In volume rendering and image segmentation, data set often possesses manifold structure. Different parts in such data sets are closely adjacent to each other and local noises exist around the boundaries, which bring great difficulty to traditional clustering algorithms. According to the non-parameter density estimation, this paper proposes a hierarchical clustering algorithm based on multi-scale information fusion. The new algorithm integrates density differences and boundary information to define a kind of similarity measurement based on multi-scale information fusion. With the graph connection of level sets, the new approach obtains a hierarchical analyzing process of the cluster structures, which outputs stable clustering results. The new algorithm is not restricted by shapes and density structures of the data set, and can detect common structural features of the data set automatically without assumption. Meanwhile, the clustering results are stable and the new algorithm is strongly robust to noises. The superiority of the proposed algorithm is demonstrated with its applications to synthetic and real data sets.