基于BP神经网络的油松人工林树高模型研究  

Establishment of Tree Height Model of Pinus tabuliformis Plantation Based on BP Neural Network

作　　者：陈佳琦 赵鹏祥 祁宁 李卫忠 CHEN Jia-qi;ZHAO Peng-xiang;QI Ning;LI Wei-zhong(College of Forestry,Northwest A&F University,Yangling 712100,Shaanxi,China)

机构地区：[1]西北农林科技大学林学院

出　　处：《西北林学院学报》2020年第1期212-217,245,共7页Journal of Northwest Forestry University

基　　金：国家重点研究发展计划项目(2016YFD0600203);陕西省林业重点项目(SHLY-2018-02)

摘　　要：通过分析比较不同算法以及不同输入层因子,构建出最佳的黄龙山区油松人工林树高预测BP神经网络模型。以陕西省延安市黄龙县44块油松人工林样地实测数据为数据源,通过对6种BP神经网络的训练方法进行训练,经过反复筛选找出最优模型并与传统胸径-树高模型作比较;最后将BP神经网络中的输入因子从2个增加到6个后,经过反复训练筛选出最优模型与2因子的BP神经网络模型作比较。结果表明:1)贝叶斯归一化(BR)算法在6种算法中表现最佳,R^2和MSE分别为0.9630和1.168;2)不同隐含层节点数的选取会对BP神经网络模型的建立产生一定的影响,BP神经网络模型的决定系数(R^2)随着隐含层节点数的增加呈现先上升后下降的趋势;均方误差(MSE)呈现先下降后上升的趋势,两者都在节点数为10时有极值,此时的模型为最优模型;3)当输入因子为胸径和优势树高时,油松人工林的最优模型结构为(输入层节点数:隐含层节点数:输出层节点数为2∶10∶1),此时BP神经网络模型对树高预测的决定系数(R^2)和均方误差(MSE)分别为0.7610和1.9847;当输入因子为胸径、优势树高、林分密度、竞争指数、坡度和坡向时,最优模型结构为6∶10∶1,此时BP神经网络模型对树高预测的决定系数(R^2)和均方误差(MSE)分别为0.8447和1.9557。由此得出,在建立油松人工林树高BP神经网络模型方面优化类算法要优于启发式下降算法;BP神经网络模型与传统模型相比,BP神经网络模型不需要目标方程结构,并且模拟和预测的精度均要优于传统模型;在原有BP神经网络模型的基础上再引入林分密度、竞争指数、坡度、坡向这些输入因子后所得到的新的BP神经网络模型对树高模型的建立和预测要优于原有BP神经网络模型。We established the optimal tree height prediction model of Pinus tabuliformis based on BP(back propagation) neural network by analyzing and comparing different algorithms and different input layer factors.We trained the data by six kinds of BP neural network algorithms based on measured data of 44 plots of P.tabuliformis plantation in Huanglong County,Yanan City,Shaanxi Province,to select the optimal model and to compare with the traditional models.After increasing the number of input factors of BP neural network from 2 to 6,we compared these 2 models.1) BR algorithm performed the best among 6 algorithms,the coefficient of determination(R^2) and mean square error(MSE) were 0.9630 and 1.168,respectively.2) Different hidden layer nodes will affect the accuracy of BP neural network models.As hidden layer nodes increased,the coefficient of determination(R^2) increased first and then decreased,the mean square error(MSE) decreased first and then increased.Both of them presented extreme values when hidden layer nodes equal to 10,and the model was the optimal one.3) When tree height and DBH were used as input factors,the optimal structure of P.tabuliformis plantation was input factor nodes∶hidden layer nodes∶output factor nodes=2∶10∶1,and the determinant coefficient(R^2) and mean square error(MSE) of BP neural network model for tree height prediction were 0.7610 and 1.9847,respectively.When tree height,DBH,stand density,competition index,slope gradient and slope direction were used as input factors,the optimal structure of P.tabuliformis plantation was 6∶10∶1 and the determinant coefficient(R^2) and mean square error(MSE) of BP neural network model for tree height prediction were 0.844 7 and 1.955 7,respectively.The optimization algorithm was superior to the heuristic descent algorithm in establishing BP neural network model of tree height of P.tabuliformis plantation.Compared with the traditional model,the BP neural network model did not need the structure of objective equation,and the accuracy of simulation

关 键 词：BP神经网络 树高模型 黄龙山 油松人工林 

分 类 号：S791.254[农业科学—林木遗传育种]