光谱分析已经在木材特性参数(例如木材树种、 气干密度、 强度、 含水率、 表面粗糙度等)检测中得到应用, 但是, 现有的木材检测研究都只是针对上述某一项参数做数学建模和预测。 如果需要检测木材多项参数, 那么需要进行多次建模, 并且每次建模预测时使用的数学模型类型(例如神经网络的类型)和内部结构参数一般各不相同。 为了提高木材质量检测效率, 提出了一种基于可见光/近红外光谱的木材树种和密度同时预测方法, 它只需要一次建模和预测就可以实现这两项参数的同时输出。 对东北5种常见木材(杨木、 桦木、 樟子松、 白松和落叶松)进行检测, 首先, 采用K/S算法划分样本集, 保证了训练集和预测集具有一定的代表性。 然后, 使用主成分分析和小波变换两种光谱降维方法, 分别与BP神经网络和偏最小二乘支持向量机相结合建立了4种木材树种和密度同时预测模型和预测精度对比。 采用美国海洋公司的Ocean Optics USB2000-VIS-NIR微型光纤光谱仪采集样本的可见光/近红外光谱并进行预测处理, 光谱范围为350～1 100 nm。 结果表明, 这四种模型都可实现对木材树种和密度的同时预测, 其中小波变换降维方法结合偏最小二乘支持向量机所建立的模型预测效果相对较好, 树种正确识别率为100%, 训练集密度的R为0.973 4, 预测集密度的R为0.940 8, 训练集密度的RMSE为0.026 13, 预测集密度的RMSE为0.038 46, 它为同时对木材多项特性参数进行预测的便携式多功能一体化木材光谱检测仪器的开发奠定了理论基础。 此外, 还采用该公司生产的另一款光谱范围为900～1 650 nm的FLAME-NIR型微型光纤光谱仪进行了同样的实验。 对比发现, 利用FLAME-NIR型光谱仪所得出的结果整体比利用USB2000-VIS-NIR型光谱仪所得到的结果好, 但是相差并不是很大。 这说明该方法可用于对木材种类与密度的同时预测, 而且具有一定的稳定性和精度, 也节约了仪器的成本。

Spectral analysis has been widely used in wood physical feature parameter detection such as wood species, density, strength, surface roughness and humidity. However, the current wood detection is used to predict the single wood parameter. If the multiple wood parameter detections are required, the single wood detection needs to be performed some times. In order to improve the wood parameter detection’s efficacy, we propose a simultaneous prediction scheme for wood species and wood density parameters with only one prediction. First, the K/S algorithm is used to divide the training and prediction sets to make them representative. Then, two dimensionality-reduction methods of principal component analysis and wavelet transform are combined with BP neural network and least squares support vector machine to establish four prediction models that can predict both wood species and density. In experiments, a small fiber spectrometer of USA Ocean Optics USB2000-VIS-NIR is used to acquire the visible/near infrared spectral curves with a spectral interval of 350～1 100 nm. The results show that all four models can achieve simultaneous prediction of wood species and density, and the model established by wavelet transform dimensionality-reduction method combined with least squares support vector machine is relatively better. The correct recognition rate of wood species based on the combination of wavelet transform and partial least squares support vector machine is 100%, the density correlation coefficient of training set is 0.973 4, the density correlation coefficient of prediction set is 0.940 8, the density training root mean square error is 0.026 13, and the prediction root mean square error is 0.038 46. It lays a theoretical foundation for the development of portable real-time on-line detection instruments that can simultaneously predict several parameters of wood physical feature. Moreover, another spectrometer of FLAME-NIR with a spectral interval of 900～1 650 nm is also used to perform the same prediction experiments. By comparisons, we find that the prediction results with the FLAME-NIR model are slightly superior to those with the USB2000-VIS-NIR model. Therefore, our simultaneous prediction of wood species and wood density is practical with a definite stability, accuracy, and a low instrumentation cost.