Hyperspectral Unmixing with Endmember Variability using Partial Membership Latent Dirichlet Allocation

Sheng Zou, Alina Zare

Abstract

The application of Partial Membership Latent Dirichlet Allocation (PM-LDA) for hyperspectral endmember estimation and spectral unmixing is presented. PM-LDA provides a model for a hyperspectral image analysis that accounts for spectral variability and incorporates spatial information through the use of superpixel-based 'documents.' In our application of PM-LDA, we employ the Normal Compositional Model in which endmembers are represented as Normal distributions to account for spectral variability and proportion vectors are modeled as random variables governed by a Dirichlet distribution. The use of the Dirichlet distribution enforces positivity and sum-to-one constraints on the proportion values. Algorithm results on real hyperspectral data indicate that PM-LDA produces endmember distributions that represent the ground truth classes and their associated variability.

Introduction

In hyperspectral imaging, the spatial resolution of sensors is often coarser than the objects on the ground, leading to pixels that capture a mixture of different materials. The objective of hyperspectral unmixing is to decompose these mixed pixels into their constituent "pure" materials, known as endmembers, and their corresponding relative proportions, or abundances.

Traditional approaches often rely on the Linear Mixing Model, which assumes each material has a single, fixed spectral signature. However, real-world materials exhibit spectral variability—natural fluctuations in a material's signature due to factors like moisture, health, and lighting conditions. Failure to account for this variability results in noisy and inaccurate material maps. This paper introduces Partial Membership Latent Dirichlet Allocation (PM-LDA) to address these challenges by treating endmembers as statistical distributions and incorporating spatial context.

Methodology

PM-LDA Framework

The authors adapt Latent Dirichlet Allocation (LDA), a technique originally developed for natural language processing, to the domain of hyperspectral imagery. In this adaptation, the image is partitioned into "superpixels"—clusters of similar, neighboring pixels—which are treated as "documents." Each pixel within a superpixel is considered a "word."

Unlike standard LDA, where each word belongs to a single topic, PM-LDA allows a pixel to have partial membership in multiple endmembers (topics) simultaneously. This probabilistic approach accurately reflects the nature of mixed pixels in remote sensing.

Graphical model representation of the PM-LDA framework showing the relationships between hyperparameters, latent variables, and observed pixels.

Accounting for Variability

To handle spectral variability, the model employs the Normal Compositional Model (NCM). Endmembers are modeled as Normal distributions rather than fixed vectors, allowing the system to estimate both the average signature of a material and its variance. The model utilizes a Metropolis-within-Gibbs sampler to simultaneously estimate endmember means, covariances, and material proportions for every pixel.

Experimental Evaluation

The proposed method was evaluated using sub-images from the University of Pavia dataset, collected by the ROSIS sensor. This urban dataset contains 103 spectral bands. Two 50x50 pixel sub-images were used to compare PM-LDA against state-of-the-art methods: NCM-Bayes and S-PCUE.

Results and Discussion

Qualitative Comparison

Experimental results demonstrate that PM-LDA produces visually smoother proportion maps. A key advantage of PM-LDA is its ability to accurately separate similar materials, such as blue roofs and sidewalks, which often confuse other models. For instance, NCM-Bayes frequently groups these two distinct materials into a single class.

Comparison of estimated endmember means for NCM-Bayes, S-PCUE, and PM-LDA. Note how PM-LDA and S-PCUE successfully distinguish between Blue Roof and Sidewalk signatures.

Quantitative Performance

PM-LDA outperformed competing models across several metrics:

  • Proportion Entropy: PM-LDA achieved significantly lower entropy (1041 and 715 for the test sub-images) compared to NCM-Bayes (1542, 1411) and S-PCUE (2023, 1624), indicating cleaner unmixing results.
  • Log-Likelihood: The model showed the highest data log-likelihood (e.g., $15.35 \times 10^5$), suggesting a superior fit to the actual hyperspectral data.

Conclusion

PM-LDA offers a robust framework for hyperspectral unmixing by leveraging spatial context through superpixels and modeling spectral variability via Normal distributions. While the Bayesian sampling approach is computationally intensive and dependent on initial superpixel segmentation quality, its ability to simultaneously estimate all model parameters makes it highly effective for complex scenes. Future work may focus on optimizing computational efficiency and exploring more advanced segmentation techniques to further enhance the model's performance.

Summary updated

Abstract

The application of Partial Membership Latent Dirichlet Allocation(PM-LDA) for hyperspectral endmember estimation and spectral unmixing is presented. PM-LDA provides a model for a hyperspectral image analysis that accounts for spectral variability and incorporates spatial information through the use of superpixel-based 'documents.' In our application of PM-LDA, we employ the Normal Compositional Model in which endmembers are represented as Normal distributions to account for spectral variability and proportion vectors are modeled as random variables governed by a Dirichlet distribution. The use of the Dirichlet distribution enforces positivity and sum-to-one constraints on the proportion values. Algorithm results on real hyperspectral data indicate that PM-LDA produces endmember distributions that represent the ground truth classes and their associated variability.
BibTeX
@inproceedings{zou2017hyperspectral,
  author = {Sheng Zou and Alina Zare},
  title = {{Hyperspectral Unmixing with Endmember Variability using Partial Membership Latent Dirichlet Allocation}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2017},
  doi = {10.1109/icassp.2017.7953348},
}
Figure