Abstract: To further improve the classification accuracy of hyperspectral images, a multiscale hyperspectral image classification method based on superpixel reconstruction is proposed. This method utilizes the spatial similarity between pixels to extract low-dimensional features of hyperspectral data more effectively. First, to fully capture the multiscale spatial features in the image, superpixel segmentation technology was used to extract the homogeneous region and generate multiscale segmentation results. In the superpixel region of each scale, the spatial neighborhood reconstruction model was established using adjacent pixels that belong to the same superpixel block. Then, principal component analysis based on superpixels was used to reduce the dimensions of different scale superpixel regions. In the classification process, the support vector machine is used to train the low-dimensional features of each scale, and then the majority voting mechanism is fused to make full use of the information of different scales. The experimental analysis on three real datasets proved the superiority of the proposed method.