Image segmentation is the foundation of seismic information extraction from high-resolution remote sensing images. While the complexity of the seismic image brings great challenges to its segmentation. Compared with the traditional pixel-level approaches, the region-level approaches are found prevailing in dealing with the complexity. This paper addresses the seismic image segmentation problem in a region-merging style. Starting from many over-segmented regions, the image segmentation is performed by iteratively merging the neighboring regions. In the proposed algorithm, the merging criterion and merging order are two essential issues to be emphatically considered. An effective merging criterion is largely depends on the region feature and neighbor homogeneity measure. The region’s spectral histogram represents the global feature of each region and enhances the discriminability of neighboring regions. Therefore, we utilize it to solve the merging criterion. Under a certain the merging criterion, a better performance could be obtained if the most similar regions are always ensured to be merged first, which can be transformed into a least-cost problem. Rather than predefine an order queue, we solve the order problem with a dynamic scheme. The proposed approach mainly contains three parts. Firstly, starting from the over-segmented regions, the spectral histograms are constructed to represent each region. Then, we use the homogeneity that combines the distance and shape measure to conduct the merge criterion. Finally, neighbor regions are dynamically merged following the dynamic program (DP) theory and breadth-first strategy. Experiments are conducted using the earthquake images, including collapsed buildings and seismic secondary geological disaster. The experimental results show that, the proposed method segments the seismic image more correctly.