In order to solve the problem of information redundancy and high time complexity of hyperspectral imagery (HSI) classification in high-dimensional situation, this paper proposes a extreme learning machine method of edge preserving residual. Specifically, HSIs are segmented through fully supervised and semi supervised superpixel multi-scale guidance, while multi-scale features are fused using residual cascades. Then, the semi-supervised classification map is subjected to edge preserving filtering to obtain edge filtering full class labels, which are further processed as a new segmentation scale standard for HSIs. Finally, the fully supervised and semi supervised features are combined for mean collaborative learning to extract image edge features while strengthening the connection between prior labels. This can reduce the complexity of classifier features to a certain extent and improve classification accuracy. On the one hand, in order to comprehensively consider the impact of different scales on classification results in superpixel multi-scale segmentation, this paper introduces residual structure, which has the advantage of multi-scale features. On the other hand, considering the positive impact of edge preserving filtering on hyperspectral classification results, this paper introduces a directional filter for denoising HSIs. In order to verify the effectiveness of this model, this paper selects three data sets: Indian Pines, University of Pavia and Salinas. The results show that this optimization model has better classification performance than some traditional classification methods in the case of small sample training, and time complexity also has a certain competitiveness.