Abstract
Endmember (EM) variability has an important impact on the performance of hyperspectral image (HI) analysis algorithms. Recently, extended linear mixing models have been proposed to account for EM variability in the spectral unmixing (SU) problem. The direct use of these models has led to severely ill-posed optimization problems. Different regularization strategies have been considered to deal with this issue, but none so far has consistently exploited the information provided by the existence of multiple pure pixels often present in HIs. In this work, we propose to break the SU problem into a sequence of two problems. First, we use pure pixel information to estimate an interpolated tensor of scaling factors representing spectral variability. This is done by considering the spectral variability to be a smooth function over the HI and confining the energy of the scaling tensor to a low-rank structure. Afterwards, we solve a matrix-factorization problem to estimate the fractional abundances using the variability scaling factors estimated in the previous step, what leads to a significantly more well-posed problem. Simulations with synthetic and real data attest the effectiveness of the proposed strategy.
Introduction
Hyperspectral imaging (HI) captures hundreds of narrow spectral bands, enabling precise material identification. Spectral Unmixing (SU) is the process of decomposing a pixel into its constituent "endmembers" (pure materials) and their corresponding "abundances" (proportions). Traditional Linear Mixing Models (LMM) assume constant spectral signatures, but real-world factors like topography and illumination cause significant variability. While models like the Generalized Linear Mixing Model (GLMM) address this, they often result in ill-posed optimization problems due to the high number of unknowns.
Proposed Methodology
This work introduces a hybrid strategy that decouples the unmixing process into two sequential steps to improve stability and accuracy:
- Estimating the Scaling Tensor: The method leverages "pure pixels" as anchors to estimate a 4D interpolated scaling tensor ($\Psi$). This tensor models how each spectral band of each material scales across the image. Two constraints are applied: smoothness (modeled via a low-rank tensor structure) and proximity to unity (reflecting the assumption that variability is typically small).
- Abundance Estimation: With variability factors pre-calculated, the unmixing becomes a more stable matrix-factorization problem. Spatial regularization (Total Variation) is included to ensure realistic, cohesive abundance maps.
Experimental Results
The proposed method was evaluated against standard algorithms, including FCLS, PLMM, ELMM, and GLMM, using both synthetic and real-world datasets.
Quantitative Performance
On synthetic data, the proposed method achieved a Root Mean Squared Error for abundances ($RMSE_A$) of 0.0233, a significant improvement over the GLMM's 0.0343. The results across various metrics are summarized in the table below.
Qualitative Analysis
The Houston dataset, featuring vegetation, concrete, asphalt, and metal roofs, was used for real-world validation. The proposed method produced cleaner abundance maps, particularly in complex areas like stadium stands where distinguishing between concrete and vegetation is challenging.
Conclusion
By utilizing pure pixels to anchor variability estimation and modeling spectral changes as a low-rank tensor, the proposed approach transforms a complex, ill-posed problem into a manageable two-step process. While computationally more intensive than some alternatives, the gains in accuracy and visual fidelity make it a robust solution for hyperspectral analysis. Future work will focus on scenarios where pure pixels are scarce.
Summary updated