基于BP神经网络确立森林健康快速评价指标.pdfVIP

第43卷 第 l2期 林 业 科 学 Vo1．43．No．12 2 0 0 7年 l2月 SCIENTIA SILVAE SINICAE Dee．，2 0 0 7 基于BP神经网络确立森林健康快速评价指标* 甘 敬 朱建刚 张国祯 余新晓 (1．北京林业大学 北京 100083；2北京市园林绿化局 北京 100029) 摘 要： 拟定森林健康快速评价(RAFH)指标，通过对训练样本的模式识别来构建一个BP神经网络，观察其能否 收敛，并以测试样本为新的输入项进行模拟，采用误差百分比法、线性回归检验法和Nash—Sutcliffe效率法对模拟值 与期望值的吻合程度进行检验，以此验证拟定指标的合理性。结果表明：在隐含层神经元 n≥l6时，网络能较好地 收敛，说明该网络输入项——林分层次结构、病虫害程度和土壤厚度3个指标的训练样本值与目标输出项——森 林健康精准评价(PAFH)结果的非线性相关程度高；模拟值与期望值的相对误差均值为一6．140 9％，回归方程斜率 为0．968 3，截距为0．049 0，Nash—Suteliffe效率为0．905 4，均表明二者之间吻合较好。因此，林分层次结构、病虫害程 度和土壤厚度可以作为森林健康快速评价(RAFH)的指标。 关键词：森林健康快速评价；指标；BP神经网络；合理性检验 中图分类号：$758．8 文献标识码：A 文章编号：1001—7488(2007)12—0001—07 Establishing Indices for Rapid Assessment of Forest Health Based on BP Neural Networks Gan Jing Zhu Jiangang Zhang Guozhen Yu Xinxiao (1 Beifing Forestry University Beijing 100083；2．Beifing Municipal Bu~au ofParks andAfforestation Beijing 100029) Abstract： The indices of rapid assessment of forest health(RAFH)were given and their rationalities were tested based on BP neural networks in terms of convergence effects of BP networks established according to pattern recognition of Iraining data and the consistency tests between simulation outputs and expected outputs with three methods including percent error，linear regression and Nash—Sutcliffe efficiency．The resuhs of convergence effects showed that the networks could converge properly with 16 or more neurons in hidden layer，which indicated that there was a significant，nonlinear correlation between the inputs derived from training data values of 3 indices consisting of stand structure，severity of pest and disease and soil thickness and the target outputs res