Hyperspectral Image Classification via Dynamic Adaptive Graph Convolutional Network

Dynamic Graph Convolutional Network (DGCN) can represent temporal evolutionary features. Its compatibility with the spectral-dimensional characteristics of hyperspectral images (HSI), such as continuous gradients and local abrupt changes in spectral signatures, makes it valuable in this field. This study introduces dynamic graph modeling into HSI analysis. By mapping spectral dimensions onto virtual time series, we propose the Dynamic Adaptive Graph Convolutional Network (DAGCN). The core idea is to use dynamic graph evolution to simulate continuous variations and local abrupt changes in spectral sequences, thereby capturing subtle spectral features of the Earth’s surface. The framework includes three components: The Dynamic Graph Sequence Construction (DGS) strategy uses superpixel segmentation and spatio-temporal graph modeling to construct slice graphs for each spectral band. The Multi-Scale Adaptive Graph Convolution (MAGC) module dynamically generates multi-scale adjacency matrices through adaptive convolutions, learning pixel features within homogeneous regions while preserving multi-scale context. The Spatio-Temporal Graph Collaborative Fusion (STGC) strategy includes a Dynamic Segmentation-based Graph Update (DSG Update) module and a Multi-Spectral Channel Attention (MSCA) mechanism. DSG Update dynamically optimizes the graph structure of adjacent temporal phases using MAGC’s adjacency matrix, extending independent multi-graph learning to spatio-temporally coupled multi-graph learning for modeling spectral-spatial evolution. MSCA assigns self-attention weights to each temporal phase to enhance discriminative band selection. Experiments on public HSI datasets show that DAGCN outperforms state-of-the-art methods in classification accuracy, especially in capturing subtle spectral variations, confirming its effectiveness and superiority.