Abstract
Kernel-based nonlinear mixing models have been applied to unmix spectral information of hyperspectral images when the type of mixing occurring in the scene is too complex or unknown. Such methods, however, usually require the inversion of matrices of sizes equal to the number of spectral bands. Reducing the computational load of these methods remains a challenge in large scale applications. This paper proposes a centralized method for band selection (BS) in the reproducing kernel Hilbert space (RKHS). It is based upon the coherence criterion, which sets the largest value allowed for correlations between the basis kernel functions characterizing the unmixing model. We show that the proposed BS approach is equivalent to solving a maximum clique problem (MCP), that is, searching for the biggest complete subgraph in a graph. Furthermore, we devise a strategy for selecting the coherence threshold and the Gaussian kernel bandwidth using coherence bounds for linearly independent bases. Simulation results illustrate the efficiency of the proposed method.
Introduction
Hyperspectral sensors capture hundreds of narrow spectral bands, providing a unique "fingerprint" for materials on Earth. Hyperspectral Unmixing (HU) is the process of decomposing a pixel's total reflectance into the proportions (abundances) of pure materials (endmembers) it contains. While linear models are simple, real-world scenes often involve complex light scattering—particularly in dense vegetation or urban areas—requiring nonlinear unmixing techniques.
Kernel-based methods represent the state-of-the-art for nonlinear unmixing, but they suffer from a significant computational bottleneck: they typically require inverting matrices with dimensions proportional to the number of spectral bands ($L \times L$). With modern sensors providing hundreds of bands, this becomes computationally prohibitive. This report details a novel Band Selection (BS) method specifically designed for nonlinear unmixing in the Reproducing Kernel Hilbert Space (RKHS).
Methods: Clique Coherence-based Band Selection (CCBS)
The proposed approach, Clique Coherence-based Band Selection (CCBS), frames the selection of optimal spectral bands as a global optimization problem. The process involves mapping spectral bands into a high-dimensional feature space (RKHS) and evaluating their redundancy.
The Coherence Criterion and Graph Construction
The method uses a "coherence" metric to measure the similarity between bands in the RKHS. An undirected graph is constructed where each vertex represents a spectral band. An edge is established between two vertices only if their coherence is below a predefined threshold ($\mu_0$). This thresholding ensures that an edge exists only between bands that are "sufficiently different" or non-redundant.
Solving the Maximum Clique Problem
The task of finding the optimal subset of bands is equivalent to solving the Maximum Clique Problem (MCP). A clique is a subset of vertices where every vertex is connected to every other vertex in the subset. By finding the maximum clique, the algorithm identifies the largest possible set of bands that are all mutually dissimilar.
The researchers utilized the maxCQL algorithm to solve the MCP. Additionally, they developed a strategy to automatically calibrate the coherence threshold and kernel bandwidth based on the desired dictionary size.
Experimental Evaluation
The CCBS method was evaluated using a variety of datasets to ensure robustness:
- Synthetic Data: Generated using mineral signatures from the USGS Cuprite library and urban signatures from the ROSIS Pavia University dataset.
- Laboratory Data: The RELAB dataset, containing controlled mineral mixtures (e.g., Olivine and Enstatite), provided ground truth for abundance estimation.
- Real-World Imagery: AVIRIS scenes of the Cuprite mining field (Nevada) and ROSIS scenes of Pavia University.
Key Results and Discussion
The experimental results highlight two primary advantages of the MCP-based approach:
- Computational Efficiency: For the Cuprite dataset (420 bands), the proposed method reduced the average CPU time for unmixing from approximately 301 seconds to just 3 seconds—a 100x speedup.
- Accuracy Preservation: Despite using only about 10% of the available bands, the Root Mean Square Error (RMSE) in abundance estimation remained nearly identical to the full-band approach (e.g., 0.0637 vs 0.0680).
Unlike "greedy" coherence-based methods that select bands sequentially, the CCBS method performs a global optimization. This results in dictionaries with lower overall coherence and improved unmixing stability compared to standard techniques like Kernel k-means.
Conclusion and Future Work
By framing band selection as a Maximum Clique Problem, this research demonstrates that a carefully selected subset of ~40 bands can capture nearly all the information required for accurate nonlinear unmixing. While the MCP is NP-complete, current solvers like maxCQL are highly effective for the scale of modern hyperspectral sensors. Future research may explore heuristic or metaheuristic algorithms to accommodate future sensors with thousands of spectral bands or to meet real-time processing requirements.
Summary updated