Abstract: The detection accuracy of traditional pixel-level change detection algorithms is seriously influenced by radiometric difference, registration error and the determination of classification threshold for a different image, and it is difficult to differentiate the true changes of interest from various kinds of detected changes. Therefore, a novel two-step change detection algorithm combining feature-level and pixel-level techniques is proposed to detect changes of man-made objects in multi-temporal remote sensing images. Large-size images are divided into overlapping sub-images, and the changed regions containing man-made objects are extracted by supervised sub-image classification. Then, a pixel-level change detection algorithm is developed to obtain quantitative detection results. Experimental results demonstrate the feasibility and effectiveness of the proposed algorithm.