丘陵区土壤有机质空间分布预测的神经网络方法  

Predict the Spatial Distribution of Soil Organic Matter for a Hilly Region with Radial Basis Function Netural Network

作　　者：李启权[1] 王昌全[1] 张文江[2] 余勇[3] 李冰[1] 杨娟[1] 白根川[1] 刘泳宏[1] 

机构地区：[1]四川农业大学资源环境学院,成都611130 [2]四川大学水力学与山区河流开发保护国家重点实验室,成都610065 [3]四川农业大学林学院,四川雅安625014

出　　处：《农业环境科学学报》2012年第12期2451-2458,共8页Journal of Agro-Environment Science

基　　金：国家杰出青年科学基金(40825003);国家自然科学基金项目(41201214;40801175)

摘　　要：土壤性质空间分布信息的准确表达是土壤资源优化利用和土壤环境保护的需要。为模拟川中丘陵区县域尺度上土壤有机质的空间分布格局,构建了以地理坐标、地形和植被因子为网络输入的径向基函数神经网络模型(RBFNN_E),并将该方法与普通克里格法(OK)、多元回归模型(MLR)和仅以地理坐标为网络输入的神经网络模型(RBFNN_C)相比较。结果表明:RBFNN_E对479个验证点模拟结果的平均绝对误差(MAE)、平均相对误差(MRE)和均方根误差(RMSE)较MLR分别降低了1.74%、1.45%和2.64%,较OK分别降低了7.77%、12.76%和3.92%,较RBFNN_C分别降低了8.89%、9.81%和7.68%。从模拟的空间分布图来看,RBFNN_E能较好地刻画环境变化引起的土壤有机质空间变异的细节信息。因此,融合环境因子的神经网络模型(RBFNN_E)不仅具有较高的模拟精度,还能更好地揭示复杂地形下土壤有机质的空间变异,使模拟结果更符合区域地学规律与实际情况,可为复杂环境条件下土壤管理、精准农业的实施以及区域环境规划等提供科学依据。Accurate spatial information of soil properties at regional scale is essential to land use and environment management.This paper proposed a radial basis function neural network method for predicting the spatial distribution of soil organic carbon（SOM） in the typical hilly region of Sichuan Basin,which uses geographic coordinates,terrain factors and vegetation index as inputs（RBFNN_E）.Its performance was compared with that of ordinary kriging（OK）,multiple linear regression mode（l MLR） and a radial basis function neural network model only using geographic coordinates as inputs（RBFNN_C）.The results of 479 validation points showed that RBFNN_E obtained lower estimation bias.The mean absolute error（MAE）,root mean squared error（RMSE） and mean relative error（MRE） of RBFNN_E were smaller than those of MLR respectively by 1.74%,1.45% and 2.64%,smaller than those of OK respectively by 7.77%,12.76%,3.92%,and smaller than those of RBFNN_C respectively by 8.89%,9.81% and 7.68%.Moreover,RBFNN_E produced the SOM map with much more details,which were con-tributed by the adoption of environmental factors as inputs.The results suggested the method of radial basis function neural network,which adopted environmental factors as inputs,can not only improve the prediction accuracy but also respond to the spatial variation of soil organic carbon over the environment variation.Therefore,RBFNN_E can help to produce the SOM map with higher accuracy which was consistent with the true geographical information.This method provides a useful tool for the accurate prediction of soil properties for the typical hilly re-gion of Sichuan Basin,the area with complex environment.

关 键 词：土壤 植被 模型 地形因子 丘陵区 径向基函数神经网络 空间预测 

分 类 号：S153.6[农业科学—土壤学]