Dual-Branch Spectral–Spatial Attention Network for Hyperspectral Image Classification

Abstract

In order to achieve accurate hyperspectral image (HSI) classification, the convolutional neural network (CNN) has been extensively utilized. However, most existing patch-based CNN methods overlook the relationship between central pixels and their surroundings. A novel dual-branch spectral–spatial attention network (DBSSAN) is proposed, which helps suppress the impact from interference elements and enhances effective feature extraction from complex features in HSI data. The global and local spatial features are fully integrated through the proposed spatial self-attention module. More specifically, it measures the relationship between the central and surrounding pixels based on cosine similarity and Gaussian–Euclidean similarity to extract global features, while the scale information extraction (SIE) model captures the local features. Furthermore, the inclusion of Transformer model enables the extraction of spectral information from a global perspective, facilitating the capture of long-distance dependencies and nonlinear correlations in HSI. The extracted spectral and spatial features are subsequently classified using a multilayer perceptron (MLP). Five publicly available hyperspectral datasets were used to present experimental evaluations, namely, Indian Pines, Kennedy Space Center, Pavia University, Houston2013, and Houston2018. The comparative results demonstrate the superior performance of the proposed network compared to several state-of-the-art methods.