Multistage Superpixel-Guided Hyperspectral Image Classification With Sparse Graph Attention Networks
Weiming Li, Qikang Liu, Shuaishuai Fan, Hongyang Bai, Mingrui Xin
Abstract
Graph Neural Networks (GNNs), as an effective learning framework for graph structure data representation, have been applied to hyperspectral images (HSIs) classification tasks. Among the variants of GNNs, Graph Attention Networks (GATs) have achieved state-of-the-art node prediction performance by learning to assign dense attention coefficients to all node neighbors for feature aggregation. However, due to the complexity of land distribution and the high dimension of HSIs data, it is difficult to identify different coverage categories by employing GATs directly, and the inadequacy of spectral information also affects the classification accuracy. To address the above application problems, we proposed a multi-stage superpixel-guided sparse GAT for HSI classification (named MSG-SGAT). Specifically, we create the adjacency connection graphs of different stages from the superpixel representation so as to effectively utilize the spatial topology. A sparse GAT (SGAT) module is designed to trim the graph by spectral sparsity to remove some task-irrelevant edges and assign a unique attention coefficient to each remaining edge. Sparse subgraphs are obtained by identifying the noise/task-independent edges and aggregating information. Moreover, SGAT is concatenated with the spectral branch for feature fusion and update, and pixel-level feature refinement is performed at the end of this network. MSG-SGAT mines the features of HSIs from the perspective of multi-scale hierarchy with low calculation cost and high efficiency. A series of comparative experiments on four benchmark datasets demonstrate that the performance of our proposed method is competitive with others.