A SVM ensemble approach combining pixel-based and object-based features for the classification of high resolution remotely sensed imagery

According to the `salt and pepper´ effect of pixel-based multi-feature classification and over-smoothing of ground details of object-based image analysis, in this paper, an approach, which fuses pixel-based features and multi-scale object-based features is proposed to improve the accuracy of image classification. (1) Firstly, mean shift algorithm is used to segment the image to obtain over-segmentation regions. Multi-scale segmentation results are obtained by merging the over-segmentation results. The relation between segmentation scales and classification accuracy on each scale is analyzed, and an optimal scale is found. (2)Secondly, objects´ spectral features of the optimal scale, pixel-based spectral features and objects´ spectral features of initialization segmentation scale are normalized. (3)Finally, the classification method based on pixel-based and object-based features is implemented by means of support vector machine ( SVM ). The experiment results demonstrate that our method can not only effectively reduce the `salt and pepper´ effect of pixel-based algorithm, but also maintain the integrity of the ground objects and preserve details. The classification accuracy of categories that are easily confused (e.g. shadow vs. streets) is also improved.