Abstract: Large-scale pesticide spraying will raise costs in agriculture and will cause environmental pollution. In order to realize quantitative and directional spraying, weed identification using image-processing technology is one of the focus problems in the precision-agriculture field. The foundation of automated identification is feature extraction. Because the dimensions of the feature are usually very high, before identification the dimensions must be reduced. The performance of any dimension-reduction method will directly affect the recognition results. The traditional dimension reduction method is a linear method, so it is very difficult to grasp the nonlinear nature of the original data. Locally linear embedding (LLE) is one kind of emerging manifold method. Compared with the traditional method, it is a nonlinear method, but it still has some limitations. The locally linear embedding is essentially an unsupervised method, and it cannot utilize the category information of the train samples. The traditional locally linear embedding method has defects in dealing with the classification problem, so the recognition accuracy is not satisfactory. In order to overcome this defects above, a supervised locally linear embedding method based on Fisher projection (FS-LLE) was chosen to reduce the feature dimension. The samples are projected by Fisher transformation first. Then, the projection coordinates are obtained, and projection distances are computed. The Euclidean distance is used to select the sample points' neighbors in the traditional locally linear embedding. Compared with the Euclidean distance, the Fisher projection distances can characterize the category attributes between different types of samples, so it is chosen to determine the neighborhood structure. In order to verify the effectiveness of this dimension reduction method, the experiment is designed as follows. The weed images are collected from the field, and the grayscale features are obtained. The influence of land background is excluded by super green characteristics at first, then the individual green plant region is selected by a morphological method. The gray scale data ensemble of each region is the feature whose dimension will be reduced. For comparison, dimension reduction was accomplished by PCA (principal component analysis), LLE (locally linear embedding), WLLE (weighted locally linear embedding) and FS-LLE (supervised locally linear embedding based on the Fisher projection), respectively. Through second visual dimension reduction results, it can be found clearly that the FS-LLE method achieves better low-dimensional data-clustering effects. This illustrates that the method proposed in this article finds the samples' intrinsic class features at the same time it reduces feature dimension. This virtue is more conducive to solve identification problems. At last the task of classification for recognition is fulfilled by the support vector machine. Further comparison and analysis of the identification results obtained by different dimension reduction methods, the average recognition rate of corn of this method reaches 97.2%, while the average recognition rates of PCA, LLE, and WLLE were 70.1%、77.1% and 86.8%, respectively. In the same way, the average recognition rate of weed of this method reaches 77.8%, while the average recognition rates of PCA, LLE, and WLLE were 38.4%、61.2% and 60.4%, respectively. Therefore, the method proposed by this article is a significant improvement over the other traditional methods. This result also shows that the method has certain advantages in solving the classification problem.