Deep semantic network for hyperspectral image classification using enhanced bifurcation and light self-Gaussian attention mechanism

Convolutional neural networks (CNNs) are widely used in hyperspectral image (HSI) classification due to their strong feature extraction capabilities. Nevertheless, CNN-based classification methods face challenges in capturing deep semantic features effectively, and as the number of layers increases, computational expenses escalate significantly. In addition, spatial locations and spectral bands in HSI data have varying discriminative power. To address these challenges, we propose a novel approach, DSNet, which integrates a multiway attention mechanism with a CNN-based deep semantic network to improve HSI classification. First, we extract shallow abstract features using a module that combines 3D and 2D convolution layers, followed by transformation through a Gaussian-weighted feature tokenizer. Second, we introduce an enhanced bifurcation attention mechanism with dynamically adjustable perceptual ranges to capture spatial semantic information from both horizontal and vertical perspectives. Third, a convolutional multilayer perceptron module is employed to extract high-level feature representations in both spatial and spectral dimensions, which are then refined by a light self-Gaussian attention module. Finally, a linear layer is used to determine the sample labels. Experiments conducted on three publicly available hyperspectral datasets illustrate that the proposed DSNet achieves superior performance compared with several contemporary state-of-the-art methods, even when trained with limited samples.