1: \begin{abstract}

2: This paper presents a nonlinear mixing model for hyperspectral image

3: unmixing. The proposed model assumes that the pixel reflectances are

4: post-nonlinear functions of unknown pure spectral components

5: contaminated by an additive white Gaussian noise. These nonlinear

6: functions are approximated using polynomials leading to a polynomial

7: post-nonlinear mixing model. A Bayesian algorithm is proposed to

8: estimate the parameters involved in the model yielding an

9: unsupervised nonlinear unmixing algorithm. Due to the large number

10: of parameters to be estimated, an efficient Hamiltonian Monte Carlo

11: algorithm is investigated. The classical leapfrog steps of this

12: algorithm are modified to handle the parameter constraints. The

13: performance of the unmixing strategy, including convergence and

14: parameter tuning, is first evaluated on synthetic data. Simulations

15: conducted with real data finally show the accuracy of the proposed

16: unmixing strategy for the analysis of hyperspectral images.

17: \end{abstract}

18: