[Objective]Twelve plots of natural mixed spruce-fir coniferous and broadleaf forests located in Jin'gouling Forest Farm of Jilin Province were investigated to establish height prediction models for main tree species based on 12 953 data of tree height, diameter and dominant height by using BP neural network.[Method] After determining the hidden nodes, an optimum model structure was developed by training BP models of Larix olgensis, Picea spp., Abies nephrolepis, Pinus koraiensis and two deciduous groups repeatedly.Then, they were compared with two traditional height-diameter equations in which the parameters were solved with the same input datasets from 8 plots to establish BP models, and the validation datasets from the other 4 plots were used to test the models.[Result] The results show that the optimal network structure of L.olgensis and Picea spp.(nodes in input layers: nodes in hidden layers: nodes in output layers) are both 2:5:1, the optimal network structure of Pinus koraiensis, one deciduous group (Betula platyphylla, Populus ussuriensis, Ulmus pumila and other tree species) are both 2:4:1, the optimal network structure of A.nephrolepis is 2:8:1, and the optimal network structure of the other deciduous group (Acer mono, Fraxinus mandschurica, Phellodendron amurense, Tilia amurensis, and Betula costata) is 2:7:1.[Conclusion] Compared with traditional methods, the BP models need not rely on existing functions or choose model forms.The R2 of BP models are higher than that of the traditional models, and both the mean absolute error and root mean square error of BP models are less than that of the traditional models.The fitting accuracy and prediction effect of BP neural network models are better than those of traditional equations, and thus can predict tree height effectively.%[目的]以吉林省汪清林业局金沟岭林场12块天然云冷杉针阔混交林样地为对象,基于12 953对实测树高-胸径数据,结合林分优势高分树种(组)建立基于BP神经网络的标准树高模型.[方法]在确定隐层节点数后经过反复训练得到各树种(组)的适宜模型结构,使用相同的建模数据(8块样地)求解两个传统的树高方程,再利用未参与建模的4块样地分别验证模型.[结果]表明:落叶松、云杉的适宜模型结构(输入层节点数:隐藏层节点数:输出层节点数)为2:5:1;红松、中阔(白桦、大青杨、榆树和杂木)的适宜模型结构为2:4:1;冷杉的适宜模型结构为2:8:1;慢阔(色木、水曲柳、黄檗、紫椴和枫桦)的适宜模型结构为2:7:1.[结论]与传统方法相比,BP模型不依赖现存函数,不需要筛选模型形式,而且BP模型各树种R2高于传统模型,平均绝对误差、均方根误差均小于传统模型,其拟合精度和预测效果均优于传统方程,可以有效地预测树高.
展开▼
