Beyond dimensionality explosion: A latent diffusion framework for hyperspectral image classification

Meiyun Lu

doi:10.1016/j.neucom.2025.131249

Beyond dimensionality explosion: A latent diffusion framework for hyperspectral image classification

2025-08-09

0

PRE

AI

DOI：10.1016/j.neucom.2025.131249

Original

Original request for help

Save

Abstract

En 中文

In recent years, deep generative learning-based methods for hyperspectral image (HSI) classification have achieved significant progress. However, the inherently high dimensionality of hyperspectral data, combined with the modeling complexity of generative approaches, often leads these methods to suffer from a curse of dimensionality explosion during feature extraction. To address this issue, this paper proposes an HSI classification and generation framework based on a latent diffusion model. The framework simulates and reconstructs the spectral–spatial high-dimensional feature distribution through a generative diffusion mechanism in a low-dimensional latent space, and realizes feature extraction via latent features. This approach effectively avoids feature dimensionality explosion while achieving efficient spectral–spatial feature fusion. Specifically, the framework consists of a spectral–spatial diffusion module and an attention-based classification module. The spectral–spatial diffusion module first maps the high-dimensional HSI data to a low-dimensional latent space, where spectral–spatial diffusion modeling and noise injection are performed. Subsequently, the high-dimensional features are restored through a reverse denoising process, facilitating efficient fusion of spatial and spectral information. After training, low-dimensional features extracted from the latent space are fed directly into the classification module. The core advantage of this method lies in its latent space diffusion modeling and feature extraction, which effectively integrates spectral–spatial information while avoiding dimensionality explosion, thereby improving classification performance. Experimental results on multiple public datasets demonstrate that our framework outperforms previous state-of-the-art methods.

Keywords:

hyperspectral image classification

latent diffusion model

spectral-spatial feature fusion

dimensionality reduction

generative modeling

Beyond dimensionality explosion: A latent diffusion framework for hyperspectral image classification

Abstract

Journal

Organization