三种森林生物量估测模型的比较分析.PDF

植物生态学报 2011, 35 (4): 402–410 doi: 10.3724/SP.J.1258.2011.00402 Chinese Journal of Plant Ecology 三种森林生物量估测模型的比较分析 * 范文义 张海玉 于 颖 毛学刚 杨金明 东北林业大学林学院, 哈尔滨 150040 摘 要 森林生物量的定量估算为全球碳储量、碳循环研究提供了重要的参考依据。该研究采用黑龙江长白山地区的TM影 像和133块森林资源一类清查样地的数据, 选取地学参数、遥感反演参数等71个自变量分别构建多元逐步回归模型、传统BP (back propagation)神经网络模型和基于高斯误差函数的BP神经网络改进模型(Gaussian error function, Erf-BP), 进而估算该地 区的森林生物量, 并进行比较分析。结果表明, 多元逐步回归模型估测的森林生物量预测精度为75%, 均方根误差为26.87 t ·m–2 ; 传统BP神经网络模型估测森林生物量的预测精度为80.92%, 均方根误差为21.44 t·m–2 ; Erf-BP估测森林生物量的预测 精度为82.22%, 均方根误差为20.83 t·m–2 。可见, 改进后的Erf-BP能更好地模拟生物量与各个因子之间的关系, 估算精度更 高。 关键词 生物量, BP神经网络模型, 基于高斯误差函数的BP神经网络改进模型, 多元逐步回归 Comparison of three models of forest biomass estimation * FAN Wen-Yi , ZHANG Hai-Yu, YU Ying, MAO Xue-Gang, and YANG Jin-Ming School of Forestry, Northeast Forestry University, Harbin 150040, China Abstract Aims Quantitative estimation of forest biomass is significant to studies of global carbon storage and carbon cy- cle. Our objective is to develop models to estimate forest biomass accurately. Methods Multi-stepwise regression model, traditional back propagation (BP) neutral network model and BP neutral network model based on Gaussian error function (Erf-BP) were developed to estimate forest biomass in Changbai Mountain of Heilongjiang, China according to TM imagery and 133 plots of forest inventory data. There were 71 dependent variables of geoscience and remote sensing. Important findings The precisions and root mean square errors of multi-stepwise regression model, traditional –2 –2 –2 BP neutral network model and Erf-BP were 75%,