Abstract
To support high-level analysis of spaceborne imaging spectroscopy (hyperspectral) imagery, spectral unmixing has been gaining significance. However, spectral variability caused by illumination, topography, and atmospheric effects makes it difficult to accurately estimate abundance maps. Classical methods like the Linear Mixing Model (LMM) fail to robustly handle complex spectral variability. We propose a subspace-based unmixing model using a low-rank learning strategy, called Subspace Unmixing with Low-Rank Attribute embedding (SULoRA). Unlike previous approaches that unmix signatures in the original space, SULoRA jointly estimates subspace projections and abundance maps to find a raw subspace more suitable for unmixing. By projecting data into a low-rank subspace, SULoRA effectively addresses various spectral variabilities. We adopt an Alternating Direction Method of Multipliers (ADMM) based algorithm to solve the optimization problem. Experiments on synthetic and real datasets demonstrate the superiority of SULoRA over state-of-the-art methods.
Introduction
In hyperspectral imaging, a single pixel often covers a large area containing multiple materials, such as grass, soil, and water. Spectral unmixing is the process of decomposing these "mixed pixels" into their constituent pure materials (endmembers) and their corresponding proportions (abundance maps). The standard Linear Mixing Model (LMM) assumes material spectral signatures are constant. However, real-world spectral variability is inevitable due to environmental factors and sensor noise. While models like the Extended LMM (ELMM) attempt to account for scaling factors, they often struggle with complex, non-linear variations. SULoRA addresses this by treating variability as a structural perturbation that can be filtered out in a learned subspace.
Methodology: The SULoRA Framework
SULoRA shifts the unmixing process from the original high-dimensional spectral space to a robust "raw" subspace. The framework is built on three pillars:
- Subspace Projection: The model learns a projection matrix ($\Theta$) that maps raw data into a subspace where spectral variability is minimized.
- Low-Rank Attribute Embedding: By enforcing a low-rank constraint, the model assumes that pixels of the same material class are strongly correlated. This allows the system to treat spectral variability as complex noise that can be "projected away."
- Joint Optimization: SULoRA employs a closed-loop system to jointly estimate the subspace and abundance maps, ensuring structural consistency with the original data.
To solve the resulting non-convex optimization problem, the authors utilize an ADMM-based algorithm, breaking the problem into manageable sub-problems such as Singular Value Thresholding.
Experimental Evaluation
The model was evaluated against eight state-of-the-art methods (including FCLSU, SUnSAL, and ELMM) across three datasets:
- Synthetic Data: Created with five endmembers and introduced scaling factors (0.75 to 1.25) plus 25 dB of Gaussian noise. SULoRA achieved the lowest abundance root mean square error (aRMSE of 0.0220).
- Urban Area (HYDICE): A real-world dataset of Copperas Cove, Texas. SULoRA achieved an Overall Accuracy (OA) of 86.71%, significantly outperforming the next best method, SSUnSAL (71.26%), and the variability-focused ELMM (62.41%).
- MUUFL Gulfport Campus: A complex dataset with 64 spectral bands and 8 classes, testing the model's ability to handle high-resolution spatial structures.
Key Insights and Results
SULoRA's primary advantage lies in its treatment of variability as "complex noise" rather than trying to model every physical variation explicitly. This flexibility allows it to produce cleaner abundance maps with higher contrast, successfully identifying materials like asphalt and grass that other models often confuse due to shadows. Furthermore, SULoRA demonstrated remarkable stability against sparse noise, where other methods saw performance plummet.
Conclusion and Future Work
SULoRA provides a robust general framework for hyperspectral unmixing in the presence of complex spectral variability. While highly effective, the current implementation relies on pre-extracted endmembers (e.g., via VCA) and involves $O(D^2N)$ computational complexity due to SVD operations. Future research will focus on integrating joint endmember estimation directly into the subspace learning process and optimizing the algorithm for extremely large-scale satellite imagery.
Summary updated