Hyperspectral unmixing with endmember variability using Partial Membership Latent Dirichlet Allocation

Sheng Zou, Alina Zare

TL;DR

This paper introduces a novel approach to hyperspectral unmixing called Partial Membership Latent Dirichlet Allocation (PM-LDA). By treating endmembers as statistical distributions rather than fixed points and incorporating spatial context through superpixels, the model effectively accounts for spectral variability—the natural changes in a material's signature due to lighting or environmental conditions—resulting in more accurate and "smoother" material maps.

What Problem Does This Paper Address?

In hyperspectral imaging, each pixel often captures a mixture of different materials (e.g., a mix of grass and soil) because the sensor's spatial resolution is coarser than the objects on the ground. The goal of unmixing is to identify the "pure" materials (endmembers) and their relative proportions (abundances) in every pixel.

The traditional "Linear Mixing Model" assumes each material has one single, fixed spectral signature. However, in the real world, the signature of "grass" varies depending on moisture, health, and shadows. This is known as spectral variability. If a model doesn't account for this "wiggle room," the resulting material maps become noisy and inaccurate.

Methods

The authors adapt a technique from natural language processing called Latent Dirichlet Allocation (LDA). In text analysis, LDA assumes a document is a mix of topics, and each word belongs to one topic.

The authors use an extension called PM-LDA and apply it to imagery:

  • Superpixels as Documents: The image is first partitioned into "superpixels" (clusters of similar, neighboring pixels). Each superpixel is treated as a "document."
  • Pixels as Words: Each pixel within that superpixel is a "word."
  • Partial Membership: Unlike standard LDA, PM-LDA allows a "word" (pixel) to have partial membership in multiple "topics" (endmembers) simultaneously. This perfectly mirrors the reality of mixed pixels.
  • Normal Compositional Model (NCM): To handle variability, endmembers are modeled as Normal distributions rather than fixed vectors. This allows the model to estimate not just the average signature of a material, but also its variance.
  • Parameter Estimation: The model uses a Metropolis-within-Gibbs sampler, a sophisticated statistical tool, to simultaneously estimate the endmember means, their covariances, and the proportion of each material in every pixel.

Datasets

The researchers evaluated their method using sub-images from the University of Pavia dataset.

  • Source: Collected by the ROSIS (Reflective Optics System Imaging Spectrometer) sensor.
  • Specs: The original scene covers an urban area in Italy with 103 spectral bands.
  • Test Cases: Two 50x50 pixel sub-images were used to compare PM-LDA against existing state-of-the-art methods (NCM-Bayes and S-PCUE).

Key Insights and Ideas

  • Spatial Context Matters: By grouping pixels into superpixels, the model leverages the fact that neighboring pixels are likely to contain similar materials. This "spatial smoothing" prevents the noisy, speckled results often seen in pixel-by-pixel unmixing.
  • Distributions over Points: Modeling endmembers as statistical distributions (the "endmembers as distributions" approach) provides a mathematically robust way to handle the inherent "fuzziness" of real-world spectral signatures.
  • Simultaneous Estimation: PM-LDA is unique because it can estimate endmember means, covariances, and abundances all at once, whereas many competing methods require some of these values to be known or fixed beforehand.

Key Results

The paper compared PM-LDA against two other prominent models (NCM-Bayes and S-PCUE) using both qualitative and quantitative metrics:

  • Qualitative: PM-LDA produced proportion maps that were visually "smoother" and more accurately separated similar materials (like blue roofs vs. sidewalks) which confused the other models.
  • Proportion Entropy: This measures how "clean" the unmixing is (lower is better). PM-LDA achieved the lowest entropy (1041 and 715 for the two sub-images) compared to NCM-Bayes (1542, 1411) and S-PCUE (2023, 1624).
  • Log-Likelihood: This measures how well the model fits the actual data (higher is better). PM-LDA consistently showed the highest overall data log-likelihood (e.g., $15.35 \times 10^5$ for the combined sub-images).

Limitations and Future Work

While the results are promising, the paper notes a few areas for consideration:

  • Computational Complexity: Bayesian sampling methods like the Metropolis-within-Gibbs sampler can be computationally expensive and slow compared to simpler geometric unmixing methods.
  • Superpixel Dependency: The quality of the unmixing depends on the initial superpixel segmentation. If the segmentation is poor, it could negatively impact the "document" structure of the model.
  • Future Directions: The authors suggest that PM-LDA's ability to estimate all parameters simultaneously makes it a strong candidate for more complex scenes where ground truth data is scarce.

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},
}

Loading related content...

Figure