SVM-Based Spectral Recognition of Corn and Weeds at Seedling Stage in Fields
Deng Wei,Zhang Lu-da,He Xiong-kui,J. Mueller,Zeng Ai-jun,Song Jian-li,Liu Ya-jia,Zhou Ji-zhong,Chen Ji,Wang Xu
DOI: https://doi.org/10.3964/j.issn.1000-0593(2009)07-1906-05
2009-01-01
Abstract: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.