To address the low accuracy and low stability of a single feature for weed recognition, a multi-feature fusion method based on support vector machine (SVM) and DS (Shafer-Dempster) evidence theory was proposed. Firstly, three types of plant leaf features such as shape, texture and fractal dimension were extracted from the plant leaves after a series of image processing. Then the SVM classification results of each single feature were used as evidences to construct the basic probability assigned (BPA), and the method of DS fusion based on matrix analysis was used for decision fusion. Finally, recognition results were given based on fusion results and classification thresholds. The experimental results showed that the accuracy of multi-feature fusion method was 96。11% which has good performance on accuracy and stability compared with the single feature method in weed recognition．