The invention discloses a hyperspectral image dimension reduction method, which comprises the following steps: firstly, dividing an original hyperspectral image into non-overlapped superpixels by using an over-segmentation method; next, as the pixel point in one super-pixel usually belongs to the same kind of objects, describing the spatial information by using an intra-class graph in the invention; and finally, introducing the intra-class graph based on the super-pixel level into an LGDE model as a regular item. In addition, in order to effectively capture the nonlinear characteristics of thehyperspectral image, the invention expands the linear LGDE into a kernel version. An original pixel point classification method (RAW), a principal component analysis (PCA), a linear discriminant analysis (LDA) method, a spectral space linear discriminant analysis (SSLDA) method, a local reservation projection (LPP) method, a collaborative graph-based discriminant analysis (CGDE) method, a sparsegraph-based discriminant analysis (SGDE) method, and a local graph-based discriminant analysis (LGDE) method are compared. Under the same experiment conditions, the classification result of the methodis more accurate.