DiffusionAAE: Enhancing hyperspectral image classification with conditional diffusion model and Adversarial Autoencoder

Abstract

Hyperspectral image (HSI) classification is essential for ecological monitoring, but faces significant challenges due to high dimensionality, complex spectral–spatial relationships, and limited labeled data. This study introduces DiffusionAAE, a novel framework that uniquely combines Adversarial Autoencoders (AAE) with conditional diffusion models to address these challenges. Unlike existing approaches, DiffusionAAE incorporates spectral similarity constraints and class label guidance into the diffusion process, ensuring the generation of physically realistic synthetic samples. Our framework’s key innovation lies in its two-stage architecture: first, an AAE extracts robust latent features capturing intricate spectral–spatial relationships; second, a conditional diffusion model refines these features through progressive denoising, enabling class-specific feature generation while maintaining physical constraints inherent to hyperspectral data. Comprehensive experiments on three benchmark datasets demonstrate DiffusionAAE’s superior performance: compared to state-of-the-art methods, our approach achieves significant improvements with an overall accuracy (OA) of 96.77% on Indian Pines (3.21% higher than CNN-based methods), 99.56% on University of Pavia (1.24% improvement over Transformer-based approaches), and 99.62% on Salinas (0.98% better than the best competing method). Notably, DiffusionAAE shows remarkable performance on minority classes, with an average 7.35% accuracy improvement across underrepresented classes in the Indian Pines dataset. The framework demonstrates particular strength in scenarios with limited training data, maintaining 95.3% accuracy even when using only 5% of available labeled samples. These results establish DiffusionAAE as a significant advancement for ecological informatics applications, especially for biodiversity monitoring and land cover classification where labeled data scarcity and class imbalance are prevalent challenges.