Data Augmented Graph Convolutional Network for Hyperspectral Image Classification

Labeled hyperspectral images (HSIs) data is hard to access, which becomes a great difficulty for the classification task. Graph convolutional networks can efficiently process labeled and unlabeled data via a semi-supervised fashion. To further strengthen the model classification performance, we propose a data augmented graph convolutional network (DAGCN) method. First, we use an efficient graph convolutional network to collect and extract spectral-spatial data. Then, we utilize spatial sample random reset (SSRR) method to extend spectral-spatial data with better use of abundant spatial information. Finally, we adopt the broad learning network to strengthen the width expansion of the data. Experiments prove that DAGCN outperforms the contrast methods.