Introduction
Hyperspectral unmixing is a critical task in remote sensing, aimed at decomposing mixed pixels into pure constituent materials, known as endmembers, and their corresponding fractional abundances. While traditional endmember finding (EF) algorithms have relied primarily on spectral information, they often struggle with spectral variation caused by noise, lighting, or topography. Recent research highlights that spatial context—the relationship between a pixel and its neighbors—is a vital piece of missing information. This paper proposes three novel Spatial Exploiting (SE) strategies to enhance spectral-based EF (sEF) algorithms by integrating this spatial distribution data.
Proposed Spatial Exploiting Strategies
The core idea is to "clean" or revise spectral signatures by leveraging local neighborhoods. The authors introduce three specific methods:
- Local Linear Representation (LLR): Reconstructs a pixel as a linear combination of its neighbors, effectively averaging out local noise while preserving the underlying spectral shape.
- Local Sparse Representation (LSR): Employs a sparsity constraint (L1-norm) to ensure the pixel is reconstructed using only the most relevant neighbors, which is crucial in heterogeneous areas.
- SVD-Based Strategy: Operates in the feature domain by performing Singular Value Decomposition on a local neighborhood and using dominant singular vectors to filter out noise.
A key innovation is the "Anomaly Switch." To prevent the "oversmoothing" common in spatial methods—which can erase rare, small-scale targets—the algorithm compares the revised pixel to the original using Spectral Angle Distance. If the difference is too large, the pixel is flagged as an anomaly, and the original signature is preserved. These strategies are versatile, functioning as either a preprocessing step for the entire image or a postprocessing step to refine identified endmembers.
Conceptual Impact on the Data Simplex
In the spectral feature space, hyperspectral data typically forms a simplex where the vertices represent pure endmembers. Noise and spectral variation cause pixels to deviate from this ideal structure, making it difficult for sEF algorithms to identify the true vertices.
As shown in the figure above, SE preprocessing moves the observed pixels (blue) toward a more representative "revised" state (red). This tightening of the data cloud allows the EF algorithm to find endmembers (red stars) that are much closer to the true physical signatures than those found using raw data (blue stars).
Experimental Results
The proposed methods were tested on synthetic fractal datasets, the Cuprite mineral dataset, and the Urban HYDICE dataset.
- Accuracy and Robustness: In synthetic tests, SE-LLR and SE-LSR consistently outperformed state-of-the-art methods like SPP and SSPP, particularly in high-noise environments (10dB to 30dB SNR), achieving significantly lower Spectral Angle Distance (SAD) scores.
- Anomaly Preservation: Unlike competing spatial methods that often smooth away rare materials, the proposed strategies maintained a near 100% success rate in detecting hidden "abnormal" pixels, thanks to the anomaly switch.
- Real-World Efficacy: On the Cuprite dataset, the SE-SVD method demonstrated the lowest reconstruction error, proving its effectiveness in complex mineralogical mapping.
Conclusion
By integrating spatial context, the proposed SE strategies significantly improve the performance of traditional spectral-based unmixing. While postprocessing offers a computationally efficient alternative, preprocessing provides the highest accuracy. Future work will focus on automating parameter selection, such as window sizes and anomaly thresholds, to make these powerful tools more accessible for automated remote sensing pipelines.
Summary updated