An Efficient Topology Construction Scheme Designed for Graph Neural Networks in Hyperspectral Image Classification
Year: 2025
Type: article
Abstract: Hyperpixel-based graph neural networks (GNNs) have achieved remarkable results in hyperspectral image (HSI) classification tasks, mainly due to their ability to capture the implied topology while maintaining low computational complexity by transferring information between spatially neighboring hyperpixels. However, the representation assumption that pixel features within a hyperpixel are treated as identical may limit the expressive power of the model because feature regions in hyperspectral imagery are usually irregular, leading to difficulties in homogeneity of pixels within a hyperpixel and determination of the scale of the hyperpixel. To overcome this problem, we re-understand the implied topology between pixels through their spectral feature similarity and spatial location dependence. Considering the computational bottleneck, we adopt a new subgraph delineation method and sorting and filtering techniques to select important relation pairs from the graph, so as to construct an a priori topology that favors the downstream tasks on HSIs of arbitrary size. Based on this a priori topology, we construct the GCN module to optimize the representation of the pixel features and integrate it into a unified framework.We conducted a series of comprehensive experiments on three benchmark datasets, and the results show that the proposed method has a clear advantage in distinguishing compact, discontinuous, and small- to medium-sized feature regions and even outperforms some of the powerful superpixel-level fusion models proposed in recent years. In addition, we find that different models exhibit different discriminative abilities when dealing with feature regions of different characteristics. In particular, the classification performance is significantly improved by combining these models with our approach. Our code is open-source on the public platform GitHub: https://github.com/LittleBlackBearLiXin/Pixel-level-graph-neural-networks-based-on-optimized-feature-representation-for-hyperspectral-image.
Source: IEEE Access
Authors: Yu Zhang, Xin Li, Yaoqun Xu
Institution: Harbin University of Commerce
Language: English
Subfield: Media Technology
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W4412985134