Abstract
Several approaches have been proposed to solve the spectral unmixing problem in hyperspectral image analysis. Among them, the use of sparse regression techniques aims to characterize the abundances in pixels based on a large library of spectral signatures known a priori. Recently, the integration of image spatial-contextual information significantly enhanced the performance of sparse unmixing. In this work, we propose a computationally efficient multiscale representation method for hyperspectral data adapted to the unmixing problem. The proposed method is based on a hierarchical extension of the SLIC oversegmentation algorithm constructed using a robust homogeneity testing. The image is subdivided into a set of spectrally homogeneous regions formed by pixels with similar characteristics (superpixels). This representation is then used to provide prior spatial regularity information for the abundances of materials present in the scene, improving the conditioning of the unmixing problem. Simulation results illustrate that the method is capable of estimating abundances with high quality and low computational cost, especially in noisy scenarios.
Introduction
Hyperspectral sensors capture data across hundreds of narrow spectral bands, providing rich information for material identification. However, their spatial resolution is often limited, leading to "mixed pixels" where multiple materials contribute to a single pixel's spectral signature. Spectral unmixing aims to decompose these pixels into their constituent materials (endmembers) and their relative proportions (abundances).
Sparse unmixing leverages large, pre-existing spectral libraries to identify these endmembers. While powerful, this approach is often ill-posed; the high dimensionality of the libraries makes the solution sensitive to noise. Incorporating spatial context—the assumption that neighboring pixels likely share similar material compositions—can stabilize the problem, but often at a high computational cost.
Methods: Homogeneity-based Multiscale Representation
The proposed Homogeneity-based Multiscale sparse Unmixing Algorithm (HMUA) addresses these challenges by creating an adaptive, hierarchical representation of the hyperspectral cube.
Adaptive Oversegmentation
Unlike traditional methods that use fixed-size superpixels, HMUA employs a hierarchical extension of the Simple Linear Iterative Clustering (SLIC) algorithm. It begins with a coarse initial segmentation and iteratively refines only those regions that fail a homogeneity test.
Robust Homogeneity Testing
To determine if a superpixel requires further subdivision, HMUA calculates the Euclidean distance between each pixel and the median spectral vector of the region. Using the median instead of the mean provides robustness against outliers and sensor noise, preventing unnecessary subdivisions in noisy but otherwise uniform areas.
Multiscale Unmixing Strategy
The unmixing process occurs in two primary stages:
- Coarse Scale: The algorithm solves the sparse unmixing problem for the average spectra of the superpixels, providing a computationally efficient global view.
- Fine Scale: The coarse results serve as a spatial prior to regularize the pixel-level unmixing, ensuring local consistency while preserving fine details.
Experimental Results
The researchers evaluated HMUA against synthetic datasets (DC1, DC2, DC3) and real-world imagery (Samson and Jasper Ridge).
Performance and Noise Resilience
HMUA demonstrates superior performance in low Signal-to-Noise Ratio (SNR) environments. In tests with 20 dB SNR, HMUA consistently outperformed standard MUA and state-of-the-art algorithms like $S^2WSU$. For instance, on the complex DC3 dataset, HMUA achieved a 5% improvement in Signal-to-Reconstruction Error (SRE) over MUA.
Computational Efficiency
A key advantage of HMUA is its speed. While the $S^2WSU$ algorithm required over 230 seconds to process synthetic scenes, HMUA completed the same tasks in 11 to 21 seconds—a significant reduction in computational overhead without sacrificing accuracy.
Conclusion
HMUA provides a robust and efficient framework for hyperspectral unmixing by intelligently adapting to the spatial complexity of the scene. By focusing computational resources on non-homogeneous regions and utilizing a noise-resistant median-based test, it achieves high-fidelity abundance maps even in challenging conditions. Future research may focus on automated hyperparameter selection or integrating this multiscale approach into deep learning architectures for hyperspectral analysis.
Summary updated