Multilayer non-negative matrix factorization （MLNMF） can not fully use the spatial-spectral features of hyperspectral remote sensing images， and the ubiquitous noise in hyperspectral images. To solve the problem， this paper proposed a new spatial-spectral reweighted sparse MNMF unmixing algorithm. Firstly， the subspace clustering algorithm was used to construct spatial weights according to the spatial characteristics of hyperspectral images. Secondly， the superpixel segmentation algorithm was used to segment the hyperspectral image， and the similarity between superpixels was calculated. The KMEANS++ algorithm was used to cluster the superpixels， and then the pixel-level similarity was calculated in the superpixel to construct the spectral weight. The spatial weight and spectral weight were fused， and the fused spatial-spectral weight was used to characterize the spatial-spectral information of the hyperspectral image. Then， the SUnSAL algorithm was used to calculate the sparse noise reduction weight， which can effectively reduce the influence of noise on the unmixing performance. Finally， the norm was used to constrain the endmembers and abundance of the model to improve the unmixing performance of the model. Compared with the experimental results of five unmixing algorithms， the mean Spectral Angle Distance and Root Mean Square Error of the proposed algorithm on the synthetic dataset were optimal. It also achieves good unmixing results on two real datasets Jasper Ridge and Cuprite. The endmember estimation error of the proposed algorithm is reduced by 1.49% to 4.68%， and the abundance estimation error is reduced by 1.83% to 4.18% on each dataset.