Hyperspectral and SAR Image Classification via Graph Convolutional Fusion Network

Abstract

Hyperspectral and synthetic aperture radar (SAR) image classification, aiming to merge multisource information to boost the precision and reliability of land cover classification, has gained increasing attention. Nevertheless, current techniques still exhibit certain limitations in extracting discriminative features and integrating heterogeneous features. In this work, a graph convolutional fusion network (GCFNet) is proposed for hyperspectral and SAR image classification. First, a spectral residual neural network is employed to extract the spectrum information. Then, a dual-branch graph convolutional network (GCN) is developed to extract the spatial information from hyperspectral and SAR images. Finally, a cross-contextual transformer fusion module is created to merge the spectral and spatial information followed by a dense layer to yield the final prediction outcome. To confirm the performance of the GCFNet, experiments on three datasets (e.g., Berlin, Augsburg, and Yellow River) demonstrate that the GCFNet significantly surpasses other representative methods.