Normalized cut segmentation with edge constraint for high resolution remote sensing imagery

In this paper, a framework of object-based classification with Normalized Cut segmentation method, combined with edge information, is presented for high spatial resolution images. Normalized Cut, which is a useful segmentation method for natural images, also performs well in high resolution images if affinity measurement is carefully chosen. Taking the characteristics of abundant geometric information for high resolution images into consideration, the combined affinity model excels the spectral-based and edge-based ones. Furthermore, the majority voting strategy is employed for segmentation map with a pixel-based classification result of support vector machine (SVM). Compared with watershed transform segmentation, the experimental results show better stability and effectiveness of the proposed method.