Abstract:Hypersepctral unmixing (HU) has been one of the most challenging tasks in hyperspectral image research. Recently, nonnegative matrix factorization (NMF) has shown its superiority in hyperspectral unmixing due to its flexible modeling and little prior requirement. But most NMF algorithms tend to use least square function as the objective, which is sensitive to outliers and different kinds of noise. In this article, we propose a modified Huber (mHuber) NMF model to achieve robustness to outliers and different kinds of noise. Under this robust model, we accelerate the half-quadratic optimization algorithm by replacing multiplicative updating rule with a projected nonlinear conjugated gradient rule, which achieves much faster convergence rate. Moreover, a new tuning parameter, rather than a fixed one, is given to adapt to mHuber loss function. Finally, we perform algorithm analysis and experiments in the synthetic and real-world datasets, which confirms the effectiveness and superiority of the proposed method when compared with several state-of-the-art NMF methods in HU.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that in hyperspectral image unmixing, the existing non - negative matrix factorization (NMF) methods are sensitive to outliers and different types of noise. Specifically, most NMF algorithms tend to use the least - squares function as the objective function, which is very sensitive to outliers and various noises in practical scenarios, resulting in insufficient robustness of the model. To solve this problem, the author proposes an NMF model based on the modified Huber loss function (mHuber NMF) to improve the robustness to outliers and noise, and accelerates the optimization process through the projected non - linear conjugate gradient algorithm, thereby achieving a faster convergence speed. ### Main contributions: 1. **Introduction of the modified Huber loss function**: It not only improves the robustness to outliers and different types of noise, but also ensures the stability and better performance of the model. 2. **Proposing the projected non - linear conjugate gradient algorithm**: Update the matrix \( S \) in the semi - quadratic optimization framework, making up for the shortcoming of the slow convergence speed of the multiplicative update rule. 3. **Proposing a new parameter - tuning method**: Introduce a new adjustment parameter to adapt to the modified Huber loss function instead of using the fixed median method. ### Model and algorithm: - **Linear Mixture Model (LMM)**: Assume that each pixel is composed of a linear combination of several pure end - members. The model is represented as \( \mathbf{x}=\mathbf{A}\mathbf{s}+\boldsymbol{\epsilon} \), where \( \mathbf{A} \) is the end - member matrix, \( \mathbf{s} \) is the abundance vector, and \( \boldsymbol{\epsilon} \) is the noise. - **Huber loss function**: The traditional Huber loss function behaves as a squared loss when the error is small and as an absolute value loss when the error is large. The modified Huber loss function further improves the stability. \[ H(E_{ij})=\begin{cases} \frac{1}{2}E_{ij}^2 & \text{if } \|E_{ij}\|^2\leq c\\ c\|E_{ij}\|-\frac{1}{2}c^2 & \text{otherwise} \end{cases} \] The modified Huber loss function is: \[ H_M(E_{ij})=\begin{cases} c^2\left[1 - \cos\left(\frac{E_{ij}}{c}\right)\right] & \text{if } \left|\frac{E_{ij}}{c}\right|\leq\frac{\pi}{2}\\ c|E_{ij}|+c^2\left(1-\frac{\pi}{2}\right) & \text{otherwise} \end{cases} \] - **Optimization objective function**: Introduce the nuclear norm and graph regularization terms to ensure low - rank and sparsity while maintaining the intrinsic manifold structure of the abundance. \[ f(\mathbf{A},\mathbf{S})=\frac{1}{2}\sum_{i = 1}^{M}\sum_{j = 1}^{N}H_M((\mathbf{X}-\mathbf{A}\mathbf{S})_{ij})+\lambda_1(\|\mathbf{A}\|_F^2+\|\mathbf{S}\|_F^2)+\lambda_2\text{Tr}(\mathbf{S}\mathbf{L}\mathbf{S}^T) \] - **Update rules**: Adopt the semi - quadratic optimization framework, and update the matrices \( \mathbf{A} \) and \( \mathbf{S} \) respectively by introducing auxiliary variables and the conjugate gradient method.

A Modified Huber Nonnegative Matrix Factorization Algorithm for Hyperspectral Unmixing

Spectral–Spatial Robust Nonnegative Matrix Factorization for Hyperspectral Unmixing

Two-Dimensional Robust Nonnegative Matrix Factorization for Hyperspectral Unmixing.

Hyperspectral Unmixing Using Robust Deep Nonnegative Matrix Factorization

Hypergraph-Regularized Sparse NMF for Hyperspectral Unmixing.

Nonnegative matrix factorization with piecewise smoothness constraint for hyperspectral unmixing

PIXEL-WISE LINEAR/NONLINEAR NONNEGATIVE MATRIX FACTORIZATION FOR UNMIXING OF HYPERSPECTRAL DATA

Spectral-Spatial Constrained Nonnegative Matrix Factorization for Spectral Mixture Analysis of Hyperspectral Images

Nonnegative Matrix Factorization with Data-Guided Constraints for Hyperspectral Unmixing.

Minimum Variance Block-Based Nonnegative Matrix Factorization Algorithm for Hyperspectral Unmixing.

Nonlinear Hyperspectral Unmixing Based on Constrained Multiple Kernel NMF

Kernel Based Sparse NMF Algorithm for Hyperspectral Unmixing

Constrained Nonnegative Matrix Factorization for Robust Hyperspectral Unmixing.

Adaptive $textit{L}_{bf 1/2}$ Sparsity-Constrained NMF With Half-Thresholding Algorithm for Hyperspectral Unmixing

Parallel Adaptive Sparsity-Constrained NMF Algorithm for Hyperspectral Unmixing.

Non-Negative Matrix Factorization with Averaged Kurtosis and Manifold Constraints for Blind Hyperspectral Unmixing

Nonconvex Nonseparable Sparse Nonnegative Matrix Factorization for Hyperspectral Unmixing

Hyperspectral Image Unmixing Based on Sparse and Minimum Volume Constrained Nonnegative Matrix Factorization.

High-Efficiency Hyperspectral Unmixing Based on Band Selection

Nonlinear Hyperspectral Unmixing With Robust Nonnegative Matrix Factorization

Multiple Clustering Guided Nonnegative Matrix Factorization for Hyperspectral Unmixing