HDECGCN: A Heterogeneous Dual Enhanced Network Based on Hybrid CNNs Joint Multiscale Dynamic GCNs for Hyperspectral Image Classification

A hyperspectral image (HSI) classification algorithm that combines graph convolutional networks (GCNs) and convolutional neural networks (CNNs) aims to generate complementary spatial-spectral joint information at the superpixel and pixel levels. However, the CNN part is typically a single 2-D or 3-D network that cannot fully capture the middle or long-range spatial relationships between pixels. Additionally, the GCNs part is commonly under-segmented in the superpixel segmentation process and does not consider the weight between neighboring superpixels when calculating the adjacency matrix. Therefore, this article proposes a multiscale dynamic tuning parameter, where the dual superpixel segmentation GCN strategy joins the enhanced hybrid 3-D–2-D CNN framework to enhance the superpixel and pixel complementary nature. The hybrid enhanced CNN branch uses the groupable convolutions with a mixed spectral stacking and residual nonlocal block at the hybrid convolution output to overcome the accuracy degradation problem caused by long convolutional layers and poor generalization performance of a single network structure. An additional branch performs simple linear iterative clustering and entropy rate superpixel (ERS) segmentation, which are sequentially implemented on the HSI to solve the under-segmentation problem. This strategy is important, as dynamically calculating the tuning parameters for feature segmentation maps increases the number of the multiscale GCN layers and fully extracts contextual spatial information. Experiments on three public datasets, Indian Pines, Kennedy Space Center, and the University of Pavia, demonstrate that the proposed framework achieves the optimal OA, AA, and Kappa coefficients. The source code is available at
https://github.com/henulx/HDECGCN-Framework
.