Global Prototypical Network for Few-Shot Hyperspectral Image Classification

Abstract

This article proposes a global prototypical network (GPN) to solve the problem of hyperspectral image classification using limited supervised samples (i.e., few-shot problem). In the proposed method, a strategy of global representation learning is adopted to train a network (fθ) to transfer the samples from the original data space to an embedding-feature space. In the new feature space, a vector called global prototypical representation for each class is learned. In terms of the network (fθ), we designed an architecture of a deep network consisting of a dense convolutional network and the spectral-spatial attention network. For the classification, the similarities between the unclassified samples and the global prototypical representation of each class are evaluated and the classification is finished by nearest neighbor classifier. Several public hyperspectral images were utilized to verify the proposed GPN. The results showed that the proposed GPN obtained the better overall accuracy compared with existing methods. In addition, the time expenditure of the proposed GPN was similar with several existing popular methods. In conclusion, the proposed GPN in this article is state-of-the-art for solving the problem of hyperspectral image classification using limited supervised samples.

Keywords

• Deep Learning
• dense convolution
• global representations
• hyperspectral image classification
• small number of samples
• spectral–spatial attention