.png){kind=icon}
.png){kind=icon}
Return
Adaptive Convolutional Interaction-Guided Graph Attention Network for Hyperspectral Image Classification
.png){kind=icon}
PRE
AI
T
Y
R
F
L
DOI：10.1109/taes.2026.3711614.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Hyperspectral image (HSI) classification is a challenging task due to the high spatial–spectral complexity property of HSI, especially when the available training samples are scarce. Existing graph neural network (GNN) and convolutional neural network (CNN)-based methods have shown some advantages for the task, but they usually suffer from the limitations in fully revealing the spatial–spectral as well as global–local structural characteristics of HSI and exploring discriminate features for classification. For these issues, this article has contributed an adaptive convolutional interaction-aware graph attention network (AI-GAT). The network adopts a divide-and-conquer strategy, and the discriminate features are learned and refined via the collaboration of pixel-level (spectral information) and superpixel-level (joint of spatial and spectral information) feature learning. Specifically, an interaction space module is devised to interact the disjoint regions for revealing and enhancing the global long-range dependency properties of HSI, which benefits pixel-level and superpixel-level feature learning. Second, a global-aware space module and an extended graph attention network are developed, to achieve the graph node feature by generating diverse superpixel features. In addition, we have devised a pixel-level feature enhancement block with an adaptive convolution module, to cooperate with the superpixel feature flow for refining classification results. Extensive experiments on real-world datasets demonstrate that the proposed AI-GAT is competitive in both qualitative and quantitative evaluations compared with numerous state-of-the-art methods.
Keywords:
Graph attention network (GAT)
hyperspectral image (HSI)
land-cover classification
spatial–spectral
superpixel
Journal
IF:
5.7
Papers:
651
Citations:
2.4W
Organization
Cited Papers