Dual linear latent space constrained generative adversarial networks for hyperspectral image classification

Abstract

Hyperspectral image classification is a critical challenge in remote sensing due to the high dimensionality of the data and the scarcity of labeled samples. In this study, we propose a novel Dual-Line Latent Space Constrained Generative Adversarial Network (DLC-GAN) that integrates spatial and spectral feature extraction through dual pathways and incorporates latent space constraints to enhance classification robustness. Unlike existing methods, the DLC-GAN employs a bilinear structure to extract complementary information, improving both feature representation and classification accuracy. The model was evaluated on benchmark datasets such as Indian Pines, Pavia University, and Salinas, achieving state-of-the-art performance. Specifically, the DLC-GAN demonstrated improvements in overall accuracy by 5–11 % compared to recent methods and exhibited superior adaptability to limited training data. These findings underscore the potential of DLC-GAN in addressing critical challenges in hyperspectral image classification, with promising applications in environmental monitoring, agricultural management, and urban planning.