Low-rank and Sparse NMF for Joint Endmembers' Number Estimation and Blind Unmixing of Hyperspectral Images

Paris V. Giampouras, Athanasios A. Rontogiannis, Konstantinos D. Koutroumbas

Abstract

Estimation of the number of endmembers in a scene is a critical yet challenging task in hyperspectral unmixing. Traditional methods treat endmember number estimation and unmixing as separate, sequential steps, which can be computationally expensive and prone to error propagation. This paper proposes a unified framework that performs both tasks simultaneously. By formulating the problem as a low-rank and sparse Nonnegative Matrix Factorization (NMF) task with a sophisticated $\ell_2/\ell_1$ norm penalty, the algorithm can automatically prune redundant endmembers from an initial overestimation. Experimental results on simulated and real Martian hyperspectral data demonstrate the effectiveness of this joint approach.

Introduction

Hyperspectral imaging captures detailed spectral information across hundreds of bands, enabling the identification of materials within a scene. However, due to limited spatial resolution, individual pixels often represent a mixture of multiple materials, known as endmembers. The process of decomposing these mixed pixels into their constituent spectral signatures and corresponding abundance fractions is called "unmixing."

A major hurdle in unsupervised unmixing is determining the exact number of endmembers present. Most existing algorithms require this number as a predefined input. If the estimate is inaccurate, the resulting spectral signatures and abundance maps will be flawed. This paper introduces a method to estimate the model order and perform unmixing in a single, integrated optimization step.

Methodology: Overestimate and Prune

The core idea of the proposed approach is to start with an overestimation of the number of endmembers and then use a specialized penalty to "prune" the unnecessary ones.

The Low-Rank and Sparse NMF Framework

The problem is modeled using the Linear Mixing Model (LMM) and cast as an NMF optimization. To achieve joint estimation and unmixing, the authors introduce a group-sparsity ($\ell_2/\ell_1$) norm penalty. Unlike the standard nuclear norm, which can induce unwanted linear dependence between signatures, the $\ell_2/\ell_1$ penalty encourages entire columns of the endmember and abundance matrices to go to zero if they do not contribute significantly to explaining the data.

Optimization Algorithm

The resulting cost function is non-smooth and non-separable, making it difficult to solve with standard methods. The authors developed an alternating proximal Newton-type algorithm that utilizes quadratic approximations and a smoothing constant to ensure convergence.

Algorithm 1 and recovered endmember signatures for simulated data.

Experimental Evaluation

The algorithm was tested on both synthetic and real-world datasets to verify its robustness.

  • Simulated Data: A dataset was created using four signatures from the USGS library. When initialized with ten endmembers ($r=10$), the algorithm successfully identified and removed the six redundant components, converging to the true count of four. The recovered signatures closely match the ground truth.
  • Real Data (Mars): The method was applied to OMEGA hyperspectral imagery of the Martian South Polar Cap. Starting with an initial guess of eight endmembers, the algorithm correctly pruned the model down to the three known primary components: H2O ice, CO2 ice, and mineral dust. The results aligned with established planetary science literature.

Key Insights

  • Robustness to Overestimation: By intentionally overestimating the rank, the algorithm avoids the pitfalls of under-fitting and uses the $\ell_2/\ell_1$ penalty as an "elegant pruning shear."
  • Signature Integrity: The proposed penalty shrinks the rank without blurring or inducing linear dependence between the extracted spectral signatures, a common issue with nuclear norm regularization.

Conclusion

This research demonstrates that endmember number estimation and blind unmixing can be effectively combined into a single optimization framework. By leveraging low-rank and sparse NMF with a group-sparsity penalty, the proposed method reduces computational complexity and improves the reliability of hyperspectral data analysis in complex environments, such as planetary exploration.

Summary updated

Abstract

Estimation of the number of endmembers existing in a scene constitutes a critical task in the hyperspectral unmixing process. The accuracy of this estimate plays a crucial role in subsequent unsupervised unmixing steps i.e., the derivation of the spectral signatures of the endmembers (endmembers' extraction) and the estimation of the abundance fractions of the pixels. A common practice amply followed in literature is to treat endmembers' number estimation and unmixing, independently as two separate tasks, providing the outcome of the former as input to the latter. In this paper, we go beyond this computationally demanding strategy. More precisely, we set forth a multiple constrained optimization framework, which encapsulates endmembers' number estimation and unsupervised unmixing in a single task. This is attained by suitably formulating the problem via a low-rank and sparse nonnegative matrix factorization rationale, where low-rankness is promoted with the use of a sophisticated $\ell_2/\ell_1$ norm penalty term. An alternating proximal algorithm is then proposed for minimizing the emerging cost function. The results obtained by simulated and real data experiments verify the effectiveness of the proposed approach.
BibTeX
@inproceedings{giampouras2017low,
  author = {Paris V. Giampouras and Athanasios A. Rontogiannis and Konstantinos D. Koutroumbas},
  title = {{Low-rank and Sparse NMF for Joint Endmembers' Number Estimation and Blind Unmixing of Hyperspectral Images}},
  booktitle = {EUSIPCO},
  year = {2017},
  url = {https://openreview.net/forum?id=Scb0cfxPgu},
}
Figure