Median Filter Algorithm combined with K–means clustering was employed to segment lesion area of wheat powdery mildew, stripe rust and leaf rust. Color moments and gray–level co–occurrence matrix (GLCM) were used to extract color features and texture features. Variance algorithm and sequential floating forward search (SFFS) algorithm were used for selection of optimal feature subset with which classification and recognition of the 3 kind of wheat diseases were achieved. Experiment was done based on SVM using the feature subset, and the classification accuracy was up to 99%. Compared with PCA method which classifying feature subset obtained by dimension reduction, the method used in this study could reduce the feature space and improve recognition accuracy effectively.