Enhancing remote target classification in hyperspectral imaging using graph attention neural network

Abstract

The method of target classification known as hyperspectral imaging (HSI) is commonly used in the field of remote sensing. However, recent research has shown that categorizing HSI can be problematic due to the limited availability of labelled data. There is significant interest in applying this technique to hyperspectral data. Previous graph neural network (GNN)-based methodologies often used a graph filter to obtain HSI properties, but the potential advantages of various graph neural networks and graph filters have not been fully exploited. GNNs often operate under the assumption that a node’s neighbours are independent of each other, neglecting potential interactions among them. To overcome these limitations, graph attention neural network-based remote target classification (GANN-RTC) has been proposed. It has the ability to handle both the labelled and unlabelled datasets. To evaluate the performance of GANN-RTC, we compared it with existing methods using performance measures such as individual class accuracy, overall accuracy, and the Kappa coefficient. The findings indicate that the GANN-RTC yields enhancements in OA, ICA, and KC by 2.32, 7.89, and 2.47% for the Cuprite dataset and 4.79, 11.85, and 2.82% for the Pavia University dataset.

Research highlights

• The research focuses on remote target classification in hyperspectral imaging using a Graph Attention Neural Network.

• Previous methods in this field have not fully utilized the potential advantages of graph filters and graph neural networks.

• The proposed approach overcomes limitations by considering interactions between neighbouring nodes and can handle both labelled and unlabelled datasets.

• Performance evaluation shows significant improvements in overall accuracy, individual class accuracy, and the Kappa coefficient compared to existing state-of-the-art methods.