PaperlayerView exact revision on arXiv

Source-linked AI summary

Interpretable Hyperspectral Unmixing Framework with Fixed Endmember Prior and Structured Residual Refinement

Ziyi Guan, Jianping Zhang, Qian Liu

arXiv:2609.08786v1cs.CV

TL;DR

When endmember priors are fixed, unexplained spectral mismatch must be handled by downstream unmixing, with limited guidance on its allocation between abundances and residuals. I-HyperSU separates abundance estimation from structured-residual refinement; across three datasets, it reduces joint reconstruction error versus fixed-prior UCLS while maintaining comparable abundance RMSE.

  • Problem

    With the endmember matrix fixed, downstream solvers lack explicit guidance on allocating unexplained spectral energy between abundance corrections and structured residual refinement.

    Source: §1 · p. 2

  • Method

    I-HyperSU separates a fixed endmember prior from soft-constrained abundance estimation and a low-rank-guided DIP branch for structured-residual refinement.

    Source: §1 · pp. 2–3

  • Results

    With default N-FINDR priors, joint reconstruction error falls 61.7%–69.5% versus fixed-prior UCLS across three datasets, with comparable abundance RMSE.

    Source: §1 · p. 3, §4 · p. 15

  • Takeaways & Limitations

    Under fixed and imperfect priors, soft abundance relaxation outperforms hard simplex projection, while the residual branch captures genuine out-of-model mismatch.

    Source: §1 · p. 3, §4 · p. 15

    Evaluation focuses on three standard datasets and a frozen-A configuration; unified comparisons with variability-aware and end-to-end methods remain future work.

    Source: §4 · p. 15

Abstract

from arXiv · show

Hyperspectral unmixing decomposes mixed pixels into material endmembers and their abundances from contiguous spectral observations. In modular sensing pipelines, endmembers are often first identified and then treated as fixed during abundance estimation. When this fixed endmember prior is inaccurate, spatially structured mismatch arising from illumination changes, sensor artifacts, or material boundaries may be incorrectly captured by the abundance variables, leading to unstable decompositions. This study presents an interpretable stage-wise hyperspectral unmixing framework (I-HyperSU) under fixed endmember priors, which is explicitly decomposed into a fixed endmember matrix $\mathbf{A}$, an abundance block $\mathbf{X}$, and a structural residual refinement block $\mathbf{S}$. The X-block estimates abundances using FISTA with nonnegativity and sparsity enhancement, and a soft penalty that approximately enforces sum-to-one constraints. The S-block jointly applies low-rank SVD structural regularization and a lightweight deep image prior (DIP) to refine structured residuals. This staged design makes the interaction between abundance and residual components transparent and interpretable. Experiments on Samson, Urban, and Jasper Ridge datasets demonstrate that, under fixed and imperfect endmember priors, soft abundance relaxation consistently outperforms hard simplex projection. Under the default N-FINDR endmember prior, the proposed framework reduces the joint reconstruction error by 61.7\%--69.5\% compared with a fixed-$\mathbf{A}$ UCLS baseline, while keeping the abundance RMSE nearly unchanged, indicating that the residual refinement branch accounts for structured model mismatch without degrading the abundance estimates. For example, on Urban, the reconstruction SAM decreases from $5.99^\circ$ for the X-only model to $1.92^\circ$ for the full model.

1 Introduction

The paper addresses fixed-endmember unmixing, where unexplained spectral mismatch must otherwise be handled by downstream estimation, and proposes I-HyperSU to separate abundance estimation from structured-residual refinement. Across three datasets, it reports lower joint reconstruction error than fixed-prior UCLS while maintaining comparable abundance RMSE.

1 Introduction

  • Fixed endmember priors leave downstream unmixing to accommodate spectral mismatch, while existing approaches offer limited guidance on allocating it between abundances and residual refinement.Some mismatch-aware methods modify or jointly estimate the endmember matrix, unlike the frozen-prior setting addressed here.Source: §1 · p. 2
  • I-HyperSU separates the fixed endmember prior A, abundance branch X, and structural-residual branch S into inspectable components.The framework targets allocation of mismatch under imperfect fixed priors.Source: §1 · pp. 2–3
  • The X-block uses FISTA with nonnegativity, sparsity regularization, and a soft sum-to-one penalty, while the S-block refines structured discrepancies.The staged procedure produces intermediate abundance-residual pairs for interpreting their interaction.Source: §1 · pp. 2–3
  • With default N-FINDR priors, joint reconstruction error falls 61.7%–69.5% versus fixed-prior UCLS across three datasets, with comparable abundance RMSE.The experiments cover Samson, Urban, and Jasper Ridge.Source: §1 · p. 3
  • On synthetic data, the residual branch tracks genuine out-of-model mismatch with r ≈0.77 overall and r > 0.99 on the component orthogonal to the endmember subspace, while remaining inactive without mismatch.This tests the branch against a known ground-truth residual.Source: §1 · p. 3

2 Method

With the endmember matrix held fixed, I-HyperSU explicitly apportions unmodeled signal between abundance estimates and a structural residual. Its stage-wise A-, X-, and S-blocks use a relaxed abundance constraint and progressively refined residual models.

2.1 Problem Formulation

  • The decomposition focuses on how unmodeled signal should be apportioned between abundance matrix X and residual term S once endmember matrix A is fixed.The model represents the residual as a structured component alongside small random perturbations.Source: §2.1 · p. 4

2.2 Proposed I-HyperSU Framework

  • A-block freezes a geometric endmember prior, X-block estimates abundances, and S-block progressively refines structured residuals across stages.The final output is abundance map X_T and structural residual S_T.Source: §2.2 · pp. 4–5, Figure 1 · p. 5, Fixed Endmember Extraction (A-Block) · pp. 5–6
  • X-block uses FISTA with nonnegativity, sparsity, and a soft sum-to-one penalty that permits controlled deviations from simplex membership.This relaxation leaves structured discrepancies to be modeled explicitly in the S-block rather than requiring abundances to account for all fixed-prior mismatch.Source: Soft Abundance Relaxation (X-Block) · pp. 6–7
  • S-block models the residual after abundance estimation, making structured discrepancies explicit instead of absorbing them into abundances.Its structural priors range from element-wise soft thresholding to low-rank truncated SVD and an SVD-guided DIP branch.Source: §2.2 · p. 4, SVD-Guided DIP Residual Modeling… · p. 7
  • The low-rank-guided DIP reparameterizes S through a neural mapping constrained by dominant SVD modes while allowing nonlinear variations within the induced subspace.The paper uses a compact per-pixel two-layer MLP; unless stated otherwise, its defaults are residual guidance rank 3 and two stages.Source: SVD-Guided DIP Residual Modeling… · pp. 7–8

3 Experiments

Experiments evaluate I-HyperSU on three benchmark datasets using fixed endmember priors and report abundance accuracy and decomposition fidelity. The setup uses a common configuration and fixed-run protocol.

3.1 Experimental Setup

  • Samson, Urban, and Jasper Ridge are evaluated with fixed endmember priors and a uniform configuration without dataset-specific tuning.The datasets contain 156, 162, and 198 bands, respectively; the X-block uses FISTA and the default S-block uses lowrank_dip with residual rank 3.Source: §3.1 · p. 8
  • Abundance accuracy is measured by xrmse, while decomposition fidelity is measured by reconstruction SAM and relative joint error.Endmember extraction is evaluated using mean SAM in degrees.Source: §3.1 · p. 8
  • Main runs use one fixed N-FINDR realization and random seed 0, with seeds 0–2 reserved for the Fig. 3b stability analysis.Quantitative comparisons are made within each dataset because λx and µ are fixed across scenes with different scales.Source: §3.1 · p. 9

3.2 Reliability of Fixed Endmember Priors

The experiments establish a reasonable but imperfect frozen endmember prior, then test the downstream pipeline under alternative priors. N-FINDR is chosen for balance across the three datasets, not because it dominates each one.

3.2 Reliability of Fixed Endmember Priors

  • The A-block evaluation seeks a reasonable but imperfect frozen prior rather than an oracle endmember extractor.The downstream solver is designed to handle mismatch in the fixed prior.Source: §3.2 · p. 9
  • N-FINDR is adopted as the default A-block because it provides the most balanced prior across Samson, Urban, and Jasper, although it is imperfect, especially on Urban.ATGP has the lowest SAM on Urban, while N-FINDR performs best on Samson and Jasper and remains competitive with ATGP on Urban.Source: §3.2 · p. 9
  • VCA, ATGP, and FIPPI are each tested through the complete pipeline in a prior-mismatch stress test, with each prior re-extracted independently.Per-extractor SAM values in that test may differ from those reported in Table 1.Source: §3.2 · p. 9

3.3 Soft Relaxation Versus Hard Simplex

Under the same imperfect fixed N-FINDR prior, soft abundance relaxation yields better abundance and joint-error outcomes than hard simplex projection. On Urban, the relaxation is associated with regions of greater prior residual.

3.3 Soft Relaxation Versus Hard Simplex

  • The comparison examines soft sum-to-one relaxation against hard simplex projection under an imperfect prior, including where the relaxation is applied.Source: §3.3 · p. 9
  • Soft abundance relaxation yields lower abundance RMSE and reduced joint error than hard simplex projection under the same fixed N-FINDR prior.Exact sum-to-one enforcement does not necessarily produce better decompositions when the prior is inaccurate.Source: §3.3 · p. 9, Table 2 · p. 10
  • On Urban, soft-over-hard error improvement grows with prior residual (Pearson r = 0.72), with about 5× greater median improvement in the most unreliable than reliable quartile.The soft sum-to-one violation map co-localizes with the fixed-prior residual, indicating selective rather than uniform relaxation.Source: Figure 2 · p. 10, §3.3 · p. 10

3.4 Ablation Study for Hyperspectral Unmixing Enhancement

The ablations show consistent gains from progressively refining the residual branch, while abundance RMSE remains essentially unchanged. Controls indicate that the staged SVD-guided routing, rather than a generic DIP module alone, distinguishes the design.

3.4 Ablation Study for Hyperspectral Unmixing Enhancement

  • The ablation varies residual sparsity, low-rank structure, and the SVD-guided DIP branch, and includes a generic Joint-DIP control.Source: §3.4 · p. 10
  • Across all three datasets, successive S-block refinements improve SAM and joint error while xrmse remains essentially unchanged.The staged enhancement path is X-only, l1-soft, lowrank_r3, then lowrank_dip.Source: §3.4 · p. 10, Table 3 · p. 11
  • The study compares joint error with non-DIP l1-soft and analytic low-rank S-blocks (Fig. 3), using (T, rs) = (2, 3) as a globally fixed default rather than a tuned optimum.The rs=5 variant is reported only as an exploratory higher-capacity option.Source: §3.4 · p. 11
  • Frozen-A Joint-DIP has worse reconstruction SAM and joint error than the staged design on all three datasets, while three DIP seeds keep joint error within ±1% of its mean.Joint-DIP attains slightly lower abundance RMSE in some cases.Source: §3.4 · p. 11

3.5 Qualitative Residual Refinement and Abundance Stability

Structured residual models progressively reduce the final residual, with lowrank_dip achieving the lowest reconstruction SAM on each dataset. Residual refinement improves spectral fit most in high-mismatch regions while abundance drift remains small.

3.5 Qualitative Residual Refinement and Abundance Stability

  • Fig. 3 tracks DIP joint error versus residual rank and seed stability; all three seed runs stay within ±1% of the per-dataset mean.Source: Figure 3 · p. 12
  • Lowrank_dip attains the lowest reconstruction SAM on every dataset as richer residual models leave weaker final residuals.Fig. 4 compares l1-soft, lowrank_r3, and lowrank_dip under the fixed N-FINDR prior.Source: §3.5 · p. 11, Figure 4 · p. 12
  • Across fixed-prior residual quartiles, S lowers median SAM, with larger gains in higher-residual regions.For one high-residual pixel per dataset, AX + S better matches Y and reduces per-pixel SAM.Source: §3.5 · p. 11, Figure 5 · p. 12
  • Abundance drift stays within 0.003 in ∆xrmse and 0.022 in ∆Xrms across all datasets.Source: §3.5 · p. 12

3.6 Comparisons With Other Methods

Under a frozen endmember prior, I-HyperSU improves reconstruction consistency over classical solvers while preserving competitive abundance RMSE. The comparison uses downstream solvers without endmember updates or extra training data.

3.6 Comparisons With Other Methods

  • Across all three datasets, I-HyperSU attains the best reconstruction SAM and joint error while preserving abundance RMSE comparable to the strongest classical baseline.Source: §3.6 · p. 13
  • Under the frozen-prior, fixed-A regime, I-HyperSU reduces joint error relative to UCLS by roughly 69.5% on Samson, 67.0% on Urban, and 61.7% on Jasper, while maintaining competitive abundance RMSE.Source: §3.6 · p. 13
  • External baselines are restricted to solvers that neither update A nor use extra training data; SUnSAL is omitted because its results nearly matched NNLS under the adopted parameters.Source: §3.6 · p. 13

3.7 Synthetic Validation with a Known Residual

A synthetic test with known residual structure evaluates whether S recovers genuine mismatch rather than merely fitting observation residuals. S correlates strongly with the true out-of-subspace residual and reduces joint error across mismatch strengths.

3.7 Synthetic Validation with a Known Residual

  • The Jasper synthetic test adds a known non-negative rank-5 residual representing stray-light/haze and striping, partly outside span(Agt), and sweeps strengths 0, 0.1, 0.2, and 0.3.The pipeline runs with A = Agt fixed.Source: §3.7 · pp. 13–14
  • S correlates with the known residual at r≈0.77 overall and r > 0.99 on its out-of-subspace component.The out-of-subspace component cannot be explained by abundance changes.Source: Figure 6 · p. 14, §3.7 · p. 14
  • Routing residual into S substantially reduces joint error at every tested mismatch strength, while the S0=0 control shows no fabricated structure and essentially unchanged abundance RMSE.Source: §3.7 · p. 14, Figure 6 · p. 14

3.8 Prior-Mismatch Stress Test

Across four endmember extractors, the residual branch keeps reconstruction error low, but abundance accuracy degrades for severely inaccurate priors. I-HyperSU is therefore reconstruction-resilient under imperfect priors, not prior-independent.

3.8 Prior-Mismatch Stress Test

  • Abundance accuracy degrades for severely wrong priors, including FIPPI, so S cannot restore abundance identifiability once endmembers are badly inaccurate.Source: §3.8 · p. 14, Figure 7 · p. 15
  • The framework is reconstruction-resilient under imperfect priors but not prior-independent; accurate abundance recovery still requires reasonably reliable endmembers.Source: §3.8 · p. 14
  • With S, joint error and reconstruction SAM remain low across N-FINDR, VCA, ATGP, and FIPPI priors.The stress test repeats the full pipeline across the three datasets.Source: §3.8 · p. 14, Figure 7 · p. 15

4 Conclusion

I-HyperSU separates fixed endmember priors, abundance estimation, and structured residual modeling. Experiments report improved reconstruction with comparable abundance RMSE, while the study remains scoped to three standard datasets and a frozen-A configuration.

4 Conclusion

  • The framework reduces joint reconstruction error by 61.7%–69.5% relative to UCLS under the default N-FINDR prior, with negligible impact on abundance RMSE.This result is reported across Samson, Urban, and Jasper Ridge.Source: §4 · p. 15
  • Soft abundance relaxation consistently outperforms hard simplex projection under fixed and imperfect endmember priors.Source: §4 · p. 15
  • Synthetic and prior-mismatch experiments show S captures genuine out-of-model mismatch (r ≈0.77), while accurate abundance recovery still requires a reasonable prior.Source: §4 · p. 15
  • The study covers three standard datasets under a frozen-A configuration; unified comparisons with variability-aware and end-to-end deep methods remain future work.Source: §4 · p. 15
Loading 2609.08786v1…