Semiblind Hyperspectral Unmixing in the Presence of Spectral Library Mismatches

Xiao Fu, Wing-Kin Ma, José M. Bioucas-Dias, Tsung-Han ChanIEEE TGRS 2016 · DOI

TL;DR

This paper tackles the challenge of spectral signature mismatches in hyperspectral unmixing (HU) by proposing a novel dictionary-adjusted, nonconvex sparsity-encouraging regression (DANSER) framework. DANSER effectively accounts for these mismatches by introducing dictionary correction variables and employing a more robust sparsity-promoting function, leading to improved unmixing performance.

What Problem Does This Paper Address?

Hyperspectral unmixing (HU) aims to identify the constituent materials within a hyperspectral image (HSI) by decomposing its pixels into spectral signatures. A popular approach is dictionary-aided semiblind HU, which leverages spectral libraries as dictionaries. However, a critical assumption in these methods is that the spectral signatures in the scene perfectly match those in the library. In reality, spectral signatures can vary due to environmental conditions, calibration errors, or differences in spectral resolution, leading to "spectral signature mismatches." These mismatches significantly degrade the performance of existing HU algorithms, including dictionary pruning methods like MUSIC. This paper addresses this crucial limitation by developing an HU framework that explicitly accounts for these spectral mismatches.

Methods

The core contribution is the Dictionary-Adjusted Nonconvex Sparsity-Encouraging Regression (DANSER) framework. Key methodological aspects include:

  • Modeling Mismatches: Instead of assuming perfect matches, DANSER models spectral signature mismatches as bounded error vectors ($\epsilon_n$) added to the dictionary signatures ($d_{k_n} = a_n + \epsilon_n$), with the constraint that the norm of these errors is bounded ($||\epsilon_n||_2 \le \delta$).
  • Dictionary Adjustment: The framework incorporates "dictionary correcting variables" ($e_k$) to adjust the dictionary signatures ($d'_k = d_k + e_k$) to better match the true spectral signatures in the scene.
  • Nonconvex Sparsity Promotion: DANSER utilizes a nonconvex $l_2/l_p$ quasi-norm ($0 < p < 1$) as the sparsity-promoting function for the abundance matrix. This is motivated by the observation that nonconvex quasi-norms can offer better sparsity performance in challenging scenarios, such as high-coherence dictionaries, compared to traditional convex norms.
  • Alternating Optimization Algorithm: A computationally efficient algorithm is developed for DANSER using an alternating optimization strategy. This involves iteratively updating the abundance matrix ($C$), the adjusted dictionary ($D'$), and auxiliary variables.
  • Robust Dictionary Pruning (RMUSIC): To address the issue of large dictionary sizes and the sensitivity of traditional MUSIC to mismatches, a robust MUSIC (RMUSIC) criterion is proposed. This method aims to identify spectral samples that are "close" to the true signatures rather than exactly matching them, making it more resilient to mismatches. RMUSIC reformulates the problem into a single-variable optimization, enabling efficient implementation.

Datasets

  • Synthetic Hyperspectral Images: Used for extensive simulations to evaluate the performance of the proposed algorithms under controlled conditions. Ground-truth spectra were randomly selected from a subset of the U.S.G.S. library.
  • AVIRIS Cuprite Data Set: A real-world hyperspectral dataset captured in Nevada, 1997, which is known to have calibration mismatches between real image spectra and the U.S.G.S. library. This dataset is used to validate the proposed methods on real-world data.

Key Insights and Ideas

  • Mismatches are Damaging: The paper strongly emphasizes that spectral signature mismatches are a significant practical problem that severely impacts existing dictionary-aided HU methods, including dictionary pruning.
  • Nonconvexity for Robustness: The use of nonconvex $l_p$ quasi-norms, while computationally more challenging, can lead to improved sparsity and better performance in the presence of dictionary imperfections.
  • Dictionary Adjustment is Key: Directly adjusting the dictionary to account for mismatches, rather than solely relying on robust estimation of abundances, is a powerful strategy.
  • RMUSIC for Pruning: The proposed RMUSIC method offers a more robust way to prune large spectral libraries, making it suitable for scenarios with spectral mismatches, unlike standard MUSIC.
  • Trade-offs in Dictionary Size: The paper highlights that while a larger pruned dictionary might seem safer, it can sometimes degrade sparse regression performance, suggesting a need for careful selection of the pruned dictionary size.

Key Results

  • Improved Unmixing Performance: RMUSIC-DANSER consistently yields higher Signal-to-Reconstruction Error (SRE) performance compared to MUSIC-CSR and DANSER with MUSIC pruning across various scenarios (different DMERs, SNRs, and number of materials).
  • Better Detection Probabilities: RMUSIC demonstrates significantly better detection probabilities for identifying ground-truth materials from the dictionary, especially under higher dictionary mismatch levels (DMERs), compared to standard MUSIC.
  • Sparser Solutions: DANSER, due to its nonconvex sparsity-promoting function, tends to identify a sparser set of active spectra compared to CSR.
  • Computational Efficiency: Despite the added complexity of dictionary adjustment and nonconvexity, the proposed alternating optimization algorithm for DANSER is designed to be computationally efficient with low per-iteration complexity.
  • Real Data Validation: Experiments on the AVIRIS Cuprite dataset show that RMUSIC-DANSER provides comparable or improved abundance maps for prominent materials compared to RMUSIC-CSR, and it identifies a number of materials close to established methods.

Limitations and Future Work

  • Initialization Sensitivity: The paper notes that since Problem (10) is nonconvex, a good initialization can help DANSER converge to a better solution. Using the CSR solution for initialization is suggested.
  • Computational Complexity: While the algorithm is efficient per iteration, operations can scale with the dictionary size ($K$). The authors suggest that using a smaller dictionary or pruning it in advance is motivated, which is addressed by RMUSIC.
  • Parameter Selection: The choice of parameters like $\lambda$, $\tau$, $\mu$, $\epsilon$, and $p$ can influence performance, and optimal selection might require further investigation or adaptive methods.
  • Future Work: The paper implicitly suggests that further exploration of nonconvex sparsity-promoting functions and more advanced dictionary adjustment strategies could lead to even better performance. The robustness of the RMUSIC criterion to different types and levels of spectral variability could also be a direction for future research.

Summary updated

BibTeX
@article{fu2016semiblind,
  author = {Xiao Fu and Wing-Kin Ma and José M. Bioucas-Dias and Tsung-Han Chan},
  title = {{Semiblind Hyperspectral Unmixing in the Presence of Spectral Library Mismatches}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2016},
  doi = {10.1109/tgrs.2016.2557340},
}
Figure