PaperlayerView exact revision on arXiv

Source-linked AI summary

Structured Sparse Method for Hyperspectral Unmixing

Feiyun Zhu, Ying Wang, Shiming Xiang, Bin Fan, Chunhong Pan

arXiv:1403.4682v1cs.CVcs.AI

TL;DR

Hyperspectral unmixing needs to exploit spatial relationships and sparse mixing because existing methods often fail to use spatial information. SS-NMF combines graph-Laplacian structure modeling with lasso-regularized NMF, and experiments report better performance than state-of-the-art methods across noise levels.

  • Problem

    Existing HU methods often fail to exploit spatial information, despite hyperspectral pixels containing manifold structures and sparse mixtures.

  • Method

    SS-NMF combines a graph-based manifold constraint with lasso-regularized nonnegative matrix factorization to learn regionally smooth, sparse abundances.

  • Results

    SS-NMF outperforms or matches competing methods across noise levels, including 59.12% lower SAD and 8.22% lower RMSE than the best non-proposed baselines on Urban data.

  • Takeaways & Limitations

    The structured sparse constraint transfers manifold structure into abundance space while learning expressive endmembers, with robustness to different noise levels.

Abstract

from arXiv · show

Hyperspectral Unmixing (HU) has received increasing attention in the past decades due to its ability of unveiling information latent in hyperspectral data. Unfortunately, most existing methods fail to take advantage of the spatial information in data. To overcome this limitation, we propose a Structured Sparse regularized Nonnegative Matrix Factorization (SS-NMF) method from the following two aspects. First, we incorporate a graph Laplacian to encode the manifold structures embedded in the hyperspectral data space. In this way, the highly similar neighboring pixels can be grouped together. Second, the lasso penalty is employed in SS-NMF for the fact that pixels in the same manifold structure are sparsely mixed by a common set of relevant bases. These two factors act as a new structured sparse constraint. With this constraint, our method can learn a compact space, where highly similar pixels are grouped to share correlated sparse representations. Experiments on real hyperspectral data sets with different noise levels demonstrate that our method outperforms the state-of-the-art methods significantly.

I. INTRODUCTION

Hyperspectral unmixing decomposes mixed pixel spectra into constituent endmembers and abundances, but conventional methods often neglect spatial structure and sparse mixing. SS-NMF addresses these gaps with a graph-based structured sparse constraint and reports stronger performance than state-of-the-art methods.

  • Motivation: Hyperspectral imaging records hundreds of contiguous narrow spectral bands, but limited spatial resolution causes neighboring substances to merge into mixed pixels.HU decomposes each pixel spectrum into constituent spectra, such as tree, water, and soil, and their corresponding proportions.
  • Prior methods: NMF is physically suited to HU because nonnegative factors represent endmembers and abundances while allowing additive combinations.Its solution space is non-convex and parts-based representations may be insufficiently expressive for HU.
  • Priors: Pixels within the same image structure can have similar spectra and be sparsely mixed by a common set of relevant endmembers.The introduction uses road regions and color-coded abundances to illustrate these spatial and sparse priors.
  • Proposed method: SS-NMF encodes latent structures with a weighted graph and combines graph regularization with a lasso penalty on abundances.The resulting constraint encourages related neighboring pixels to share correlated sparse representations.
  • Results: SS-NMF comparisons on hyperspectral data with different noise levels show stronger quantitative and qualitative performance than state-of-the-art methods.The paper describes the method as effective for HU and presents extensive experiments and comparisons.

B. Nonnegative Matrix Factorization

The paper extends NMF for HU by preserving manifold structure among similar pixels while enforcing sparse abundances. Its graph construction uses local spatial and spectral similarity, and its structured penalty transfers those relationships into abundance space.

  • Nonnegative Matrix Factorization: NMF decomposes a nonnegative data matrix into nonnegative factor matrices, providing a parts-based representation suited to hyperspectral data.The joint objective is non-convex even though it is convex in each factor separately.
  • Limitations of NMF: NMF does not model the sparse abundance prior or manifold structures that are essential for hyperspectral unmixing.SS-NMF introduces a structured sparse constraint to address both limitations.
  • Graph structure: The weighted graph groups highly related neighboring pixels and transfers their local manifold relationships into the learned abundance space.Local overlapping structures are combined to capture similarity information between pixel pairs.
  • Graph construction: Neighbors are selected within a 7 × 7 window and among the top 30% by spectral-angle similarity.This selection reduces graph connections across dissimilar targets at image edges.
  • Structured sparse constraint: The lasso penalty makes abundances sparse, while graph regularization encourages sparse abundance vectors from the same structure to remain similar.The graph weight controls how strongly two pixels share similar abundances.

B. Optimization for SS-NMF

SS-NMF is optimized by alternating updates of the endmember and abundance matrices under nonnegativity constraints. The implementation initializes factors and graph structure, rescales factors, and iterates until the objective change falls below a threshold.

  • Optimization: The joint SS-NMF objective is non-convex in the endmember and abundance matrices, so the algorithm seeks local minima iteratively.The method alternates updates for the two factor matrices.
  • Optimization: The optimization incorporates Lagrange multipliers and Karush-Kuhn-Tucker conditions for the nonnegative endmember and abundance constraints.These conditions yield equations used to derive the multiplicative update rules.
  • Update rules: Using the Laplacian relation L = D − W, the derived equations produce update rules for the two factor matrices.The updates are applied alternately to M and A.
  • Normalization: Scaling factor matrices removes the diagonal-scale ambiguity shared by NMF solutions.SS-NMF applies this normalization strategy after each iteration for computer realization.
  • Algorithm: The algorithm initializes M from dissimilar pixels, initializes A with positive values, constructs W, D, and L, then iterates until the objective change meets threshold τ.After each iteration, M and A are scaled before the final unmixing result is returned.

C. Proof of Convergence

The convergence analysis uses an auxiliary upper-bound function to show that alternating updates make the objective non-increasing and converge to local minima after finite iterations.

  • C. Proof of Convergence: The optimization objective is non-increasing under the proposed updating rules and converges to local minima after finite iterations.The proof focuses on the abundance update because the endmember update matches NMF's established convergence proof.
  • C. Proof of Convergence: An auxiliary function G(A, A′) is defined as an upper bound for the objective function F(A).The construction satisfies the auxiliary-function relationship used in the convergence argument.
  • C. Proof of Convergence: Minimizing the auxiliary function produces a suitable A(t+1) that makes F(A) non-increasing.The update rule is obtained by minimizing G(A, A′).
  • C. Proof of Convergence: The abundance objective is represented as a quadratic function, and the auxiliary function is constructed from its Taylor expansion.This construction links the update rule to minimization of the upper-bound function.
  • C. Proof of Convergence: The proof establishes the upper-bound inequality by comparing quadratic terms and showing two component terms are nonnegative.The argument separately verifies nonnegativity for the terms identified in the inequality decomposition.

D. Comparison with Gradient Descent Method

The paper relates SS-NMF's alternating updates to gradient descent, identifying them as special gradient-descent rules under learning-rate conditions that preserve meaningful local minima.

  • D. Comparison with Gradient Descent Method: Gradient-descent learning rates must be sufficiently small to obtain local minima and preserve the nonnegative factors during each iteration.The learning rates are denoted [u_lk] and [v_kn].
  • D. Comparison with Gradient Descent Method: SS-NMF updating rules (19) and (20) are special updating rules obtainable through the Gradient Descent Method.The relationship is established for the optimization problem defined by the structured sparse objective.

E. Computational Complexity Analysis

SS-NMF requires extra per-iteration work to construct pixel-structure relationships, but its total computational complexity is reported as very close to NMF because it needs fewer iterations.

  • E. Computational Complexity Analysis: SS-NMF adds O(LNp) cost for constructing structure relationships between pixel pairs.The total cost also depends on the number of iterations t and the factor dimensions L, N, and K.
  • E. Computational Complexity Analysis: When both methods use t iterations, SS-NMF has slightly higher computational complexity than NMF.This comparison concerns the iterative algorithms and their per-iteration costs.
  • E. Computational Complexity Analysis: The total computational complexity of SS-NMF is very close to that of NMF because SS-NMF requires fewer iterations for convergence.The paper states that the iteration comparison is shown in a later subsection.
  • E. Computational Complexity Analysis: The complexity analysis counts additions, multiplications, and divisions in floating-point precision and summarizes them for each iteration in Table II.Table I lists the parameters used in the analysis.
  • E. Computational Complexity Analysis: Because K ≪ min{N, L}, multiplying (MA)A^T is reported as more costly in the stated matrix-multiplication comparison.The comparison contrasts approximately 2LNK operations with (N + L)K^2 operations.

IV. EXPERIMENTS

The experiments evaluate SS-NMF for hyperspectral unmixing and are designed to demonstrate that the algorithm is well suited to the problem.

  • IV. EXPERIMENTS: Several experiments are conducted to evaluate the effectiveness of SS-NMF for the hyperspectral unmixing problem.The experiments are intended to show that the proposed algorithm is well suited for HU.

A. Evaluation Metrics

The evaluation uses SAD for estimated endmembers and RMSE for estimated abundance maps, with lower values indicating better performance.

  • Evaluation Metrics: SAD measures the angle distance between an estimated endmember and its corresponding ground truth.Smaller SAD values indicate better endmember estimation.
  • Evaluation Metrics: RMSE measures the error between an estimated abundance map and its corresponding ground truth.Smaller RMSE values indicate better abundance-map estimation.
  • Evaluation Metrics: Abundance maps are presented in both pseudo-color and grayscale formats.

B. Compared Algorithms

The study compares SS-NMF with geometrical and statistical HU methods, while tuning sparsity and graph-regularization parameters for relevant algorithms.

  • Compared Methods: The comparison includes SS-NMF, VCA, ℓ1/2-NMF, G-NMF, W-NMF, and EDC-NMF among eight evaluated algorithms.VCA is geometrical; the other seven methods are statistical.
  • Compared Methods: VCA requires pure pixels for each endmember and estimates abundances separately through a constrained least-squares problem.The other listed methods estimate endmembers and abundances simultaneously.
  • Parameter Settings: SS-NMF uses α to control the lasso penalty and λ to control the structured constraint.
  • Parameter Settings: The sparsity-related parameter α is estimated from hyperspectral-data sparsity and refined by searching 50 equally spaced values.The same parameter-estimation approach is used for ℓ1-NMF and ℓ1/2-NMF.
  • Parameter Settings: The reference λ0 is estimated from similarities within 100 randomly selected 5 × 5 patches, averaging 2400 central-neighbor values before a 50-value search.The graph-constraint parameters in G-NMF and W-NMF are set similarly.

D. Real Data Sets

Experiments use Urban and Jasper Ridge hyperspectral data, evaluate robustness across seven Gaussian-noise levels, and report consistently strong SS-NMF performance.

  • Real Data Sets: Urban contains 307 × 307 pixels and 162 retained spectral bands, while Jasper Ridge uses a 100 × 100 subimage with 198 retained bands.Jasper Ridge is reduced to a subimage because the original data are too complex for obtaining ground truth.
  • Real Data Sets: Ground truths are constructed by determining the endmember number, manually selecting spectra, and solving a constrained convex optimization problem for abundances.
  • Noise Robustness: Each data set is tested with i.i.d. zero-mean white Gaussian noise at seven SNR levels, with every experiment repeated 50 times.Mean results and standard deviations are reported.
  • Quantitative Results: On Urban, SS-NMF is consistently best or comparable across noise levels and achieves 59.12% lower SAD and 8.22% lower RMSE than the strongest alternatives.The comparison baselines are ℓ1/2-NMF for SAD and EDC-NMF for RMSE.
  • Quantitative Results: Sparse-regularized NMF methods perform better or comparably to graph-only NMF methods for estimating endmembers.The paper attributes this observation to the relative importance of sparse constraints for endmember estimation.
  • Quantitative Results: NMF-based methods significantly outperform VCA because they do not require pure pixels and retain a parts-based representation.
  • Quantitative Results: On Jasper Ridge, SS-NMF is consistently best or comparable and achieves 30.68% lower SAD and 35.34% lower RMSE than the strongest alternatives.The comparison baselines are ℓ1/2-NMF for SAD and ℓ1-NMF for RMSE.

2) Qualitative Results and Analysis:

SS-NMF produces abundance maps that more closely match ground truth while preserving regional smoothness and sparse mixing. Its performance depends on regularized parameter settings, with strong results near the selected values but reduced robustness under extreme settings.

  • SS-NMF achieves abundance maps most similar to the ground truths, with regional smoothness and sparse mixing evident in its color patterns.These visual properties support the effectiveness of the structured sparse constraint.
  • Sparse regularized NMF methods produce noisier abundance maps than SS-NMF and G-NMF because they lack the graph-based smoothness constraint.The graph structure constraint encourages learned abundance maps to be smooth.
  • SAD and RMSE are averaged across estimated endmembers and abundance maps, respectively, to evaluate performance on Urban and Jasper Ridge data.Figure 8 varies λ and α across nine settings around selected values.
  • SS-NMF performance first improves and then decreases as λ and α vary across the tested ranges.The experiments use parameter values from 2^-4 to 2^4 times the selected settings.
  • SS-NMF outperforms the other algorithms greatly under the suitable setting λ = λ̂ and α = α̂.
  • ℓ1-NMF, ℓ1/2-NMF, and G-NMF appear more robust to parameter variation, while SS-NMF performs poorly under extreme settings.The paper attributes this partly to SS-NMF’s stronger sensitivity despite its better results under suitable settings.

G. Influences of Weighting Methods

The weighting-method experiments compare spatial-only with spatial-and-feature-distance graphs, while additional analyses examine convergence and overall conclusions. Spatial-only weighting yields overly smooth, hazy maps; SS-NMF otherwise achieves strong quantitative and qualitative performance and is relatively robust to noise.

  • Influences of Weighting Methods: Spatial-only weighting and spatial-plus-feature weighting are compared using abundance maps on Urban and Jasper Ridge data.The compared weighting strategies construct local graph structures from spatial distance alone or from both spatial and feature distances.
  • Influences of Weighting Methods: Spatial-only weighting produces abundance maps that are too smooth and hazy.The paper attributes this effect to relying only on spatial distance when constructing the graph.
  • Convergence Analysis: SS-NMF requires graph construction before iteration, whereas NMF does not require this graph-construction step.Table VII separates graph-construction time, iteration time, and total convergence time.
  • Convergence Analysis: SS-NMF costs less time than NMF during iteration because it requires fewer iterations before convergence.Figure 11 plots objective-function energy against iteration number for Urban and Jasper Ridge data.
  • Convergence Analysis: NMF converges faster overall on Urban data, but not on Jasper Ridge data; graph-construction overhead contributes to this dataset-dependent difference.The reported ratios of graph-construction time to convergence time are 0.38 for Urban and 0.30 for Jasper Ridge.
  • Conclusion: SS-NMF transfers manifold structure into abundance space, learns expressive and accurate endmembers, and achieves better quantitative and qualitative results than state-of-the-art methods.The conclusion also reports relative robustness to different noise levels.
  • Qualitative Results: Urban abundance maps are shown across seven noise levels, with algorithm outputs compared against ground truths.Rows represent noise levels, columns represent algorithms or ground truth, and colors encode fractional abundances of materials.
Loading 1403.4682v1…