Double-Branch Multilevel Skip Sparse Graph Attention Network for Hyperspectral Image Classification

Graph neural networks (GNNs) offer a potent framework for learning and representing data with graph structures. In the context of hyperspectral images (HSIs) classification tasks, GNNs have been successfully employed. Among the diverse variations of GNNs, graph attention networks (GATs) stand out, demonstrating cutting-edge performance in node prediction. GATs excel by acquiring the capability to assign dense attention coefficients to each neighboring node during the process of feature aggregation. First, the intricate nature of land distribution, coupled with the high dimensionality of hyperspectral imaging (HSI) data, poses a hurdle for the direct application of GATs in effectively distinguishing between various coverage categories. Second, the insufficiency of spectral information further compounds the issue by adversely affecting classification accuracy, contributing to GATs’ susceptibility to overfitting during training. Additionally, prevailing classification approaches commonly derive features either at the pixel or superpixel levels. The utilization of single-scale feature extraction techniques proves inadequate in concurrently addressing both local and global features of land cover. This limitation contributes to less-than-optimal classification outcomes. To address this issue, this article proposes a double-branch multilevel skip sparse GAT for HSI classification. The network comprises two parallel branches of sparse GATs, extracting features from the perspectives of pixels and superpixels. Via the construction of a multilevel skip graph, these two branches adeptly gather both local and global context information. Subsequently, a fully connected network is deployed to classify the integrated features derived from both branches. The effectiveness of the proposed method is validated through experimental results on three benchmark HSI datasets, showcasing its superior performance.