基于支持向量机的近红外特征变量选择算法用于树种快速识别

Fast Identification of Wood Species Using Near Infrared Spectroscopy Coupled with Variables Selection Methods Based on Support Vector Machine

将稳定度自适应重加权采样特征变量选择算法用于支持向量机定性分析(Support vector machine-stability competitive adaptive reweighted sampling,SVM-SCARS)。该算法通过对数据多次采样建模计算各变量的稳定度值,稳定度值能更加客观准确地评估变量在建模中的作用,因此可作为变量重要性的评价依据。通过循环迭代方式,采用自适应重加权采样技术逐步筛选变量,然后以每次循环所得变量子集建立SVM模型,并以模型交叉验证分类正确率(Correct classification rate of cross validation,CCRCV)评估子集优劣,确定最优特征变量子集。将该算法结合漫反射近红外光谱技术建立了制浆造纸常用木材的树种识别模型,实现了对4种桉木和2种相思木的快速识别分类。最终共筛选出15个特征变量建立分类模型,模型对各树种分类的正确率达97.9%,具有较好的分类效果。与全光谱模型和递归特征消除支持向量机模型相比,SVM-SCARS能够筛选出更少的特征变量,且模型具有更好的预测性能和稳定性。研究结果表明,SVM-SCARS算法能够有效优化光谱特征变量,提高近红外在线分析模型在木材材性分析中的稳健性和适用性。

A novel variable selection method based on stability competitive adaptive reweighted sam- piing was applied to work with support vector machines（ SVM - SCARS） for selecting informative vari- ables of near infrared spectroscopy to build more robust SVM model. This method computed the sta- bility index of each variable from a statistical analysis of weight vectors of multiple SYMs trained on subsamples of the original data by multiple sampling. The stability index represents the influence of variable on SVM modeling and could be used to evaluate the importance of variable. The variable with higher stability index was treated as informative variable that has an important effect on predictive performance of the model. Through iterations, the important variables was selected gradually by u- sing adaptive reweighted sampling technology. Then the selected variables in each iteration were stored into variable subset. The optimal variable subset was determined by assessing the correct clas- sification rate of cross validation（CCRCV） of SVM models based on all variable subsets. The SVM - SCARS algorithm combined with near-infrared diffusion reflectance spectrum technology were applied to construct wood identification model for four kinds of eucalyptus and two kinds of acacia. Experi- mental results showed that the SVM -SCARS model has a superior performance for identifying differ-ent wood species, in comparison to the full spectrum model and the support vector machine recursive feature elimination（ SVM - RFE） model, both in terms of prediction ability and selected variables size. As a result, fifteen variables were selected by SVM - SCARS method to construct identification model with the correct classification rate of 97.9%. This study demonstrates that SVM - SCARS could effectively extract important characteristic variables from near infrared spectrum to improve the robustness and applicability of NIR online detection model for wood property analysis.