Abstract
Hyperspectral unmixing aims to decompose mixed pixels into pure endmember signatures and their corresponding abundances. Selecting the optimal number of endmembers is a critical challenge: too few lead to high reconstruction error (residuum), while too many result in numerical instability (high condition number). This paper introduces the condition-residuum diagram, a visualization tool to navigate this trade-off. Furthermore, a greedy reduction algorithm is proposed to prune redundant endmembers from an over-complete set. Experiments on standard datasets demonstrate that the "L-shape" of the condition-residuum curve effectively identifies the intrinsic dimensionality of the scene, often outperforming traditional direct extraction methods.
Introduction
Hyperspectral imaging captures a wealth of spectral information, allowing for the identification of materials at a sub-pixel level. Unmixing is the standard technique used to estimate the fractional abundances of pure materials, known as endmembers (EMs), within each pixel. However, determining the correct number of EMs is non-trivial.
Researchers face a "Goldilocks" problem when choosing the number of EMs. A model with insufficient EMs suffers from high residuum (reconstruction error), failing to capture the scene's complexity. Conversely, an over-complete set of EMs introduces redundancy. When EMs are highly correlated, the linear system for abundance estimation becomes ill-conditioned. In such cases, the condition number ($\kappa$) of the EM matrix spikes, making the solution extremely sensitive to noise and leading to physically implausible abundance estimates.
The Condition-Residuum Diagram
To address this, we propose the condition-residuum diagram. This 2D plot maps the reconstruction error (RMSE) on the x-axis against the condition number ($\kappa$) on the y-axis. This visualization allows researchers to quantify the stability-accuracy trade-off. A stable system has a low $\kappa$, while a high $\kappa$ indicates that the EMs are nearly linearly dependent. By observing how these metrics change as EMs are added or removed, one can identify the "sweet spot" for a given dataset.
Greedy Endmember Reduction
Rather than estimating the number of EMs a priori, our approach starts with an "over-complete" set—typically twice the expected number. We then apply a greedy reduction algorithm that iteratively removes the EM whose absence most improves the balance between stability and accuracy. This is guided by a multi-objective function that seeks to maximize the gain in stability while minimizing the loss in reconstruction accuracy. This pruning process continues until a single EM remains, generating a trajectory on the condition-residuum diagram.
Experimental Evaluation
The method was evaluated using five benchmark hyperspectral datasets. These datasets provide a variety of spatial and spectral resolutions, allowing for a robust assessment of the proposed visualization and reduction techniques.
Results and Discussion
A key finding is the characteristic "L-shape" of the resulting curves across different datasets and extraction algorithms (e.g., OSP, N-FINDR, VCA). The "elbow" or "kink" of this L-curve typically corresponds to the reference number of endmembers ($m^{ref}$) established in literature.
As shown in Figure 1, the reference points consistently align with the region where the residuum begins to plateau while stability remains manageable. Interestingly, popular automated estimators like HySime often suggest a much higher number of EMs, placing them deep into the unstable vertical branch of the L-curve. This suggests that our diagram serves as a valuable visual "sanity check" for dimensionality estimation.
Furthermore, starting with an over-complete set and pruning down often yields a more stable subset than extracting the "correct" number of EMs directly. This is because the initial over-complete set is more likely to capture all relevant spectral signatures, which can then be refined through the greedy selection process.
Conclusion
The condition-residuum diagram provides a robust framework for evaluating hyperspectral unmixing results. By explicitly considering numerical stability alongside reconstruction error, researchers can produce more reliable abundance maps. While the greedy reduction approach is computationally more intensive than direct extraction, its ability to find stable, high-fidelity subsets makes it a powerful tool for complex scenes. Future work will focus on optimizing the algorithm's complexity and exploring automated methods for tuning the trade-off between accuracy and stability.
Summary updated