Abstract
Hyperspectral unmixing (HU) is the process of decomposing mixed pixels into pure spectral signatures (endmembers) and their fractional proportions (abundances). While Nonnegative Matrix Factorization (NMF) is a popular technique for HU, it often struggles with local minima and sensitivity to initialization. This paper presents MCG-NMF, a self-supervised framework that leverages multiple clustering tasks to automatically extract prior knowledge from hyperspectral data. By integrating K-means clustering directly into the NMF optimization loop, the method generates Proximate Endmember Signatures (PESs) and Proximate Abundance Distributions (PADs) to guide the estimation process. Evaluations on synthetic and real-world datasets show that MCG-NMF provides more stable and accurate unmixing results without requiring external spectral libraries.
Introduction
Hyperspectral sensors capture data across hundreds of narrow spectral bands. Due to limited spatial resolution, individual pixels often represent a mixture of multiple materials, such as soil, vegetation, and water. The goal of Hyperspectral Unmixing (HU) is to identify these constituent materials (endmembers) and their relative concentrations (abundances).
Nonnegative Matrix Factorization (NMF) is widely used for this task but faces three significant challenges:
- Non-convexity: Optimization frequently converges to local minima.
- Sensitivity: Performance is highly dependent on initial values.
- The "Prior" Gap: While "priors" (known spectral signatures) can improve NMF, external libraries often fail to account for site-specific atmospheric or lighting conditions.
MCG-NMF addresses these issues by "mining" priors directly from the image cube using a self-supervised approach.
Methods: The MCG-NMF Framework
The core innovation of this work is the Multiple Clustering Guided framework. Rather than treating clustering as a separate pre-processing step, the authors embed it within the NMF optimization process to provide continuous guidance.
The framework consists of several key components:
- Proximate Endmember Signatures (PESs): The algorithm applies K-means clustering to the hyperspectral data. For each cluster, it identifies a "mean pixel" and its $K$-nearest neighbors. These are treated as "proximate" versions of the true endmembers, forming the PES matrices.
- Proximate Abundance Distributions (PADs): Based on the clusters and initial least-squares estimates, the model generates a coarse map of material locations.
- Self-Supervised Constraints: PESs and PADs are incorporated as regularization terms in the NMF objective function. The NMF is penalized if its estimated endmembers deviate significantly from the cluster-derived guidance.
- Iterative Interaction: The NMF and clustering processes are interdependent. Periodically, NMF results are used to re-initialize the K-means clustering, which then provides refined guidance back to the NMF. This mutual reinforcement helps the model escape local minima and find more physically meaningful solutions.
Experimental Evaluation
The authors evaluated MCG-NMF against state-of-the-art (SOTA) methods, including GLNMF and MVCNMF, using both synthetic and real-world datasets:
- Synthetic Data: Created using the USGS spectral library to test robustness against varying noise levels (SNR) and mixing degrees.
- Washington DC Mall: A real-world dataset from the HYDICE sensor featuring urban endmembers like roofs, water, and grass.
- Cuprite Dataset: The benchmark for mineral unmixing, captured by the AVIRIS sensor, containing 10 distinct mineral types.
Key Results
- Superior Accuracy: MCG-NMF consistently achieved the lowest average Spectral Angle Distance (SAD) on the Washington DC Mall data, indicating highly accurate signature estimation.
- Robustness to Noise: In synthetic tests, the method maintained stable performance even at high noise levels (15dB), where geometric-based methods typically fail.
- Mineral Mapping: On the Cuprite dataset, abundance maps were visually cleaner and more consistent with geological ground truth compared to other NMF variants.
- Computational Efficiency: While more intensive than simple geometric algorithms, MCG-NMF is significantly faster than complex tensor-based methods, making it a practical choice for large-scale applications.
Conclusion
MCG-NMF introduces a robust self-supervised approach to hyperspectral unmixing by exploiting the inherent spatial and spectral locality of materials through clustering. By using clustering as a surrogate task, the model effectively bridges the "prior gap" and overcomes the traditional limitations of NMF. Future work may involve extending the framework to nonlinear mixing models and incorporating more sophisticated spatial contexture information, such as superpixel segmentation.
Summary updated