An agent graph-attention–enhanced transformer for semi-supervised hyperspectral image classification

Abstract

Hyperspectral image classification (HSIC) is a key topic in remote sensing research, but acquiring a sufficient quantity of high-quality labeled samples is costly and, in some cases, unattainable. We propose a Semi-Supervised classification network that integrates Agent Graph Attention with a Transformer (AGAT-SS) to fully exploit both labelled and unlabelled samples to improve HSIC performance. The network is composed of three core components: a Feature Alignment Module (FAM), an Agent Graph Attention Network (A-GAT), and an Agent-Enhanced Feed-forward Transformer (AEF-Transformer). FAM employs channel attention and multi-scale convolutions with the objective of enhancing the consistency between labelled and unlabelled data. This process establishes a reliable foundation for subsequent feature extraction. A-GAT introduces an agent-attention mechanism that jointly captures global and local features while markedly reducing computational complexity, yielding efficient and robust feature learning. AEF-Transformer combines Agent Attention with an Enhanced Feed-forward Module (AEFM), thereby substantially strengthening feature-extraction capacity and model expressiveness. Extensive experiments on the public Indian Pines (IP), Pavia University (PU) and Houston 2013 (HU13) datasets indicate that AGAT-SS significantly outperforms excellent algorithms such as Multiscale Dynamic Graph Convolutional Network (MDGCN), Multiscale Spectral–Spatial GAT (MSSGAT), and Dynamic Evolution GAT (DEGAT). In particular, on the IP dataset, when only 5 % of the labeled samples and 20 % of the unlabeled samples were used for training, AGAT-SS outperformed DEGAT by 1.16 % in Overall Accuracy (OA), 1.36 % in Average Accuracy (AA), and 0.91 % in Kappa coefficient. These gains further confirm its superiority in semi-supervised learning.