Revisiting Graph Convolutional Networks with Mini-Batch Sampling for Hyperspectral Image Classification

Abstract

Graph convolutional networks (GCNs) have been success-fully and widely applied in computer vision and machinelearning fields. As a powerful tool, GCNs have recentlyreceived increasing attention in the remote sensing commu-nity, e.g., hyperspectral image (HSI) classification. However,the application ability of GCNs in identifying the materi-als via spectral signatures remains limited, since traditionalGCNs fail to extract node features for large-scale graphs ef-ficiently. Also, simultaneous consideration of all samplesin GCNs tends to obtain poor representations, possibly dueto the vanishing gradient problem. To this end, we in thispaper develop a novel mini-batch GCN (miniGCN) for HSimage classification. More importantly, miniGCN not onlycan effectively train the network via mini-batch sampling ina supervised way, but also directly infer new samples (out-of-sample) without re-training GCNs. Experiments conductedon two commonly-used HSI datasets demonstrate the superi-ority of miniGCN over other state-of-the-art network archi-tectures.