Abstract
Spectral unmixing aims at recovering the spectral signatures of materials, called endmembers, mixed in a hyperspectral image (HSI) or multispectral image, along with their abundances. A typical assumption is that the image contains one pure pixel per endmember, in which case spectral unmixing reduces to identifying these pixels. Many fully automated methods have been proposed in recent years, but little work has been done to allow users to select areas where pure pixels are present manually or using a segmentation algorithm. Additionally, in a nonblind approach, several spectral libraries may be available rather than a single one, with a fixed number (or an upper or lower bound) of endmembers to chose from each. In this letter, we propose a multiple-dictionary constrained low-rank matrix approximation model that addresses these two problems. We propose an algorithm to compute this model, dubbed multiple matching pursuit alternating least squares, and its performance is discussed on both synthetic and real HSIs.
Introduction
Hyperspectral imaging (HSI) captures hundreds of narrow spectral bands, providing a detailed "fingerprint" for materials. However, due to sensor resolution limits, individual pixels often represent a mixture of several materials—such as grass, soil, and asphalt. The process of decomposing these mixed pixels into their pure constituent materials (endmembers) and their relative proportions (abundances) is known as spectral unmixing.
Traditional unmixing methods generally fall into two categories: "blind" methods that attempt to find endmembers automatically from the data, and "non-blind" methods that rely on pre-existing spectral libraries. This paper identifies a critical gap: current methods lack flexibility. They often cannot handle multiple distinct libraries simultaneously or allow users to incorporate spatial intuition (e.g., "there is a pure pixel of concrete somewhere in this specific parking lot").
The M2PALS Algorithm
To address these limitations, the authors introduce M2PALS (Multiple Matching Pursuit Alternating Least Squares). This algorithm frames endmember selection as a constrained assignment problem, allowing for a "user-in-the-loop" workflow.
The algorithm operates through an iterative three-step process:
- Least Squares Estimation: It first estimates what the endmembers and abundances should look like without constraints.
- The Assignment Problem: It "projects" these estimates onto the available dictionaries. Using the Hungarian algorithm, it finds the best match in the libraries that satisfies user-defined constraints (such as picking a specific number of atoms from each dictionary).
- Update: It updates the abundances based on these library-matched atoms and repeats until the solution converges.
This framework is uniquely flexible because a "dictionary" can be an external lab library or a set of pixels from a user-selected region of the image.
Experimental Evaluation
The researchers tested M2PALS on both synthetic and real-world datasets:
- Synthetic Urban HSI: A controlled environment where six spectra were mixed with varying levels of Gaussian noise. This allowed for a rigorous comparison of "miss-selection rates" against state-of-the-art methods.
- Urban HSI: A real-world dataset (307x307 pixels, 162 bands) representing an urban landscape with materials like asphalt, grass, and trees.
Key Insights and Results
Spatial Regularization
One of the most powerful features of M2PALS is treating image segments as dictionaries. By segmenting an image—either manually or through algorithms like H2NMF—users can create "local" dictionaries. This acts as a form of spatial regularization, forcing the algorithm to find endmembers within specific geographic bounds.
Robustness to Noise
Because M2PALS leverages a priori information (from libraries or spatial constraints), it is significantly more robust to noise than purely geometric "blind" methods. In synthetic tests with high noise (10dB SNR), M2PALS achieved a miss-selection rate of approximately 59%, whereas standard methods like SPA or NFINDR struggled with rates between 79% and 86%.
Computational Efficiency
Despite the added complexity of the assignment problem, M2PALS remains computationally efficient. Its complexity is linear relative to the number of pixels ($n$), spectral bands ($m$), and endmembers ($r$). In practical tests, the algorithm ran in roughly 0.1 seconds on a standard laptop, making it comparable in speed to much simpler geometric methods.
Conclusion
M2PALS bridges the gap between automated and manual spectral unmixing. By allowing for multiple dictionaries and spatial constraints, it provides researchers with a flexible tool that is robust to noise and computationally efficient. Future work will focus on proving the mathematical identifiability of the model and adapting it for more complex noise models and data fusion tasks.
Summary updated