Fast and Structured Block-Term Tensor Decomposition For Hyperspectral Unmixing

Meng Ding, Xiao Fu, Xi-Le Zhao

Hyperspectral unmixing (HU) is a cornerstone of remote sensing, enabling the identification of material signatures (endmembers) and their spatial distributions (abundances) within a scene. While traditional Matrix Factorization (MF) has been the workhorse of HU, it often struggles with identifiability in the absence of "pure pixels." Tensor-based approaches, specifically the LL1 block-term decomposition, offer a mathematically robust alternative that guarantees identifiability under milder conditions. However, their adoption has been hindered by high computational costs and difficulty in incorporating physical constraints.

This paper introduces GradPAPA, a high-performance framework designed to unlock the potential of tensor decomposition for HU. By re-parameterizing the LL1 model and employing accelerated gradient projection, GradPAPA achieves significant speedups and superior accuracy compared to both matrix and existing tensor baselines.

The Problem with Traditional Unmixing

Standard HU methods rely on MF, which is frequently non-identifiable. To reach a unique solution, researchers often assume the presence of pure pixels—pixels containing only a single material. In complex, highly mixed environments, this assumption is unrealistic. The LL1 tensor model addresses this by exploiting the multi-way structure of hyperspectral data, but existing algorithms are notoriously slow. They typically use three-factor Alternating Least Squares (ALS) methods that require expensive matrix unfoldings and converge slowly, making it nearly impossible to integrate essential physical priors like spatial smoothness or the abundance sum-to-one (STO) constraint.

The GradPAPA Framework

The core innovation of GradPAPA is a two-block re-parameterization of the LL1 model. Instead of the traditional three-factor approach ($A, B, C$), the authors collapse the spatial factors into a single abundance map factor ($S_r$) for each material $r$. This map is explicitly constrained to be low-rank, reflecting the spatial correlation inherent in real-world scenes.

Illustration of the LL1 model and its application to hyperspectral data decomposition.

Key components of the GradPAPA method include:

  1. Gradient Projection Alternating Projection (GradPAPA): A first-order optimization scheme that alternates between updating endmembers and abundance maps.
  2. Two-Block Optimization: By reducing the number of blocks, the algorithm avoids the high-dimensional matrix operations that plague traditional LL1 solvers, drastically reducing the "flops" required per iteration.
  3. Structured Projections: An inner Alternating Projection (AP) solver handles nonnegativity, STO constraints, and low-rank constraints simultaneously. Unlike previous methods that used soft penalty terms, GradPAPA treats these as hard constraints, ensuring physical feasibility throughout the optimization process.
  4. Total Variation (TV) Regularization: To further exploit spatial context, the model incorporates a smoothed TV term, encouraging neighboring pixels to have similar compositions and resulting in cleaner abundance maps.

Experimental Results

The authors evaluated GradPAPA against synthetic and real-world datasets, including the Terrain, Urban, and Moffett Field scenes.

Speed and Convergence

One of the most striking results is the impact of Nesterov-style extrapolation. As shown in the convergence plots, the accelerated version of GradPAPA reaches a stable objective value in approximately 100 iterations, whereas the unaccelerated version requires nearly 800.

Convergence comparison between accelerated and unaccelerated GradPAPA under different low-rank constraints.

In practical terms, GradPAPA processed the Terrain dataset in 5–6 minutes, a task that took existing LL1 baselines over an hour. On the Moffett Field dataset, it demonstrated a 22x speedup over state-of-the-art tensor methods.

Accuracy and Visual Quality

GradPAPA consistently outperformed both matrix-based (e.g., SISAL) and tensor-based (e.g., MVNTF) competitors in terms of Mean Squared Error (MSE) for both endmembers and abundances. Visually, the abundance maps produced by GradPAPA are significantly less noisy, with sharper boundaries between different materials, thanks to the integrated TV regularization and low-rank constraints.

Conclusion

GradPAPA represents a significant leap forward for hyperspectral unmixing. By combining the theoretical strengths of tensor decomposition with an efficient, two-block optimization framework, it overcomes the speed and complexity barriers that previously limited tensor-based HU. While challenges remain—such as the nonconvexity of certain projections and the need for automated rank selection—GradPAPA provides a robust and scalable foundation for next-generation remote sensing analysis.

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