利用近红外光谱结合误差反向传播神经网络（BP）对三种人工林木材（尾叶桉、 马尾松、 南方无性系I-72杨）进行识别, 探讨隐含层神经元个数、 光谱预处理方法、 光谱范围对BP网络模型的影响, 并与SIMCA法所建模型做比较。 结果表明: （1）BP网络结合全波段（780～2 500 nm）近红外光谱数据建模, 识别正确率达到97.78%, 并确定隐含层神经元数为13; （2）全波段光谱建模比短波段（780～1 100 nm）和长波段（1 100～2 500 nm）光谱建模识别效果好, 其识别正确率分别为97.78%, 95.56%和96.67%, 用一阶导数和二阶导数对全波段光谱进行预处理后, BP网络模型识别正确率分别为93.33%和71.11%; 用多元散射校正（MSC）对全波段光谱进行预处理后, BP网络模型识别正确率为98.89%, （3）在三种波段（780～2 500, 780～1 100和1 100～2 500 nm）光谱建模的情况下, BP网络建模识别正确率分别为95.56%, 96.67%和97.78%, SIMCA模型识别正确率分别为76.67%, 81.11%和82.22%, BP网络建模比SIMCA法建模对三种人工林木材的识别正确率高。

In this study, the near infrared spectroscopy coupled with Back-Propagation (BP) network was used for the recognition of three kinds of plantation wood (Eucalyptus urophylla, Pinus massoniana, Populus×euramericana (Dode) Guineir cv. “San Martino” (1-72/58)). The study considered the effects of hidden layer neurons number, spectral pretreatment method and spectral regions on BP model, which are compared with SIMCA model simultaneously. The results showed that, (1) the recognition rate was 97.78% achieved by BP network model with hidden layer neurons number 13 and the spectral region of 780～2 500 nm. (2) BP model with spectral region of 780～2 500 nm was more robust than the other two BP models with spectral regions of 780～1 100 and 1 100～2 500 nm, of which recognition rates were 97.78%, 95.56% and 96.67%, respectively. After the full spectra was pretreated with the first derivative and the second derivative methods, the recognition rates of BP models fell down to 93.33% and 71.11%. However, the recognition rate of BP model rose to 98.89% with the full spectra being pretreated by the multiplicative scatter correction (MSC). (3) Compared with SIMCA models that recognition rates of three spectral regions (780～2 500, 780～1 100 nm, and 1 100～2 500 nm) were 76.67%, 81.11% and 82.22% respectively, BP network work models had higher recognition rates.