论文部分内容阅读
讨论了基于高光谱成像技术光谱及纹理特征在识别早期柑橘黄龙病中的应用。使用一套近地高光谱成像系统采集了176枚柑橘叶片的高光谱图像作为实验样品,其中健康叶片60枚,黄龙病叶片60枚,缺锌叶片56枚。手工选取每幅叶片高光谱图像的病斑位置作为样品感兴趣区域(regions of interest,ROI),计算其平均光谱反射率,并以此作为样品的反射光谱,光谱范围为396~1 010nm。样品光谱分别经过主成分分析(PXA)及连续投影算法(SPA)进行数据降维,再结合最小二乘支持向量机(LS-SVM)分类器建立分类模型。相比原始光谱,由PCA选取的前四个主成分及SPA选取的一组最佳波长组合(630.4,679.4,749.4和899.9 nm)建立的模型拥有更好的分类识别能力,其对三类柑橘叶片平均预测准确率分别为89.7%和87.4%。同时,从被选四个波长的每幅灰度图像中提取6个灰度直方图的纹理特征以及9个灰度共生矩阵的纹理特征再次构建分类模型。经SPA优选的10个纹理特征值进一步提高了分类效果,对三类柑橘叶片的识别正确率达到了100%,93.3%和92.9%。实验结果表明,同时包含光谱信息及空间纹理信息的高光谱图像在柑橘黄龙病的识别中显示了很大的潜力。
The application of spectral and texture features based on hyperspectral imaging in the identification of early citrus Huanglongbing was discussed. A set of near-surface hyperspectral imaging system was used to collect 176 pieces of hyperspectral images of citrus leaves, including 60 healthy leaves, 60 leaves of Huanglongbing and 56 zinc deficient leaves. The location of the lesion in each leaf hyperspectral image was manually selected as the regions of interest (ROI), and the average spectral reflectance was calculated and used as the reflectance spectrum of the sample with a spectral range of 396-1 010 nm. The spectrum of the sample was dimensionally reduced by principal component analysis (PXA) and continuous projection algorithm (SPA) respectively, and the classification model was established by LS-SVM classifier. Compared with the original spectrum, the model established by the first four principal components selected by PCA and the best combination of wavelengths selected by SPA (630.4, 679.4, 749.4 and 899.9 nm) has better classification and recognition ability, The average prediction accuracy of leaves was 89.7% and 87.4% respectively. At the same time, the texture features of six gray histograms and the texture features of nine gray-level co-occurrence matrices were extracted from each gray image of four selected wavelengths to construct the classification model again. The 10 texture feature values optimized by SPA further improved the classification effect, and the recognition rates of the three types of citrus leaves reached 100%, 93.3% and 92.9%. The experimental results show that hyperspectral images containing both spectral information and spatial texture information show great potential in citrus Huanglongbing identification.