DSR-GCN: Differentiated-Scale Restricted Graph Convolutional Network for Few-Shot Hyperspectral Image Classification

Graph convolutional networks (GCNs) have shown great potential for few-shot hyperspectral image (HSI) classification. Mainstream GCNs construct graphs according to single-scale segmentation, which usually ignores subtle adjacency relations between small regions, leading to an unreliable initial local graph. To overcome the above issue, we propose a differentiated-scale restricted GCN (DSR-GCN) for HSI classification. First, we propose a differentiated-scale graph construction method considering both the subtle and relative wider range spectral–spatial relation. Second, restricted fusion loss is designed to restrict the fusion of features extracted with differentiated-scale GCN branches. Finally, we design a lightweight spectral-spatial siamese network (S3Net) to remedy local pixel-level features. The proposed DSR-GCN can better model spatial structure with a reliable and refined graph, and it can capture more discriminate features in few-shot learning (FSL) scenarios. Extensive experiments conducted on four benchmark datasets demonstrate that DSR-GCN outperforms the other deep learning methods in terms of classification accuracy and generalization performance, with improvements in terms of overall accuracy (OA) around 6.20%
$\sim $
23.41% (Indian Pines), 4.45%
$\sim $
16.48% (University of Pavia), 4.25%
$\sim $
11.85% (Salinas), and 2.0%
$\sim $
17.23% (University of Houston) under five labeled samples per class.