Feature-guided dynamic graph convolutional network for wetland hyperspectral image classification

Abstract

Recently, graph convolutional networks (GCNs) has attracted wide attention on the wetland classification with limited samples. However, traditional approaches of superpixel generation rely on artificial experience and the spatial information is ignored during the construction of graph structure, which limits the classification performance. To address these problems, a feature-guided dynamic graph convolutional network (FG-DGCN) is proposed for wetland classification. First, a learnable superpixel generation module is proposed to generate adaptive superpixel boundaries, which composed of a pixel-wise feature enhancement block and a superpixel generation block. The former is utilized to improve the discrimination of features and the latter is applied to adjust the representation of superpixels by training. Second, a feature-guided adjacency matrix update mechanism is designed to dynamically capture and fuse the spectral and spatial correlations of graph nodes, promoting the aggregation of neighborhood information. Finally, the features are differentially projected back to the pixel space for wetland classification. Experiments on three wetland datasets demonstrate the superiority of FG-DGCN over state-of-the-art methods.

Keywords

• Wetlands classification
• Hyperspectral image (HSI)
• Dynamic graph convolutional networks(DGCN)
• Learnable superpixel generation
• Feature-guided update