Simultaneously sparse and low-rank abundance matrix estimation for hyperspectral image unmixing

Paris Giampouras, Konstantinos Themelis, Athanasios A. Rontogiannis, Konstantinos KoutroumbasIEEE TGRS 2015 · DOI · OpenReview

TL;DR

This paper introduces a novel approach to hyperspectral image unmixing by simultaneously modeling the abundance matrix as both sparse and low-rank. Two new algorithms, IPSpLRU and ADSpLRU, are proposed to solve this problem, demonstrating improved accuracy and robustness compared to existing methods on both simulated and real-world data.

What Problem Does This Paper Address?

Hyperspectral images (HSIs) capture detailed spectral information across numerous bands, making them valuable for various applications. However, each pixel in an HSI is often a mixture of several materials (endmembers), and determining the proportion of each endmember (abundance) in each pixel is a challenging task known as spectral unmixing. Traditional methods often rely on a linear mixing model (LMM) and impose constraints like nonnegativity and sum-to-one on the abundance vectors. This paper addresses the limitations of these methods by recognizing that the abundance matrix (which stacks abundance vectors of neighboring pixels) often possesses inherent structural properties: sparsity (only a few endmembers contribute to a pixel) and low-rankness (neighboring pixels in homogeneous regions tend to have similar spectral compositions, leading to linearly dependent abundance vectors). The challenge is to effectively exploit both these properties simultaneously for more accurate unmixing.

Methods

The core of the proposed method lies in formulating the unmixing problem as a sparse reduced-rank regression problem. This is achieved by defining a novel cost function that regularizes the standard least-squares fitting term with a mixed penalty term. This penalty term is a combination of:

  • Weighted $\ell_1$ norm: Encourages sparsity in individual abundance vectors.
  • Weighted nuclear norm: Promotes low-rankness in the abundance matrix.

To solve this complex, non-smooth optimization problem, two novel algorithms are developed:

  1. Incremental Proximal Sparse Low-Rank Unmixing (IPSpLRU): This algorithm leverages the proximal operators of each term in the cost function and applies them sequentially in a cyclic manner. It's inspired by existing incremental proximal methods.
  2. Alternating Direction Method of Multipliers (ADMM) Sparse and Low-Rank Unmixing (ADSpLRU): This algorithm utilizes the ADMM framework to decompose the problem and iteratively update variables, offering a different approach to optimization.

A key aspect of both algorithms is the use of reweighting schemes for the $\ell_1$ and nuclear norms. This dynamic adjustment of weights based on current estimates is shown to significantly improve accuracy, although it can make the problem non-convex.

Datasets

The paper evaluates its proposed algorithms on:

  • Simulated Datacubes: Generated using the linear mixing model with synthetic abundance matrices exhibiting various combinations of sparsity and low-rankness. The endmembers were selected from the USGS spectral library.
  • Real Hyperspectral Data: A portion of the Salinas vegetation scene acquired by the AVIRIS sensor. This dataset contains eight different vegetation species.

No new benchmark datasets were introduced.

Key Insights and Ideas

  • Simultaneous Sparsity and Low-Rankness is Crucial: The paper strongly demonstrates that exploiting both sparsity and low-rankness simultaneously leads to significantly better unmixing results than considering either property in isolation. This is a key conceptual advance.
  • Spatial Correlation is Key: The low-rankness assumption effectively captures the spatial correlation present in homogeneous regions of hyperspectral images, a crucial piece of information often overlooked.
  • Reweighting Enhances Performance: Dynamically reweighting the $\ell_1$ and nuclear norms based on intermediate estimates proves to be highly effective in improving accuracy, even though it introduces non-convexity.
  • ADSpLRU Offers Faster Convergence: While both algorithms are effective, ADSpLRU generally converges faster than IPSpLRU, albeit with a slightly higher per-iteration computational cost.

Key Results

  • Improved Accuracy: Both IPSpLRU and ADSpLRU consistently outperform state-of-the-art methods (CSUnSAL, MMV-ADMM, BiICE) on simulated data, achieving lower RMSE and higher SRE, especially when the abundance matrix is both sparse and low-rank.
  • Robustness to Noise: The proposed algorithms demonstrate robustness to both white and correlated Gaussian noise across a range of Signal-to-Noise Ratios (SNRs).
  • Qualitative Improvements on Real Data: On the Salinas dataset, the abundance maps generated by IPSpLRU and ADSpLRU show patterns that are closer to the principal components of the image and better highlight the presence of specific vegetation species, particularly brocolli_a and brocolli_b.
  • Parameter Sensitivity: The optimal choice of sparsity ($\gamma$) and low-rankness ($\tau$) parameters is shown to be dependent on the inherent structure of the abundance matrix, necessitating fine-tuning.

Limitations and Future Work

  • Non-convexity with Reweighting: The reweighting scheme, while beneficial for accuracy, introduces non-convexity into the optimization problem, for which theoretical convergence guarantees are not established. However, the authors note that the algorithms exhibited robust convergence behavior in practice.
  • Computational Complexity: While IPSpLRU has lower per-iteration complexity than ADSpLRU, both are more computationally intensive than methods that don't exploit spatial correlation. The paper mentions that deriving more computationally efficient schemes, potentially alleviating the need for Singular Value Decomposition (SVD), is an area for future investigation.
  • Exploiting Sparsity Structure: The paper suggests that further research could explore exploiting the specific structure or pattern of sparsity that might be implicitly imposed by the low-rankness property, which could lead to even better estimation performance.

Summary updated

Abstract

In a plethora of applications dealing with inverse problems, e.g. in image processing, social networks, compressive sensing, biological data processing etc., the signal of interest is known to be structured in several ways at the same time. This premise has recently guided the research to the innovative and meaningful idea of imposing multiple constraints on the parameters involved in the problem under study. For instance, when dealing with problems whose parameters form sparse and low-rank matrices, the adoption of suitably combined constraints imposing sparsity and low-rankness, is expected to yield substantially enhanced estimation results. In this paper, we address the spectral unmixing problem in hyperspectral images. Specifically, two novel unmixing algorithms are introduced, in an attempt to exploit both spatial correlation and sparse representation of pixels lying in homogeneous regions of hyperspectral images. To this end, a novel convex mixed penalty term is first defined consisting of the sum of the weighted $\ell_1$ and the weighted nuclear norm of the abundance matrix corresponding to a small area of the image determined by a sliding square window. This penalty term is then used to regularize a conventional quadratic cost function and impose simultaneously sparsity and row-rankness on the abundance matrix. The resulting regularized cost function is minimized by a) an incremental proximal sparse and low-rank unmixing algorithm and b) an algorithm based on the alternating minimization method of multipliers (ADMM). The effectiveness of the proposed algorithms is illustrated in experiments conducted both on simulated and real data.
BibTeX
@article{giampouras2015simultaneously,
  author = {Paris Giampouras and Konstantinos Themelis and Athanasios A. Rontogiannis and Konstantinos Koutroumbas},
  title = {{Simultaneously sparse and low-rank abundance matrix estimation for hyperspectral image unmixing}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2015},
  doi = {10.1109/tgrs.2016.2551327},
}
Figure