Abstract
Hyperspectral unmixing (HSU) is a critical task in remote sensing, aimed at decomposing mixed pixels into pure spectral signatures (endmembers) and their fractional proportions (abundances). Traditional methods often struggle with high-dimensional search spaces and the neglect of spatial correlations between neighboring pixels. This paper introduces ICoNMF-TV, a robust framework that combines Non-Negative Matrix Factorization (NMF) with collaborative row sparsity and Total Variation (TV) regularization. By enforcing spatial consistency and pruning the material search space, ICoNMF-TV significantly improves unmixing accuracy and noise robustness.
Introduction
Hyperspectral sensors capture data across hundreds of narrow spectral bands, providing rich information for material identification. However, due to limited spatial resolution, individual pixels often represent a mixture of multiple materials. The Linear Mixing Model (LMM) is commonly used to address this, but it faces two significant hurdles:
- High Dimensionality: The vast number of potential material combinations makes finding an optimal solution computationally difficult.
- Spatial Disregard: Most algorithms treat pixels independently, ignoring the physical reality that adjacent pixels often share similar material compositions.
Methodology
The proposed ICoNMF-TV algorithm addresses these limitations by augmenting standard NMF with two key constraints:
Collaborative Row Sparsity
This constraint leverages the fact that only a small subset of a large spectral library is typically present in any given scene. By promoting row sparsity in the abundance matrix, the algorithm effectively prunes the search space, focusing only on relevant endmembers.
Total Variation (TV) Regularization
TV regularization serves as a spatial prior, penalizing sharp, erratic changes in abundance between adjacent pixels. This encourages the generation of smooth, physically plausible abundance maps, which is particularly beneficial in noisy environments.
The resulting non-convex objective function is solved using Proximal Alternating Optimization (PAO), which iteratively updates the endmember and abundance matrices until convergence.
Experimental Evaluation
The authors evaluated ICoNMF-TV using both simulated and real-world datasets, including the USGS spectral library, Jasper Ridge, and the Cuprite mineralogical site. These tests allowed for precise control over noise levels and performance assessment on complex natural terrain.
Key Results and Discussion
ICoNMF-TV consistently outperformed baseline methods like R-CoNMF, NCLS, and SUnSAL. In a 20dB noise simulation, ICoNMF-TV achieved a Signal-to-Reconstruction Error (SRE) of 10.16, compared to 8.05 for R-CoNMF.
Visually, the abundance maps produced by ICoNMF-TV exhibit clearer boundaries and significantly fewer noise artifacts. The inclusion of the TV term proved vital for maintaining performance in high-noise scenarios, where spectral data alone is often insufficient for accurate unmixing. Furthermore, the method demonstrated improved efficiency by optimizing a "right-sized" matrix rather than iterating through massive external libraries.
Conclusion
ICoNMF-TV provides a powerful approach to hyperspectral unmixing by effectively balancing statistical factorization with spatial and collaborative priors. While the method is computationally more intensive than simple NMF, its gains in accuracy and robustness are substantial. Future research will focus on optimizing computational speed and exploring higher-order neighborhood systems to better capture complex spatial features.
Summary updated