基于近红外高光谱成像及信息融合的小麦品种分类研究

高光谱成像技术因具有图谱合一的特点在作物品种鉴别方面具有较大潜力，但目前研究大多只提取利用了光谱信息，对图像信息没有进行有效利用。本文利用近红外高光谱成像仪采集了强筋、中筋、弱筋3个类型共计6个品种的单粒小麦种子高光谱图像，提取了长、宽、矩形度、圆形度、离心率等12个形态特征，并对图像中的胚乳和胚区域进行分割建立掩膜，提取了胚乳和胚区域的平均光谱信息。采用PLSDA和LSSVM方法建立基于图像信息的判别模型，结果表明强筋、弱筋两者二分类的识别率能达到98%以上，强筋、中筋两者二分类的识别率只能达到74.22%，说明近红外高光谱图像的形态信息能够反映品种间差异，但单独利用图像信息进行分类时准确度可能欠佳。采用SIMCA，PLSDA和LSSVM方法建立了胚乳和胚区域光谱信息的多分类模型，胚乳区域的分类效果较胚区域略好，说明籽粒不同部位的形状差异会影响分类效果。进一步融合光谱信息和图像信息，采用SIMCA，PLSDA和LSSVM方法建立融合模型，识别率较单独的图像或光谱信息模型均略有提升，PLSDA方法从原来的96.67%提升到98.89%, 表明充分挖掘高光谱图像所包含的形态特征和光谱特征可有效提高分类效果。

The Classification of Wheat Varieties Based on Near Infrared Hyperspectral Imaging and Information Fusion

Hyperspectral imaging technology has great potential in the identification of crop varieties because it contains both image information and spectral information for the object. But so far most studies only used the spectral information, the image information has not been effectively utilized. In this study, hyperspectral images of single seed of three types including strong gluten wheat, medium gluten wheat, and weak gluten wheat were collected by near infrared hyperspectra imager, 12 morphological characteristics such as length, width, rectangularity, circularity and eccentricity were extracted, the average spectra of endosperm and embryo were acquired by the mask which was created by image segmentation. Partial least squares discriminant analysis (PLADA) and least squares support vector machine (LSSVM) were used to construct the classification model with image information, results showed that the binary classification accuracy between strong gluten wheat and weak gluten wheat could achieve 98%, for strong gluten wheat and medium gluten wheat, it was only 74.22%, which indicated that hyperspectral images could reflect the differences of varieties, but the accuracy might be poor when recognizing the varieties just by image information. Soft independent modeling of class analogy (SIMCA), PLSDA and LSSVM were used to established the classification model with spectral information, the classification effect of endosperm is slightly better than the embryo, it demonstrated that the grain shape could influence the classification accuracy. Then, we fused the spectral and image information, SIMCA, PLSDA and LSSVM were used to established the identification model, the fusion model showed better performance than the individual image model and spectral model, the classification accuracy which used the PLSDA raise from 96.67% to 98.89%, it showed that digging the morphological and spectral characteristics of the hyperspectral image could effectively improve the classification effect.