Object-based multispectral image segmentation and classification

In this paper an object-based method for multispectral image segmentation and classification is proposed. Normally, in remote sensing a scene is represented by pixel-based features. It is possible to reduce data redundancy by a segmentfeature extraction process, where the segment-features, rather than the pixel-features, are used for multispectral scene representation and classification. Object-based algorithms partition the observation space into a set of disjoint segments (called objects). Then, pixels belonging to each segment are represented by segment features. In this paper, an unsupervised segmentation algorithm based on statistical region merging (SRM) framework is presented. Also a partial differential equations (PDE) algorithm is suggested as a preprocessing phase to improve the segmentation results. Illustrative examples are presented, and the performance of the extracted object features for classification purposes is investigated. Results show significant classification performance improvement.