【作者】 董敏；

【导师】 王昌全；

【作者基本信息】 四川农业大学 ， 土壤学， 2009， 硕士

【摘要】 土壤是受自然因素以及人为因素共同作用而形成。土壤并非一个匀质体,而是一个时空连续的变异体,具有高度的空间异质性。不论在大尺度还是在中小尺度上观察,土壤的空间异质性均存在。由于经济和人力的原因,采样点是有限的,要得到整个研究区的土壤属性信息就必须借助于空间插值方法。但土壤属性的复杂性和不确定性,使常用的插值方法精度往往不能满足要求,因此,改进空间插值方法,以较少样点获取更为详尽的土壤属性空间变异规律,就成为当前研究的热点。本研究以眉山太和镇和尚义镇为试验区,分别以700米和50米为间隔,采集大小两个尺度下的土壤样品80个和30个,以土壤有效铜和有效锌为研究对象,将两个样本分为独立的训练样本和检验样本数据集,在训练样本集的基础上共设计了4种样点布局方案,以广泛应用的普通克里格法和RBF神经网络作为对照,研究GARBF神经网络技术对土壤属性信息的空间插值能力。GARBF神经网络即遗传径向基函数神经网络具有很强的非线性计算能力,是解决非线性系统预测问题行之有效的研究工具。本研究将地理坐标(X,Y)和距离插值点最近的5个点作为网络的输入值,即网络的输入节点数为7,输出为采样点或者未知点的土壤属性值,输出节点为1。模型在对现有的样本进行学习和训练中,利用遗传算法优化计算RBF神经网络隐层到输出层的权值,直到网络掌握了这些输入—输出之间的对应关系为止,再用训练好的GARBF神经网络来预测各个插值点元素的含量,将预测结果存为记事本格式文件,借助ArcGIS9.0软件转成Raster文件,生成插值栅格图。研究结果表明:(1)在大、小个尺度不同布局方案(a、b、c、d)下,由三种方法(GARBF、RBF、Ordinary Kriging)的散点图可知,GARBF神经网络利用遗传算法优化RBF神经网络权值,提高了对训练样本(有效铜、有效锌)的网络拟合能力,三种插值方法对训练样本的拟合能力为GARBF＞RBF＞Ordinary Kriging。(2)以平均绝对误差和误差均方根作为插值精度的评价指标,无论在大尺度还是在小尺度中GARBF神经网络的插值精度均高于RBF神经网络和普通克里格插值法。主要结果如下:在大尺度中GARBF与RBF神经网络相比,有效铜和有效锌训练样本的逼近误差分别降低0.02～0.01(a-Cu)、0.20～0.22(b-Cu)、0.22～0.25(a-Zn)、0.10～0.11(b-Zn),检验样本的插值误差分别降低0.23～0.26(a-Cu)、0.16～0.12(b-Cu)、0.13～0.11(a-Zn)、0.02～0.13(b-Zn);GARBF与Ordinary Kriging相比,训练样本的逼近误差分别降低1.18～1.43(a-Cu)、0.98～1.16(b-Cu)、1.19～1.47(a-Zn)、1.46～1.87(b-Zn),检验样本的插值误差分别降低0.57～0.30(a-Cu)、0.02～0.05(b-Cu)、0.14～0.20(a-Zn)、0.51～0.24(b-Zn)。在小尺度中GARBF与RBF神经网络相比,有效铜和有效锌训练样本的逼近误差分别降低0.01～0.01(c-Cu)、0.21～0.25(d-Cu)、0.10～0.12(c-Zn)、0.10～0.11(d-Zn),检验样本的插值误差分别降低0.12～0.12(c-Cu)、0.01～0.04(d-Cu)、0.04～0.07(c-Zn)、0.05～0.05(d-Zn);与Ordinary Kriging相比,训练样本的逼近误差分别降低0.49～0.69(c-Cu)、1.19～1.31(d-Cu)、0.38～0.53(c-Zn)、0.36～0.60(d-Zn);检验样本的插值误差分别降低0.37～0.42(c-Cu)、0.53～0.59(d-Cu)、0.01～0.08(c-Zn)、0.08～0.27(d-Zn)。GARBF神经网络的各项误差最小,故插值精度最高。(3)不同尺度下,三种方法的空间插值图有一定的差异。大尺度中三种方法的插值图都呈现出相似的分布规律。小尺度中三种插值方法的插值图都有一定的差异性,具体体现在总体趋势的不同和代表元素含量高低的栅格分布的范围不同,这可能与小尺度采样点减少有关。总体而言,以神经网络的插值图更能表现土壤元素的空间异质性尤其是GARBF神经网络在尊重原始实测数据值的情况下,整体分布相对离散,斑块更丰富,突出了数据分布的波动性,更能反映土壤属性的实际空间分布状况。这主要是因为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面积,在一定程度上能摆脱类似克里格插值的“平滑效应”。更多还原

【Abstract】 Soil property is the result of combined action of nature and human factors.soil is non-homogeneous body, but space-time successive variant,with higher spatial heterogeneity. Soil has spatial heterogeneity in large scale or small scale. Because of economy and manpower, sampling sites are definite, so in order to obtain soil spatial information, we should rely on interpolation method. Due to the complexity and uncertainty of soil property, the precision of common interpolation method always can not often meet request. So, improving the spatial interpolation method to obtain the laws of soil spatial variability with lesser sampling sites became a hotspot in the research field now.TaiHe town and Shangyi town, MeiShan county were selected for the research. Two samples were collected in different scale, one was collected in large scale including 80 soil points,their interval was 700m, the other one was collected in small scale including 30 soil points and their interval was 50m. Available zinc and available cuprum were selected as the object for research.The samples were divided into training and validation datum sets. In order to research the performance of GARBF Neural Networks for soil spatial information interpolation, 4 sampling schemes were designed based on the two training sample set and the performance of GARBF Neural Networks was compared with RBF Neural Networks and Ordinary Kriging method which were widely applied.GARBF Neural Networks(Genetic Algorithm Radias Basis Function Neural Networks), which had strong nonlinear computing competence, was an effective tool for solving the nonlinear system problem. In this paper, the coordinate of the soil points and the values of the 5 soil points which were close to the points needed interpolating were designed as the input of net, so there were 7 nodes in the input layer, the value of the soil property of sampling sites or unknown soil point was the output of the net and there was 1 node in the output layer. In the course of studying and being trained, the model optimized the weights between RBF neural network hidden layer and output layer, using Genetic Algorithm.until the Network learns the relationship between input and output, GARBF can predict the content of each interpolation site. The result of predict value save as .txt file, it can be transferred Raster file and formed the interpolation figure of grid by using the ArcGIS 9.0. The results indicated:(1) The large and small scale under four different sampling schemes (a & b & c & d),the scatter figure of three methods (GARBF and RBF and Ordinary Kriging) showed GARBF used genetic algorithm optimize weight of RBF neural network, and raised the fitting capacity for training sample .The fitting capacity of three methods applied to the same area followed the sequence of GARBF > RBF > Ordinary Kriging.(2) Average absolute error and root mean square error were chosen to judge precision of the interpolation methods, whether the points was collected in large or small scale, the interpolation precision of GARBF neural network method excelled RBF neural network and the Kriging method. The main results as follow:In large scale, GARBF as compared with RBF, the approximate errors of the training samples about available cuprum were reduced by 0.02～0.01 in Scheme a and by 0.20～0.22 in Scheme b, available zinc by 0.22～0.25 in Scheme a and by 0.10～0.11 in Scheme b. The interpolation errors of the test samples about available cuprum by 0.23～0.26 in Scheme a and by 0.16～0.12 in Scheme b, available zinc by 0.13～0.11 in Scheme a and by 0.02～0.13 in Scheme b. GARBF as compared with Ordinary Kriging, the approximation errors of the training samples about available cuprum were reduced by 1.18～1.43 in Scheme a and by 0.98～1.16 in Scheme b, available zinc by 1.19～1.47 in Scheme a and by 1.46～1.87 in Scheme b. The interpolation errors of the test samples about available cuprum by 0.57～0.30 in Scheme a and by 0.02～0.05 in Scheme b, available znic by 0.14～0.20 in Scheme a and by 0.51～0.24 in Scheme b.In small scale, GARBF as compared with RBF, the approximate errors of the training samples about available cuprum were reduced by 0.01～0.01 in Scheme c and by 0.21～0.25 in Scheme d,available zinc by 0.10～0.12 in Scheme c and by 0.10～0.11 in Scheme d, then The interpolation errors of the test samples about available cuprum by 0.12～0.12 in Scheme c and by 0.01～0.04 in Scheme d, available zinc by 0.04～0.07 in Scheme c and by 0.05～0.05 in Scheme d; GARBF as compared with Ordinary Kriging, the approximation errors of the training samples about available cuprum were reduced by 0.49～0.69 in Scheme c and by 1.19～1.31 in Scheme d ,available zinc by 0.38～0.53 in Scheme c and by 0.36～0.60 in Scheme d ,The interpolation errors of the test samples about available cuprum by 0.37～0.42 in Scheme c and by 0.53～0.59 in Scheme d , available znic by 0.01～0.08 in Scheme c and by 0.08～0.27 in Scheme d. So it was obvious that the GARBF neural network was the least in error and the highest in interpolation precision.(3) In different scale, there were some difference in the interpolation map of three methods.The interpolation map of three methods had similar distribution tendency in large scale. But in small scale, the interpolation map of three methods had some diversity. That is, total tendency was different and the rang of raster distribution represented the elements was different too, and it was possiblily related to decrease of sampling point. In general, the interpolation map of nerual network showed the elements of soil spatial heterogeneity predominantly, GARBF neural network respect for the value of the original data especially, the overall distribution was discrete relative, plaque was abundant, distribution of the data variability was prominent and reflected the soil properties of the spatial distribution in the actual situation preferably. The reason was genetic algorithm overcame the tendency of neural networks to land in local optima and expanded 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.更多还原