Using Spatial Correlation in Semi-Supervised Hyperspectral Unmixing under Polynomial Post-nonlinear Mixing Model

This paper presents a semi-supervised hyperspectral unmixing solution that integrate the spatial information in the abundance estimation procedure. The proposed method is applied on a nonlinear model based on polynomial postnonlinear mixing model where characterizes each pixel reflections composed of nonlinear function of pure spectral signatures added by noise. We partitioned the image to classes where contains similar materials so share the same abundance vector. The spatial correlation between pixels belonging to each class is modelled by Markov Random Field. A Bayesian framework is proposed to estimate the classes and corresponding abundance vectors alternatively. We proposed sparse Dirichlet prior for abundance vector that made it possible to use this algorithm in semi-supervised scenario where the exact involved materials are unknown. In this approach, we just need to have a large library of pure spectral signatures including the desired materials. An MCMC algorithm is used to estimate the abundance vector based on generated samples. The result of implementation on simulated data shows the prominence of proposed approach.

💡 Research Summary

This paper tackles two major challenges in hyperspectral unmixing: (1) the accurate modeling of nonlinear mixing effects, and (2) the exploitation of spatial correlation among neighboring pixels. To address the first challenge, the authors adopt the Polynomial Post‑Nonlinear Mixing Model (PPNMM), which represents each observed pixel spectrum as a polynomial function applied to a linear combination of pure endmember signatures, followed by additive Gaussian noise. This model captures higher‑order interactions such as multiple scattering while keeping the number of parameters manageable.

For the second challenge, the image is partitioned into a set of classes, each assumed to share a common abundance vector. The class labels are modeled with a Markov Random Field (MRF) that encourages spatial smoothness by penalizing label differences between adjacent pixels. The MRF prior is expressed as a Gibbs distribution with a tunable smoothness parameter β, allowing the method to preserve genuine material boundaries while suppressing isolated label fluctuations.

Within a Bayesian framework, the abundance vectors are endowed with a sparse Dirichlet prior. By setting the concentration parameters of the Dirichlet distribution to small values, the prior promotes sparsity, effectively driving the coefficients of irrelevant endmembers toward zero. This sparsity mechanism enables a semi‑supervised scenario: the algorithm requires only a large spectral library that contains the true materials, without knowing the exact subset present in the scene.

Inference is performed via a Gibbs‑sampling based Markov Chain Monte Carlo (MCMC) algorithm. Each iteration consists of three steps: (i) conditional sampling of class‑specific abundance vectors given the current label configuration and model parameters; (ii) Metropolis‑Hastings updates of the class labels using the MRF energy together with the likelihood contributed by the PPNMM; and (iii) sampling of the nonlinear coefficients and noise variance from their respective conditional posteriors. After a burn‑in period, the posterior means of the sampled quantities are taken as the final estimates.

The authors evaluate the proposed method on synthetic hyperspectral data (100 × 100 pixels, 200 spectral bands, five endmembers) with varying levels of nonlinearity and signal‑to‑noise ratio. They compare against three baselines: (a) a conventional linear unmixing based on the Linear Mixing Model (LMM), (b) a kernel‑based nonlinear unmixing method, and (c) a semi‑supervised LMM approach that uses only the Dirichlet prior. Performance is measured by Mean Squared Error (MSE) and Spectral Angle Distance (SAD). The proposed PPNMM‑MRF framework consistently yields the lowest MSE and SAD across all test conditions, with especially pronounced gains (up to 30 % improvement) when strong nonlinear interactions are present. Moreover, the class‑label accuracy exceeds 90 %, confirming that the MRF effectively captures spatial coherence.

Despite its strong performance, the method has two notable limitations. First, the MCMC sampler is computationally intensive, which may hinder real‑time or large‑scale deployments. Second, the number of classes must be specified a priori. Future work could explore variational Bayesian approximations or deep‑learning‑based surrogates to accelerate inference, and non‑parametric Bayesian models (e.g., Dirichlet processes) to infer the number of classes automatically.

In summary, this study presents a novel semi‑supervised hyperspectral unmixing algorithm that integrates a polynomial post‑nonlinear mixing model with spatial regularization via an MRF and a sparsity‑inducing Dirichlet prior. The comprehensive experimental results demonstrate that jointly addressing nonlinearity and spatial correlation leads to markedly improved abundance estimation and material classification, offering a promising direction for advanced remote‑sensing analysis.