Abstract: Hyperspectral image (HSI) classification methods based on deep learning usually slice hyperspectral images into local-patches as the input of the model, which not only limits the acquisition of long-distance space-spectral information association, but also brings a lot of extra computational overhead. The image-level classification method with global image as input can effectively avoid these defects. However, the detail loss during information recovery of the existing image-level classification methods based on feature serial flow pattern of fully convolutional network (FCN) will lead to problems such as low classification accuracy and poor visual effect of the classification map. Therefore, this paper proposes a high-resolution feature network (HRNet) image-level classification method for hyperspectral image, which performs parallel computation and cross fusion of multi-resolution features of images while maintaining high-resolution features throughout the whole process, thus alleviating the information loss caused by the traditional serial flow pattern of features. Simultaneously, we propose a jointly-supervised training strategy of multi-resolution feature and a vote classification strategy, so as to further improve the classification performance of the model. Four public hyperspectral image datasets are used to verify the proposed method. Experimental results show that compared with the existing advanced classification methods, the proposed method can obtain competitive classification results, significantly reduce the training and classification time at the same time, and is more time-sensitive in practical application. In order to assure the reproducibility of method, we will open the code at https://github.com/sssssyf/fast-image-level-vote.

Key words: hyperspectral image classification, image-level, fully convolution network, HRNet