A Novel Texture-Preceded Segmentation Algorithm for High-Resolution Imagery

Abstract

Image segmentation is crucial to object-oriented remote sensing imagery analysis. In this paper, a novel texture-preceded segmentation algorithm is proposed for high-resolution remote sensing imagery, in which texture clustering is first carried out as a loose constraint for later segmentation. The algorithm is based on the graph models of region adjacency graph and nearest neighbor graph, which can achieve fast node merging, depending on the global optimum. Here, a combined distance, composed of texture, spectral, and shape features, is established to measure the similarity between nodes and gives the same semantic descriptions for the texture objects. Then, the combined distance is applied to graph models, and the final segmentation result can be obtained iteratively by fast merging. During the merging process, optimal sequence merging interacts with texture clustering to refine the real edges of a texture region. This algorithm cannot only merge the homogeneous texture segments with spectral variability easily but can also detect the real object boundaries well. The experiments on high-resolution imagery show that, in terms of the same number of segments, the proposed algorithm can improve segmentation accuracy by 10%-20% compared to the results obtained by pure spectral features with Definiens Developer software.