基于BP神经网络确立森林健康快速评价指标  

Establishing Indices for Rapid Assessment of Forest Health Based on BP Neural Networks

作　　者：甘敬[1] 朱建刚[1] 张国祯[2] 余新晓[1] 

机构地区：[1]北京林业大学,北京100083 [2]北京市园林绿化局,北京100029

出　　处：《林业科学》2007年第12期1-7,共7页Scientia Silvae Sinicae

基　　金：北京市科委重大项目(D0706001000091);国家"十一五"科技支撑计划项目(2006BAD03A0201)资助

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≥16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-6.1409%,回归方程斜率为0.9683,截距为0.0490,Nash-Sutcliffe效率为0.9054,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。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 training data and the consistency tests between simulation outputs and expected outputs with three methods including percent error, linear regression and Nash-Sutcliffe efficiency. The results 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 resulting from precision assessment of forest health （PAFH）. The reults of consistency tests demonstrated with mean relative error （ - 6. 140 9% ）, the Nash-Sutcliffe efficiency （0.905 4） as well as the slope （ a = 0.968 3） and the intercept （ b = 0.0490） of the regression equation indicated high consistency between simulation outputs and excepted outputs. Therefore, stand structure, severity of pest and disease and soil thickness could be considered as indices of rapid assessment of forest health （ RAFH）.

关 键 词：森林健康快速评价 指标 BP神经网络 合理性检验 

分 类 号：S758.8[农业科学—森林经理学]