URDM: Hyperspectral Unmixing Regularized by Diffusion Models.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Hyperspectral unmixing is crucial for analyzing mixed pixels in hyperspectral imagery.
• Traditional methods rely on hand-crafted regularizers, limiting performance and data information capture.
• Deep learning methods offer capability but lack generalizability and interpretability.

Purpose of the Study:

• To develop a novel hyperspectral unmixing method that overcomes limitations of traditional and deep learning approaches.
• To enhance the generalizability and interpretability of hyperspectral unmixing.
• To leverage the strengths of both optimization algorithms and deep generative models.

Main Methods:

• Proposed a hyperspectral unmixing method regularized by a diffusion model (URDM).
• Formulated the unmixing objective function from a variational perspective.
• Integrated the objective function into a diffusion sampling process using a denoising diffusion probabilistic model (DDPM).
• Employed a splitting-based strategy to simplify the optimization of the objective function.

Main Results:

• The proposed URDM method demonstrated superior performance on both synthetic and real hyperspectral datasets.
• The integration of diffusion models introduced effective generative priors.
• The splitting strategy facilitated efficient optimization of the complex objective function.

Conclusions:

• The URDM method offers an effective approach to hyperspectral unmixing, combining optimization and generative modeling.
• The method shows improved efficiency, generalizability, and interpretability compared to existing techniques.
• This work highlights the potential of diffusion models in hyperspectral image processing.