Abstract: The low spatial resolution characteristics of hyperspectral images often make it difficult for global texture extraction techniques to obtain accurate texture information and the single-scale local texture extraction technology is not satisfactory for effectively identifying the features. In this article, a Multi-scale Superpixel Texture Preservation and Fusion is proposed for hyperspectral image classification. Specifically, the original hyperspectral image is first extracted with multi-direction and scale global texture using 2D Gabor filter, and the texture feature of each scale is merged to enhance the texture structure characterization ability. Next, texture and spectral principal component features are fused to form spectral-texture joint discriminant features. After that, the shape adaptive oversegmentation method is applied to the spectral-texture joint feature for local texture information preservation and fusion. In particular, in order to overcome the hidden irrelevance problem of neighboring pixels, a density-based nearest neighbor similarity evaluation criterion is defined, which aims to make the superpixel texture more consistent. Finally, the updated spectral-texture joint discriminant features are input into the pixel-level classifiers to obtain their corresponding class labels, and the decision fusion mechanism of majority voting is adopted to obtain the final classification result. Experiments on the real data sets of Indian Pines and Pavia University show that the classification accuracy of this method under the condition of small samples is better than eight comparison methods such as the benchmark classifier Support Vector Machine (SVM), deep learning method Gabor Filtering and Deep Network (GFDN), and the latest spatial-spectral method Spectral-Spatial and Superpixelwise Principal Component Analysis (S3-PCA), which proves fully the practicability and effectiveness of the proposed method.