Subspace clustering has become increasingly popular in recent years and has shown great success in hyperspectral band selection (BS). However, traditional subspace clustering model and its variants are inadequate in expressing the fine spatial structure information and long-range correlations of the samples. Therefore, this paper proposes a latent feature representation-based low rank subspace clustering model for BS. It employs entropy rate superpixel segmentation to obtain the spatial structure of the image. Then, it extracts the key latent features of samples in each region by graph learning to jointly represent the original image, which can maximize the retention of key information while reducing noise and data dimensionality. Additionally, considering the short-range and long-range correlations of samples, a sample-spatial structure constraint is constructed to enhance the spatial relationship of similar samples between homogeneous and heterogeneous regions, and rectify the errors in sample feature caused by the inaccurate segmentation. This is advantageous for the subsequent clustering and BS. The effectiveness and stability of this method are confirmed by experiments on three datasets.