Hyperspectral Image Classification Using Diffusion Model

Abstract

Hyperspectral image (HSI) classification is extensively used in Earth science and is important for remote sensing. Many deep learning techniques have been developed recently for HSI classification; nevertheless, difficulties are frequently encountered with high-dimensional and complex data, making it challenging for relationships between various data elements to be captured. To address this, a novel method, dubbed "SpectralDiff," is proposed, which employs diffusion models for HSI classification. In this approach, noise in the data is repeatedly reduced, creating a clearer representation of the data's structure, thereby facilitating the handling of redundant and high-dimensional data. The framework consists of two major components:
Spectral-Spatial Diffusion Module: The establishment of connections between data samples is facilitated by the spectral-spatial diffusion module, without requiring prior knowledge of the structure. Spatial (position-related) and spectral (color-related) information from the HSI data is extracted. Attention-Based Classification Module: The features gleaned from the diffusion module are then used to classify each pixel in the image. This approach, which emphasizes the connections between multiple samples, enables better classification. Tests conducted on three publicly available datasets demonstrate that SpectralDiff achieves superior performance compared to other state-of-the-art techniques.