含量总体分布趋势也是中部较高,其他区域相对较低;土壤Hg含量在区域东北部较高,由东北方向到西南方向浓度逐渐递减。且土壤重金属在区域中的分布与当地的污染源分布相对应。在样本数有限的情况下对土壤重金属进行空间变异研究时,RBFNN方法比Ordinary Kriging方法的预测精度更高更有效。

Application of RBF neural network in determining soil heavy metal spatial variability

The Radial Basis Function Neural Network (RBFNN) was used to predict the spatial variability of Cr, Cd and Hg in the top soils of Jinyuan District, Taiyuan City. The RBFNN and Ordinary Kriging interpolation methods were compared for a more appropriate method of predicting the spatial variability of soil heavy metals. The results showed that proper spread parameter of RBFNN was critical for limiting errors and improving overall model accuracy. The optimal spread parameter values for Cr, Cd and Hg were 0.08, 0.10 and 0.14, respectively. These values were usable as the basis of reference for determining the spatial distribution of heavy metals in local farmland soils in the study area. Both the RBFNN and Ordinary Kriging interpolation methods predicted the spatial distribution of soil heavy metals with similar tendencies. Although soil Cd concentration was higher in the central region of research area (and especially in the axis from northeast to southwest), it gradually decreased from the axis to the side regions. Soil Cr concentration was also higher in the central region than in other areas. Soil Hg concentration was higher in the northeast of research area, but also gradually decreased from northeast to southwest. Generally, the spatial distributions of Cr, Cd and Hg corresponded with the sources of pollution distribution in the research area. For limited sample sizes, the RBFNN method was more sensitive and suitable for predicting the spatial distribution of heavy metals than the Ordinary Kriging method.