GCN-CNN-Based Bidirectional Heterogeneous Feature Interaction Network for Hyperspectral Image Change Detection
TL;DR
A graph convolutional network-convolutional neural network-based bidirectional heterogeneous feature interaction network, termed GCMIN, which introduces a hierarchical bidirectional interaction architecture between pixel-domain convolutional features and dual-granularity graph representations and facilitates the refinement of local spectral–spatial information and nonlocal topological relationships during feature learning.
Abstract
Hyperspectral image change detection (HSI-CD) is an important task for monitoring dynamic land–surface changes. However, integrating fine-grained local spectral–spatial information with nonlocal structural dependencies remains challenging, particularly under the high-dimensional and limited-label conditions inherent to hyperspectral data. To address these challenges, we propose a graph convolutional network-convolutional neural network-based bidirectional heterogeneous feature interaction network, termed GCMIN. At its core, GCMIN introduces a hierarchical bidirectional interaction architecture between pixel-domain convolutional features and dual-granularity graph representations. This architecture enables progressive information exchange and facilitates the refinement of local spectral–spatial information and nonlocal topological relationships during feature learning. Furthermore, a cross-level spectral–spatial adaptive module is developed to integrate heterogeneous representations from different network depths through multilevel feature aggregation and adaptive weighting. It combines fine-grained information from shallow layers with high-level semantic representations from deeper layers. In addition, Kolmogorov–Arnold network-based nonlinear mappings are employed for feature transformation and classification, reducing parameter redundancy compared with conventional multilayer perceptron-based mappings. Extensive experiments on three publicly available HSI-CD datasets demonstrate that GCMIN achieves superior performance in terms of overall accuracy, Kappa coefficient, and F1-score compared with the evaluated methods. Ablation and complexity analyses further reveal the complementary contributions of the proposed components and show that GCMIN maintains a compact model size while preserving competitive inference efficiency. Moreover, experiments with varying training sample ratios demonstrate robustness under limited-label conditions. The code is available online.