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

Li Qiquan,Wang Changquan,Yue Tianxiang,Zhang Wenjiang,Yu Yong
DOI: https://doi.org/10.11867/j.issn.1001-8166.2012.02.0175
2012-01-01
Abstract: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.
What problem does this paper attempt to address?