Feature Fusion: Graph Attention Network and CNN Combing for Hyperspectral Image Classification

Abstract

Graph convolutional networks (GCNs) have attracted increasing attention in hyperspectral image classification. However, most of the available GCN-based HSI classification methods treat superpixels as graph nodes, ignoring pixel-level spectral spatial features. In this paper, we propose a novel Feature Fusion Network (FFGCN), which is composed of two different convolutional networks, namely Graph Attention Network (GAT) and Convolutional Neural Network (CNN). Among them, superpixel-based GAT can deal with the problem of labeled deficiency and extract spatial features from HSI. Attention-based multi-scale CNN can extract multi-scale pixel local features for HSI classification. Finally, the features of the two neural network models are fused and used for classification. Rigorous experiments on two real HSI datasets show that FFGCN achieves better experimental results and is competitive with other state-of-the-art methods.