Abstract: The traditional manifold learning algorithms are based on the assumption that categories of data are located in the same manifold structure;nevertheless,due to the different features of different data categories,it is more reasonable that the data are in respective different manifold structures. Hence, the assumption of multi -manifold is more applicable for data classification. This paper adopts the thought of multi - manifold spectral clustering algorithm, mainly focuses on multiple manifolds LE algorithm, and applies this algorithm to the processing of hyperspectral data. Combined with the features of the hyperspectral data, the multiple manifolds LE algorithm is further improved by adding the spatial information and data maker information. The experimental results show that,in many kinds of hyperspectral data,the multi-manifold LE algorithm has higher precision than the LE algorithm. In addition,the improved multi-manifold LE algorithm could classify data with higher precision than the LE algorithm and multi - manifold LE algorithm. The authors have reached the conclusion that the assumption of multi -manifold is in better agreement with the features of hyperspectral data and the improved algorithm is of high performance.