Fast and Structured Block-Term Tensor Decomposition for Hyperspectral Unmixing

Meng Ding, Xiao Fu, Xi-Le Zhao

Abstract

Hyperspectral unmixing (HU) is a critical task in remote sensing, aiming to decompose mixed pixels into constituent endmembers and their corresponding abundance maps. While tensor-based methods like the LL1 decomposition offer better identifiability than matrix factorization, they often suffer from high computational complexity and difficulty in incorporating physical priors. This paper presents GradPAPA, a novel framework based on a two-factor reparameterization of the LL1 model. By alternating between endmember and abundance updates and employing Nesterov-style acceleration, GradPAPA achieves significant speedups and improved accuracy over state-of-the-art methods while naturally enforcing physical constraints such as non-negativity and spatial smoothness.

Introduction

Hyperspectral imaging captures a scene across hundreds of narrow spectral bands, providing a unique "fingerprint" for every material on the ground. However, due to limited spatial resolution, a single pixel often contains a mixture of several materials, such as soil, grass, and water. Hyperspectral Unmixing (HU) is the process of decomposing these mixed pixels into their constituent materials (endmembers) and their proportions (abundances).

Conceptual decomposition of a hyperspectral data cube into endmembers and their corresponding abundance maps.

Traditional unmixing often relies on Matrix Factorization (MF), which can produce multiple mathematically valid but physically incorrect solutions without specific conditions like the existence of "pure pixels." The LL1 tensor decomposition model addresses this by treating the hyperspectral image as a 3D cube (tensor), exploiting spatial structure to ensure solution uniqueness. However, existing LL1 algorithms use a "three-factor" parameterization that is computationally heavy and difficult to regularize with physical priors, leading to slow convergence and high complexity per iteration.

Methodology: The GradPAPA Framework

The authors introduce GradPAPA (Gradient Projection Alternating Projection Algorithm), which utilizes a two-factor reparameterization to recast the tensor problem into a rank-constrained MF formulation. This innovation allows the algorithm to incorporate physical rules more naturally and efficiently.

Key technical components include:

  • Two-Block Optimization: The algorithm alternates between updating endmembers and abundance maps, mapping the math directly to the physical components of the image.
  • Alternating Projection (AP) Solver: To handle complex, overlapping constraints—such as non-negativity, the "sum-to-one" (STO) abundance rule, and low-rank (LR) constraints—the authors developed a fast inner-loop solver.
  • Extrapolation-Based Acceleration: A Nesterov-style acceleration technique speeds up convergence without increasing the per-iteration cost.
  • Smoothed TV Regularization: Total Variation (TV) ensures that neighboring pixels have similar material compositions, reflecting real-world geography.

By exploiting the specific structure of hyperspectral data, the authors reduced the computational complexity from $O(I^4)$ to $O(I^3R)$ per iteration compared to general tensor decomposition methods.

Experiments and Results

The researchers evaluated GradPAPA against six baseline methods using synthetic data, semireal data (Terrain and Urban), and real-world data from the Moffett Field AVIRIS sensor.

Convergence plots showing the objective value over iterations, highlighting the efficiency gains from Nesterov-style acceleration.

GradPAPA demonstrated massive speedups; on the Terrain dataset, it finished in about 6 minutes, while existing LL1-based competitors took over an hour. On real data, it was up to 22 times faster than some baselines. Furthermore, GradPAPA consistently achieved the lowest Spectral Angle Distance (SAD) and Mean Squared Error (MSE), especially in "highly mixed" scenarios. Qualitative results on the Moffett Field data showed cleaner abundance maps with sharper boundaries between materials compared to classic matrix-based unmixing.

Conclusion

GradPAPA represents a significant advancement in hyperspectral unmixing, offering a faster and more physically grounded approach than previous LL1-based methods. While the current model assumes a linear mixture, future work will explore extensions to account for nonlinear light scattering and improve robustness against pixel outliers.

Summary updated

BibTeX
@article{ding2023fast,
  author = {Meng Ding and Xiao Fu and Xi-Le Zhao},
  title = {{Fast and Structured Block-Term Tensor Decomposition for Hyperspectral Unmixing}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2023},
  doi = {10.1109/jstars.2023.3238653},
}
Figure