Unsupervised post-nonlinear unmixing of hyperspectral images using a Hamiltonian Monte Carlo algorithm

Area of Science:

• Remote Sensing
• Image Processing
• Signal Processing

Background:

• Hyperspectral imaging generates complex data where pixels contain mixtures of pure spectral signatures.
• Traditional linear unmixing models are insufficient for accurately representing real-world spectral mixtures.
• Nonlinear mixing phenomena in hyperspectral images necessitate advanced unmixing techniques.

Purpose of the Study:

• To develop and validate a novel nonlinear mixing model for hyperspectral image unmixing.
• To propose an unsupervised Bayesian algorithm for estimating model parameters.
• To address the challenges of parameter estimation in complex nonlinear models.

Main Methods:

• A polynomial post-nonlinear mixing model is proposed, approximating nonlinear functions with second-order polynomials.
• A Bayesian algorithm is developed for unsupervised parameter estimation.
• An efficient Hamiltonian Monte Carlo algorithm with modified leapfrog steps is employed to handle parameter constraints.

Main Results:

• The proposed unmixing strategy demonstrates accurate parameter estimation and convergence on synthetic data.
• Simulations using real hyperspectral data confirm the effectiveness of the nonlinear unmixing approach.
• The method provides accurate analysis for hyperspectral image unmixing tasks.

Conclusions:

• The polynomial post-nonlinear mixing model effectively captures spectral nonlinearities in hyperspectral images.
• The developed Bayesian algorithm offers an unsupervised and accurate solution for nonlinear spectral unmixing.
• This approach advances the analysis capabilities of hyperspectral imaging.