Abstract: The classification of hyperspectral images faces challenges like ineffective extraction of multi-scale features and easy loss of pose information. Considering these challenges, this study proposed a classification network of hyperspectral images with multi-scale feature fusion-the hierarchical multi-scale concatenation net (HMC-Net). Initially, multi-scale convolution kernels were applied for parallel computing to extract multi-level features. Meanwhile, the 1×1 convolutional kernels were employed to reduce input-output dimensions, balancing computational complexity. These operations enabled efficient feature extraction without significantly increasing the overall computational burden. Subsequently, independent capsule networks were used for parallel processing of features at various scales. The max pooling was improved via dynamic routing to enhance the translation invariance of features, thereby reducing the loss of pose information. Finally, the concatenate operation integrated feature maps of different scales, thereby achieving a precise analysis of multi-level information in the classification of hyperspectral images. Comparative experimental results demonstrate that the HMC-Net achieved an overall accuracy of 94%, 98%, and 99% on the Kennedy Space Center, University of Pavia, and Salinas datasets, respectively. Compared to the latest classification model of hyperspectral images, the HMC-Net exhibited significant performance advantages, validating its effectiveness.