基于高光谱与纹理融合的烤烟分类方法研究

  目的  为提高烤烟的分类正确率。  方法  利用高光谱成像系统采集烤烟样本，采用Savitzky-Golay卷积平滑法（SG）、多元散射校正（MSC）、一阶导数法（FD）的组合方法对数据进行预处理。通过连续投影算法（SPA）选择特征波长，利用灰度共生矩阵（GLCM）选择烤烟的纹理特征，将纹理特征与光谱特征归一化处理后进行融合，利用邻近算法(KNN)、随机森林(RF)、支持向量机(SVM)、朴素贝叶斯（NB）验证烤烟分类效果。  结果  预处理后的全波长数据分类正确率较预处理前有所提升；利用SPA选择特征波段进行分类，正确率下降；高光谱融合纹理特征后进行分类，分类效果显著提升。  结论  基于高光谱与纹理融合可准确、有效地对烤烟进行无损分类，为量化烤烟分类提供了可行方法。 

Research on classification method of flue-cured tobacco based on fusion of hyperspectral and texture features

  Objective  The purpose of this paper is to improve the classification accuracy of flue-cured tobacco.  Methods  A hyperspectral imaging system was used to collect flue-cured tobacco sample. The combination method of Savitzky-Golay(SG) smoothing filter, multiplicative scatter correction (MSC) and the first derivative (FD) was used to preprocess the data of flue-cured tobacco. Feature wavelength was selected through successive projections algorithm (SPA). The texture features of flue-cured tobacco selected by gray-level co-occurrence matrix (GLCM), and then the texture features and spectral features were normalized and fused. The classification effect of flue-cured tobacco was verified by using k-nearest neighbor(KNN), random forests (RF), support vector machine(SVM) and naive bayes(NB).  Results  The results showed that the classification accuracy of full-band hyperspectral data after preprocessing was improved compared with that before processing. 25 features bands were selected by SPA, and the classification accuracy using features bands was decreased. After the fusion of hyperspectral and texture features, the classification effect was improved significantly.  Conclusion  The proposed method can accurately and effectively classify flue-cured tobacco without damage through fusion of hyperspectral and texture, which provides a feasible method for quantifying the classification of flue-cured tobacco.