Aiming at the problem of poor image classification accuracy caused by low signal-to-noise ratio of hyperspectral images， a hyperspectral image classification method that combines multi-scale low-rank representation and two way recursive filtering is proposed. First， perform superpixel segmentation algorithm on hyperspectral images at different scales to obtain the spatial neighborhood information and segmented images. Next， low-rank representation and PCA（Principal Component Analysis） dimensionality reduction are performed in the segmented regions of each scale， the low-rank representation can impose low-rank constraints on the high correlation between spectra in the segmented regions and remove mixed noise. Then， two way recursive filtering is used to further eliminate noise in the image. Last， according to the classification results of the feature images of each scale by the Support Vector Machine， the final classification is obtained by the majority voting method. The results showed that： Compared with the classification methods using only spectral information （Support Vector Machine and PCA）， the overall accuracy of the proposed method is improved by 32.03%， 28.04% and 16.80% on average. Compared with the deep learning classification method of spatial-spectral residual network and vertex component analysis network， the average improvement is 10.99%， 8.45% and 7.08%. Compared with other spatial-spectral classification methods， the average improvement is 8.28%， 18.77% and 10.19%， it is proved that the proposed method can achieve better overall classification accuracy with fewer training samples.