基于神经网络模型的中国表层土壤有机质空间分布模拟方法

Method for Spatial Simulation of Topsoil Organic Matter in China based on a Neural Network Model

基于全国第二次土壤普查得到的6 241个典型土壤剖面数据,采用主成分分析方法和径向基函数神经网络模型建立不同植被类型—土纲单元内土壤有机质与气候、地形和植被等环境因子间的非线性关系,模拟全国表层土壤有机质的空间分析格局。结果表明,该模型具有较准确的预测能力,性能指数达到1.94。与普通克里格法、反比距离法和多元回归模型相比,神经网络模型对621个验证点模拟结果与实测值的相关系数为0.799,分别提高了0.265、0.181和0.120,平均绝对误差分别降低了4.25、4.43和2.34 g/kg,平均相对误差分别降低了30.16%、32.66%和5.93%,均方根误差则分别降低了8.61、8.24和6.24 g/kg;从模拟结果图来看,神经网络模型能够提供更多的细节信息。该方法为大尺度土壤性质空间分布模拟提供了有益的参考。

Accurate simulation of spatial information of soil properties at large scale is essential to land use and environment management.Based on 6 241 typical soil profiles collected during the second national soil survey period（1979-1994）,a radial basis function neural networks（RBFNN） method was used to capture the nonlinear relationship between topsoil organic matter and principal components converted from environmental variables including climate data,terrain attributes and vegetation index,to simulate the spatial pattern of topsoil organic matter at national scale in China.Results show that,the ratio of performance to deviation（RPD） produced by RBFNN is 1.94 which indicates that RBFNN has much stronger ability to simulate the spatial distribution of topsoil organic matter within a large scale.Error analysis based on 621 validation points show that,compared with ordinary kriging,inverse distance weighting and multiple linear regression model,correlation coefficients between the observed value increased by 0.265,0.181 and 0.120,the mean absolute error decreased by 4.25,4.43 and 2.34 g/kg,the mean relative error was reduced by 30.16%,32.66% and 5.93%,and the root mean squared error decreased by 8.61,8.24 and 6.24 g/kg.Moreover,RBFNN produces soil organic matter map with much higher level of detail,which is closer to the actual pattern of soil organic matter in the complex environment,especially for the area where the density of sample points is much smaller.The good performance of RBFNN method can be attributed to the ability of RBFNN method which can accurately capture the nonlinear relationship between soil organic matter and environmental factors.Thus,RBFNN method is considered as an effective tool to spatial simulation of soil properties at a large scale.Further researches are needed to employ more influencing factors into the method to improve the precision of simulation result.