Adaptive Multi-Order Graph Regularized NMF with Dual Sparsity for Hyperspectral Unmixing

Hui Chen, Liangyu Liu, Xianchao Xiu, Wanquan Liu · 2025

Read Adaptive Multi-Order Graph Regularized NMF with Dual Sparsity for Hyperspectral Unmixing by Hui Chen, Liangyu Liu, Xianchao Xiu, Wanquan Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-19258.txt.

Introduction

Hui Chen, Liangyu Liu, Xianchao Xiu, and Wanquan Liu unmix satellite rainbow images with a factorization that learns its own map of neighbors in this 2025 paper. Each pixel in a hyperspectral image blends several materials, and unmixing names the parts and their shares. Standard factorization methods lean on first or second neighbors and need hand tuning.

Their method builds links at several orders at once, near and far, and lets the data set the link strengths by itself. Two kinds of sparsity keep the answer clean: one prunes the material shares, the other isolates noise into its own matrix. An alternating solver with closed-form steps does the work. On fake and real scenes the method beat its rivals at recovering the true materials. The authors pitch it for mapping land cover, minerals, and crops from orbit.

Catalog record

CREATORHui Chen, Liangyu Liu, Xianchao Xiu, Wanquan Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-19258

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Adaptive Multi-Order Graph Regularized NMF with Dual Sparsity for Hyperspectral Unmixing Hui Chen, Liangyu Liu, Xianchao Xiu, Wanquan Liu Abstract Hyperspectral unmixing (HU) is a critical yet challenging task in remote sensing. However, existing nonnegative matrix factorization (NMF) methods with graph learning mostly focus on first-order or second-order nearest neighbor relationships and usually require manual parameter tuning, which fails to characterize intrinsic data structures. To address the above issues, we propose a novel adaptive multi-order graph regularized NMF method (MOGNMF) with three key features. First, multi-order graph regularization is introduced into the NMF framework to exploit global and local information comprehensively. Second, these parameters associated with the multi-order graph are learned adaptively through a data-driven approach. Third, dual sparsity is embedded to obtain better robustness, i.e., ℓ 1 / 2 \ell_{1/2} -norm on the abundance matrix and ℓ 2 , 1 \ell_{2,1} -norm on the noise matrix. To solve the proposed model, we develop an alternating minimization algorithm whose subproblems have explicit solutions, thus ensuring effectiveness. Experiments on simulated and real hyperspectral data indicate that the proposed method delivers better unmixing results. I Introduction In the field of remote sensing, hyperspectral images (HSIs) have attracted extensive attention due to their high spatial and spectral resolution. However, HSIs are usually high-dimensional and noisy, which has led to a series of techniques to extract useful information, such as denoising [ 1 , 2 , 3 ] , classification [ 4 , 5 , 6 ] , anomaly detection [ 7 , 8 , 9 ] , and hyperspectral unmixing (HU) [ 10 , 11 , 12 ] . Among them, HU plays a critical role in analyzing HSIs by separating spectral endmembers and estimating their abundance fractions, with wide applications in medical diagnostics [ 13 ] , environmental monitoring [ 14 ] , and precision agriculture [ 15 ] . Based on the spectral mixing mechanism, a large number of methods have been proposed to deal with the HU problem. The linear mixing model (LMM) and the nonlinear mixing model (NLMM) are two representatives. LMM assumes that each pixel is a linear combination of endmember spectra, while NLMM further accounts for multiple scattering and microscopic mixing effects [ 16 ] . Typical NLMM methods include the generalized bilinear model [ 17 ] , the p-order polynomial model [ 18 ] , and the multilinear mixing model [ 19 ] . Due to the inherently nonconvex nature of unmixing problems, nonlinear methods often involve high degrees of freedom, which may lead to the generation of physically meaningless endmembers [ 20 ] . Nowadays, LMM has been widely regarded as a reasonable first-order approximation of the physical measurement process due to its strong physical interpretability [ 21 ] . Generally speaking, LMM can be divided into geometry-based [ 22 , 23 ] , statistics-based [ 24 , 25 ] , regression-based [ 26 , 27 ] , and deep learning-based [ 28 , 29 ] . Geometry-based methods like vertex component analysis (VCA) [ 30 ] extract endmembers by locating extreme pixels according to the data’s simplex structure, but they usually require pure pixels. Regression-based methods such as fully constrained least squares (FCLS) [ 31 ] estimate abundances under nonnegativity and sum-to-one constraints. The combination of VCA and FCLS, known as the VCA-FCLS framework, is widely used as initialization by separately handling endmember extraction and abundance estimation. In contrast, nonnegative matrix factorization (NMF) [ 32 ] jointly estimates endmembers and abundances under nonnegativity constraints, aligning with the physical properties of hyperspectral data. Unlike VCA-FCLS, it does not require pure pixels and supports unsupervised learning. Its flexibility and extensibility have made NMF a widely used tool, as highlighted in recent surveys [ 33 , 34 ] . Fig. 1: Flowchart of our proposed method. First, input the HSI. Second, extract spatial and spectral features. Third, construct the multi-order graph, where blue solid lines, yellow dashed lines, and gray dashed lines represent first-order neighbor relationships, second-order neighbor relationships, and third-order neighbor relationships, respectively. Fourth, add dual sparse regularization terms. Finally, output the abundance map and the endmember map. To enhance the unmixing performance of NMF, researchers have developed a large number of constrained NMF methods and achieved remarkable results. For example, Miao and Qi [ 35 ] incorporated a volume constraint based on the convex cone simplex into the standard NMF and proposed MVC-NMF. Qian et al. [ 36 ] suggested \ell_{1/2} -norm [ 37 ] constrained NMF to improve the interpretability, which is denoted as SNMF in our paper without ambiguity. Subsequently, Zhu et al. [ 38 ] replaced \ell_{1/2} -norm with \ell_{q} -norm (0<q<1) to obtain the ability of adaptive selection [ 39 ] . Furthermore, Wang et al. [ 40 ] considered the sparse structure of local pixels and combined it with the smoothness constraint, which is called spatial group sparsity regularized NMF (SGSNMF). Besides, there are low-rank constrained NMF [ 41 ] , [ 42 ] , abundance separation constrained NMF [ 43 ] , [ 44 ] , etc. We would like to point out that although constrained NMF methods significantly improve the performance of traditional NMF by incorporating prior information, it is still not robust to noise and outliers [ 45 , 46 ] . To this end, Févotte et al. [ 47 ] proposed robust NMF (RNMF) by introducing additional sparse matrices to filter out sparsely distributed noise. Recently, Zhang et al. [ 48 ] constructed group sparse constrained NMF, which is denoted as IRNMF. In addition to the above constrained NMF methods, Abdolali et al. [ 49 ] transformed the minimum-volume NMF problem in the primal space to a maximum-volume problem in the dual space for a more stable global solution, which is called MV-DUAL. In summary, the above-mentioned sparse constrained NMF methods have good unmixing performance, hence we will continue to align this strategy. Recently, graph learning was incorporated with NMF to capture the low-dimensional features in HSIs [ 50 , 51 ] . Lu et al. [ 52 ] proposed a graph-regularized NMF method by combining manifold regularization with sparse constrained NMF, where the manifold regularization term was validated to strengthen the connection between original images and abundance maps. Yang et al. [ 53 ] considered spatial geometric distance and spectral geometric distance as metrics and obtained a space-spectral manifold-regularized NMF model. Wang et al. [ 54 ] introduced the hypergraph structure to capture the similarity relationships among spatially neighboring pixels and proposed hypergraph-regularized sparse NMF. Very recently, Qu et al. [ 55 ] presented a high-order graph regularized NMF method, called HGNMF-FS. However, these graph learning-based methods have two limitations. On the one hand, most methods rely on first-order or fixed-order neighborhoods, resulting in sparse graphs that fail to capture high-order pixel associations [ 56 ] . This limitation reduces the model’s ability to represent both global and local structures, such as the similarity between neighboring pixels and the consistency across distant but similar regions [ 57 ] . On the other hand, their graph construction typically depends on predefined parameters, lacking the flexibility to adaptively learn graph structures. As shown in Fig. 1 , adaptive learning allows the model to discover the optimal graph

Read the complete text