An enhanced network with parallel graph node diffusion and node similarity contrastive loss for hyperspectral image classification

Abstract

Graph neural networks (GNNs) have substantially advanced hyperspectral image (HSI) classification. However, GNN-based methods encounter challenges in identifying significant discriminative features with high similarity across long distances and transmitting high-order neighborhood information. Consequently, this paper proposes an enhanced network based on parallel graph node diffusion (PGNDE) for HSI classification. Its core develops a parallel multi-scale graph attention diffusion module and a node similarity contrastive loss. Specifically, the former first constructs a multi-head attention-forward propagation (AFP) module for different scales, which incorporates multi-hop contextual information into attention calculation and diffuses information in parallel throughout the network to capture critical feature information within the HSI. Afterward, it builds an adaptive weight computation layer that collaborates with multiple parallel AFP modules, enabling the adaptive calculation of node feature weights from various AFP modules and generating desired node representations. Moreover, a node similarity contrastive loss is devised to facilitate the similarity between superpixels from the same category. Experiments with several benchmark HSI datasets validate the effectiveness of PGNDAF across existing methods.