Abstract: Abstract: Weeds are distributed only in patches in fields, but herbicides are applied over entire fields, thus leading to over application and unnecessary pollution. To reduce herbicide application, automatic weed recognition is being developed to treat only weed patches. Weed identification is the key point of automatic weeding, and many research studies have pointed out that the reflectance rate of green plant leaves could be used to identify the varieties. The spectral reflectance of winter rape, soil (dry and wet), and five kinds of weeds (speedwell, thistle, capsella, horseweed, and cerastium viscosum) were measured within the 350-2 500 nm wavelength range by the Analytical Spectral Device (ASD) in a laboratory. Each sample was measured five times continuously, and 370×5 samples were obtained. After rejecting 2×5 samples, a total of 368×5 samples (a 278×5 training set and a 90×5 prediction set) were used for classification, and a training set and a prediction set were randomly selected. The five original spectroscopic data sets were averaged in order to eliminate random noise. First, derivative and compressing were used to pretreat the spectral data. Then, stepwise discriminant analysis was executed to reduce the redundancy spectral information and decrease the amount of calculation and improve the accuracy. Four characteristic wavelengths, 710, 755, 950, and 595 nm were selected. Then canonical discriminant analysis and Bayes discriminant analysis were applied to build recognition models for identifying these weeds, soil, and winter rape based on the four characteristic wavelengths. For the canonical discriminant function, the recognition accuracy of training was 97.84%, two miscalculations were occurred in weeds, and the recognition ratio was 97.78%. The classification accuracy of the Bayesian discriminant model was higher than the canonical discriminant model, one error was observed in the testing set and its recognition ratio was 98.89% on the condition that prior probabilities were computed from group sizes. However, its classification accuracy is the same as canonical discriminant models when prior probabilities were all equal groups. The result of the statistical analysis showed that it was feasible to use the four characteristic wavelengths as input variables to build recognition models, which was an effective approach to simplify the recognition models. Both of the canonical discriminant function and Bayesian discriminant model have a strong ability to differentiate spectra of species of plant. However, the classification accuracy of the Bayesian discriminant model was higher when prior probabilities were computed from group sizes. The research results can provide reference for the development of spectrum sensor.