In the standard spectrum clustering algorithm, the metric based on Euclidean space cannot represent the complicate space distribution feature of some data set, which might lead to the clustering result inaccuracy. While the geometric relationship between data can be described more precise by manifold space. The special expression on curved surface is researched, the feature which is more fit for measuring the distance between data is applied, and an improved spectrum clustering analysis algorithm based on the distance metric under Graasmann manifold is proposed. The similarity between data is analyzed under manifold space. The experimental results show that the proposed algorithm can cluster data set either belonging the same or different subspace more accurately, furthermore, it can cluster data set with more complicate geometric structure under manifold space efficiently.