地面成像光谱数据兼具高光谱分辨率与高空间分辨率,在田间杂草识别中具有很好的应用前景。目前基于机器视觉的杂草识别方法以形状特征为主,当作物杂草形态相似时识别的困难和利用高光谱特征以像元为单元识别时效率较低,不利于实时自动化除草,因此,本文提出一种综合面向对象与高光谱特征匹配的杂草识别方法,在对作物杂草对象样本的形状特征和光谱曲线提取分析的基础上,建立基于形状特征规则与光谱角匹配的植物对象识别决策树,用于识别实验田中的作物杂草对象。实验结果表明,当场景中某些不同种类植物对象的形态相似时,基于形状特征规则与光谱角匹配的杂草识别方法可借助高光谱特征精细区分植物对象的种类,且在形状特征规则约束下使用高光谱特征匹配法识别植物对象,可克服"同物异谱"和"同谱异物"现象带来的不确定性,该方法识别精度可优于仅使用光谱角匹配法的情况,并优于使用颜色和形状分析技术的情况。

Weed identification is a basic task in precision agriculture, as well as in the principle of variable spraying and accurate weeding. Field imaging spectrometer data, with both hyperspectral and high spatial resolutions, have potential applications in weed identification. Currently, methods for weed identification include considering shape features based on machine vision, which perform poorly when weeds and crops have similar shape features, while the use of hyperspectral features have low efficiency when it is identified in pixels. To overcome the limitations of these existing methods, a new weed identification method combining the strengths of both object-oriented and spectral feature matching approaches is proposed. The proposed method extracts and analyzes the shape features and the spectral curves of plant object samples, and builds a decision tree using shape feature rules and spectral angle mapper to identify the plant objects in the experimental field. The results show that the proposed method could identify different kinds of plant objects with similar shape features by using hyperspectral features, and could overcome the difficulties in identifying same objects with different spectra and different objects with the same spectrum by using shape feature rules. The identification accuracy of the described method is higher than both the spectral angle mapper method and the color and shape analysis.