Research on Hyperspectral Remote Sensing Image Data Augmentation Method Based on CVAEWGAN

To address the issue of dimensionality disaster caused by the scarcity of labeled samples and high-dimensional spectral features in hyperspectral remote sensing image classification, traditional supervised learning methods face significant performance degradation under small sample conditions. Therefore, this study proposes a joint framework, CVAE-WGAN, which integrates Conditional Variational Autoencoder (CVAE) and Wasserstein Generative Adversarial Network (WGAN) to enhance sample quality and distribution consistency through a multi-stage generation mechanism. Firstly, CVAE is used to model the latent space and generate samples under class conditions. Subsequently, WGAN’s adversarial discrimination mechanism is introduced, utilizing the Wasserstein distance and a gradient penalty term to optimize the spectral fidelity of generated samples and alleviate the mode collapse issue present in traditional generative models. Experimental results on three public datasets, Indian Pines, Pavia University, and Salinas, show that CVAE-WGAN improves the average spectral angle and mean squared error by over $28 \%$ and $25 \%$, respectively. Compared to well-performing classification methods, the overall classification accuracy is improved by approximately $15 \%$. During the training process, the various loss curves of CVAE-WGAN gradually converge and remain stable. This method provides a data augmentation solution for small-sample hyperspectral classification tasks that combines spectral fidelity, class controllability, and training stability.