基于支持向量机的玉米苗期田间杂草光谱识别

田间全面积均匀喷施除草剂不经济，还污染环境，精准喷施除草剂意义重大，其关键是正确识别杂草。用便携式野外光谱仪，在田间测量了玉米、马唐和稗草植株冠层在350～2 500 nm波长范围内的光谱数据，经过数据预处理，数据分析波长选为350～1 300和1 400～1 800 nm。数据处理采用支持向量机(SVM)模式识别方法。SVM具有可实现对小样本建模结构风险最小化、结果最优化、泛化能力强的优点。用线性、多项式、径向基和多层感知核函数对玉米和杂草建立二分类模型，结果表明，三阶多项式核函数SVM分类模型的正确识别率最高，达到80%以上，且支持向量比例较小。以二分类模型为基础，利用投票机制，建立了玉米、马唐和稗草的一对一多分类SVM模型，正确识别率达80%。田间光谱测量受光照、背景和仪器测量精度等条件的影响较大，但结果仍表明SVM结合光谱技术在田间杂草识别中应用潜力很大，此研究为田间杂草识别及传感器的建立提供了一种研究思路和应用基础。

SVM-Based Spectral Recognition of Corn and Weeds at Seedling Stage in Fields

A handheld FieldSpec 3 Spectroradiometer manufactured by ASD Incorporated Company in USA was used to measure the spectroscopic data of canopies of seedling corns, Dchinochloa crasgalli, and Echinochloa crusgalli weeds within the 350-2 500 nm wavelength range in the field. Each canopy was measured five times continuously. The five original spectroscopic data were averaged over the whole wavelength range in order to eliminate random noise. Then the averaged original data were converted into reflectance data, and the unsmooth parts of reflectance spectral curves with large noise were removed. The effective wavelength range for spectral data process was selected as 350-1 300 and 1 400-1 800 nm. Support vector machine (SVM) was chosen as a method of pattern recognition in this paper. SVM has the advantages of solving the problem of small sample size, being able to reach a global optimization, minimization of structure risk, and having higher generalization capability. Two classes of classifier SVM models were built up respectively using “linear”, “polynomial”, “RBF”(radial basis function), and “mlp (multilayer perception)” kernels. Comparison of different kernel functions for SVM shows that higher precision can be obtained by using “polynomial” kernel function with 3 orders. The accuracy can be above 80%, but the SV ratio is relatively low. On the basis of two-class classification model, taking use of voting procedure, a model based on one-against-one-algorithm multi-class classification SVM was set up. The accuracy reaches 80%. Although the recognition accuracy of the model based on SVM algorithm is not above 90%, the authors still think that the research on weeds recognition using spectrum technology combining SVM method discussed in this paper is tremendously significant. Because the data used in this study were measured over plant canopies outdoor in the field, the measurement is affected by illumination intensity, soil background, atmosphere temperature and instrument accuracy. This method proposes a kind of research ideology and application foundation for weeds recognition in the field.