The Importance of Endmember Estimation in Hyperspectral Image Processing
In hyperspectral image processing, one often has to initialize the endmember matrix, and an accurate estimate of that matrix has a major effect on the performance of the final result. When the problem is convex, the choice of initial value has a relatively small effect on the final result. For a nonconvex problem, however, an initial value close to the global optimum can substantially improve the algorithm’s performance.

Convex and nonconvex optimization
The Importance of the Number of Endmembers
In many models used in hyperspectral imaging, the number of endmembers plays a key role. For example, given a hyperspectral image
\(\mathcal{Z} \in \mathbb{R}^{M\times N\times B}\), we assume that its number of endmembers is \(R\). Under the nonnegative matrix factorization model, we have:
\[ \mathbf{Z} \in \mathbb{R}^{B\times MN} = (\mathbf{A} \in \mathbb{R}^{B\times R})(\mathbf{S} \in \mathbb{R}^{R\times MN}) \]Here, \(\mathbb{R}^{B\times MN}\) denotes the endmember matrix.
Under the tensor CP decomposition model, the expression becomes:
\[ \mathcal{Z} = \sum_{r=1}^{R} \omega_{r}(\mathbf{a}_{r} \circ \mathbf{b}_{r} \circ \mathbf{c}_{r}) \]Here, \(\omega_{r}\) denotes the weight of the \(r\)-th component. Matricizing \(a\) and \(b\) yields a rank-1 approximation of the spatial information matrix, while \(c\) is the matrix that contains the spectral information.
For the Tucker decomposition model, the expression is:
\[ \mathcal{Z} \in \mathbb{R}^{M\times N\times B} = (\mathcal{G} \in \mathbb{R}^{m\times n\times R}) \times_{1} \mathbf{A} \times_{2} \mathbf{B} \times_{3} (\mathbf{C} \in \mathbb{R}^{B\times R}) \]Here, \(\mathbf{C} \in \mathbb{R}^{B\times R}\) denotes the spectral factor matrix, which contains the compositional information of the spectra. Initializing it as the endmember matrix often improves both the algorithm’s performance and its convergence speed.
Finally, in the BTD (Block Term Decomposition), we obtain:
\[ \mathcal{Z} = \sum_{r=1}^{R} \mathcal{G}_{r} \times_{1} \mathbf{A}_{r} \times_{2} \mathbf{B}_{r} \times_{3} \mathbf{C}_{r} \approx \sum_{r=1}^{R} \mathbf{A}_{r} \cdot \mathbf{B}_{r}^{T} \circ \mathbf{c}_{r} = \sum_{r=1}^{R} \mathbf{E}_{r} \circ \mathbf{c}_{r} \]In BTD, a low-rank assumption is used to simplify the standard decomposition, which finally yields the product of the spatial factor matrix and the spectral factor matrix. The matrix \(\mathbf{C}\) denotes the spectral signature matrix of the hyperspectral image. Initializing it as the endmember matrix likewise improves the algorithm’s performance and convergence speed.
In these models, the choice of the number of endmembers \(R\)
is critical to the algorithm’s performance and to the accuracy of the result. Estimating the number of endmembers correctly not only improves processing efficiency, but also ensures that the final result is accurate and reliable.
