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

Graph convolutional networks (GCNs) have been successfully and widely applied in computer vision and machine learning fields. As a powerful tool, GCNs have recently received increasing attention in the remote sensing community, e.g., hyperspectral image (HSI) classification. However, the application ability of GCNs in identifying the materials via spectral signatures remains limited, since traditional GCNs fail to extract node features for large-scale graphs efficiently. Also, simultaneous consideration of all samples in GCNs tends to obtain poor representations, possibly due to the vanishing gradient problem. To this end, we in this paper develop a novel mini-batch GCN (miniGCN) for HS image classification. More importantly, miniGCN not only can effectively train the network via mini-batch sampling in a supervised way, but also directly infer new samples (out-of-sample) without re-training GCNs. Experiments conducted on two commonly-used HSI datasets demonstrate the superiority of miniGCN over other state-of-the-art network architectures. The codes of this work are available at https://github.com/danfenghong/IEEE_TGRS_GCN for the sake of reproducibility.