基于GARBF神经网络的土壤有效锌空间插值方法研究  

STUDY ON SOIL AVAILABLE ZINC WITH GA-RBF-NEURAL-NETWORK-BASED SPATIAL INTERPOLATION METHOD

作　　者：董敏[1] 王昌全[1] 李冰[1] 唐敦义[2] 杨娟[1] 宋薇平[1] 

机构地区：[1]四川农业大学资源环境学院,四川雅安625014 [2]成都市龙泉驿区农村发展局,成都610100

出　　处：《土壤学报》2010年第1期42-50,共9页Acta Pedologica Sinica

基　　金：四川省教育厅重点项目(2006ZD003;2006A013;07ZS002);四川省科技厅项目(2008JY0095)资助

摘　　要：以土壤有效锌为研究对象,构建遗传径向基函数(GARBF)神经网络对该元素属性值进行空间插值,以训练样本集的测定值与预测值之间的决定系数、逼近误差及检验样本的插值误差为评判标准,比较GARBF神经网络、径向基函数(RBF)神经网络、普通克里格(Ordinary Kriging)的拟合能力和空间插值能力。结果表明:同一区域两种抽样方案(a、b)下三种插值方法对训练样本的拟合能力为GARBF>RBF>Or-dinary Kriging。以平均绝对误差和误差均方根作为插值精度的评价指标,GARBF与RBF神经网络相比,训练样本的逼近误差分别降低0.22～0.25(a方案)和0.10～0.11(b方案),检验样本的插值误差分别降低0.13～0.11(a方案)和0.02～0.13(b方案);GARBF神经网络与Ordinary Kriging相比,训练样本的逼近误差分别降低1.12～1.40(a方案)和1.45～1.88(b方案),检验样本的插值误差分别降低0.20～0.24(a方案)和0.14～0.32(b方案),GARBF神经网络的误差最小,插值精度最高。从GARBF神经网络的插值图可以看出,遗传算法避免了神经网络容易陷入局部最优点,扩大了对土壤中相关空间信息的搜索范围,在一定程度上避免了类似克里格插值的"平滑效应"。A spatial interpolation method based on GARBF neural network was used to study available zinc in soil. Comparison between GARBF neural network, RBF neural network and Ordinary Kriging interpolation method in fitting capacity and spatial interpolation capacity, with determination coefficient between measured values and predicted values of the training sample set, approximate error and interpolation error of test samples cited as criteria for evaluation. Results show that in terms of the fitting capacity, the three methods applied to the same area under two different sampling schemes （a ＆ b） followed the sequence of GARBF 〉 RBF 〉 Ordinary Kriging. When average absolute error and root mean square error were chosen to judge precision of the interpolation methods, comparison between GARBF and RBF showed that the approximate errors of the training samples were reduced by 0.22 - 0.25 in Scheme a and by 0. 10 - 0. 11 in Scheme b, and the interpolation errors of the test samples by 0. 13 - 0. 11 in Scheme a and by 0.02 - 0. 13 in Scheme b. Comparison between GARBF and Ordinary Kriging showed that the approximation errors of the training samples were reduced by 1.12 - 1.40 in Scheme a and by 1.45 - 1.88 in Scheme b and the interpolation errors of the test samples by 0.20 - 0.24 in Scheme a and by 0. 14 - 0.32 in Scheme b. So it is obvious that the GARBF neural network is the least in error and the highest in interpolation precision. The GARBF interpolation map reveals that the application of genetic algorithm overcomes the tendency of neural networks to land in local optima and expands the scope of search of spatial information pertaining to soil, thus to a certain extent avoiding a similar problem of ＂ smooth effect＂ like Ordinary Kriging. The findings of this study could provide a practical analysis tool and decision-making basis for precision fertilization and prevention of soil pollution.

关 键 词：GARBF神经网络 RBF神经网络 普通克里格 空间插值 

分 类 号：S153.6[农业科学—土壤学]