Remote sensing image classification method based on superpixel segmentation and adaptive weighting K-Means

To improve remote sensing image classification precision, we propose a novel method which is based on superpixel and adaptive weighted K-Means. First, superpixel segmentation algorithm is used to divide input images into irregular blocks which remain their semantic information and boundaries. And then, SIFT, GIST, Census, Gabor, and Color histogram, and many other types of features are extracted. These five features represent different kinds of image characteristics. In order to obtain best feature combination, abundant experiments and analyses are performed. We also propose a novel adaptive weighted k-Means method to automatically estimate optimal weights and cluster centers for improving the representation accuracy of visual bag-of-words. An improved soft vector quantization based on sparse coding is adopted to generate the image feature. Finally support vector machine is utilized to complete the image classification. Experiments show that the new method proposed in this paper can effectively improve the classification accuracy of remote sensing images, and it also show good stability and robustness.