Semi-Supervised Classification Module for Hyperspectral Images Based on Dual-Stream Attention and Kernel Density Estimation

Deep Neural Networks (DNN) are widely used in hyperspectral image (HSI) classification due to their powerful ability to extract spatial-spectral features. However, the scarcity of labeled samples limits the performance of DNNs. In this paper, an HSI pseudo-sample generation module based on dual-stream attention and kernel density estimation (KDE) is proposed, which enhances the robustness and generalization ability of DNNs by generating and feeding a sufficient number of high-fidelity samples into the network. First, a dual-stream spatial-spectral feature extraction network with serial channel and spatial attention is constructed to map the spectral dimensions to a low-dimensional space while reconstructing discriminative features for different classes. Then, KDE is used to perceive the probability density function of the class distribution in the low-dimensional space. Finally, samples with high matching degrees to the class distribution are selected to generate pseudo-samples. More importantly, our proposed module can be embedded into any DNNs to further improve the performances of these models. Extensive experimental results show that the pseudo-samples generated by our module not only effectively alleviate performance fluctuations of the original network but also significantly improve the classification accuracy of the network.