Entropic Descent Archetypal Analysis for Blind Hyperspectral Unmixing

Abstract: In this paper, we introduce a new algorithm based on archetypal analysis for blind hyperspectral unmixing, assuming linear mixing of endmembers. Archetypal analysis is a natural formulation for this task. This method does not require the presence of pure pixels (i.e., pixels containing a single material) but instead represents endmembers as convex combinations of a few pixels present in the original hyperspectral image. Our approach leverages an entropic gradient descent strategy, which (i) provides better solutions for hyperspectral unmixing than traditional archetypal analysis algorithms, and (ii) leads to efficient GPU implementations. Since running a single instance of our algorithm is fast, we also propose an ensembling mechanism along with an appropriate model selection procedure that make our method robust to hyper-parameter choices while keeping the computational complexity reasonable. By using six standard real datasets, we show that our approach outperforms state-of-the-art matrix factorization and recent deep learning methods. We also provide an open-source PyTorch implementation: https://github.com/inria-thoth/EDAA.

• The paper introduces an entropic descent optimization-based archetypal analysis that bypasses the pure pixel assumption in hyperspectral unmixing.
• It employs an alternating minimization and ensembling strategy to achieve the lowest RMSE and competitive spectral angle distances compared to existing methods.
• The method demonstrates significant GPU speedup while ensuring physically interpretable abundance estimates and hyperparameter robustness.

Entropic Descent Archetypal Analysis for Blind Hyperspectral Unmixing

Problem Formulation and Motivation

Blind hyperspectral unmixing (HU) operates under the premise of linear mixing in remote sensing imagery, where each pixel is an unknown superposition of several endmembers—the pure spectral signatures of materials—weighted by their respective abundances. A principal challenge in unsupervised HU is that pure pixels (pixels containing only a single material) are often absent or unrepresentative due to sensor noise, atmospheric and environmental effects, or scene complexity. Classical methods relying on pure pixel extraction (e.g., VCA, PPI, N-FINDR) are inadequate in these scenarios.

Archetypal analysis (AA) provides a robust alternative: instead of requiring pure pixel assumptions, AA models endmembers as convex combinations of scene pixels. This offers enhanced interpretability and resilience to spectral variability and noise, as each estimated endmember directly relates to constituent pixels in the image. However, the main computational obstacle in AA arises from the constrained optimization over simplex domains, often addressed by slow alternating optimization or active-set methods with poor scalability.

Methodological Contributions

The proposed algorithm adopts entropic gradient descent (mirror descent with negative entropy) for archetypal analysis, targeting the blind HU setting. The key innovations are:

• Entropic Descent Optimization: By leveraging negative entropy as the Bregman divergence, optimization steps are mapped to softmax updates, guaranteeing feasibility within simplicial constraints and enabling fast, efficient GPU-compatible implementations.
• Alternating Minimization Scheme: The AA objective is non-convex but convex in each variable holding the other fixed. Alternating entropic descent updates on abundance and pixel contribution matrices yield superior empirical solutions compared to projected gradient or active-set approaches.
• Model Selection via Ensembling: Multiple runs with random initialization and variation in step sizes (drawn from a prescribed set) are aggregated. Model selection occurs in two stages: first, subset solutions with robust $\ell_1$ reconstruction error are selected; second, from these, the candidate with lowest maximal endmember spectral coherence (pairwise inner product) is prioritized, ensuring diversity and physical interpretability.
• Hyperparameter Robustness: The model selection procedure renders the approach nearly parameter-free—the only sensitive parameter is the number of endmembers, assumed known a priori.

Experimental Benchmarks

Experiments span six canonical hyperspectral datasets (Samson, Jasper Ridge, Urban4/6, APEX, Washington DC Mall), covering varied spatial sizes, spectral bands, and endmember complexities. Comparative baselines include:

• Geometrical unmixing (FCLSU + VCA)
• Deep learning (Endnet, MiSiCNet)
• Matrix factorization (NMF-QMV, Plain AA)

The metrics of evaluation are global and per-endmember abundance RMSE and spectral angle distance (SAD).

Numerical Results:

• EDAA achieves the lowest RMSE abundances across all datasets, with the performance gap increasing as mixing complexity grows.
• SAD accuracy is highly competitive, especially in challenging scenes (Urban6, APEX, WDC), outperforming both deep learning and classical matrix factorization methods.
• Plain AA is competitive in small-scale or less mixed datasets but fails on datasets with high endmember spectral overlap or absent pure pixels, underscoring the benefits of ensembling and model selection.
• NMF-QMV's lack of non-negativity preservation for endmembers results in physically implausible spectra and uniform, non-sparse abundance maps.
• Deep learning methods (Endnet, MiSiCNet) are slower and, without spatial priors or spectral regularization, can over-sparsify abundance maps or miss key endmembers.

Computational Efficiency:

• EDAA, including 50 model selection runs, is substantially faster than deep learning approaches on GPU and even traditional AA on CPU, mainly due to softmax-based entropic updates.
• The computational cost remains tractable even as dataset size or complexity increases.

Ablation Analysis:

• The method displays insensitivity to the number of alternating steps and run count in ensembling, provided a moderate number of runs ($M\geq50$) are utilized.
• Model selection via coherence and $\ell_1$ fit consistently yields robust candidates.

Theoretical and Practical Implications

The entropic descent approach aligns optimization geometry with the simplex constraints intrinsic in abundance and endmember representation. This theoretically improves convergence rates and ensures physically interpretable results. The model selection scheme, focusing on incoherence, adheres to established sparse estimation principles, promoting well-separated endmembers to enhance abundance estimation reliability.

Practically, the approach achieves highly competitive accuracy without requiring pure pixels, extensive hyperparameter tuning, or complex preprocessing (beyond $\ell_2$ normalization). The open-source GPU implementation ensures scalability for large-scale remote sensing applications, addressing the growing demand for automated, robust HU in domains such as environmental monitoring, precision agriculture, and urban mapping.

Speculation on Future Directions

Potential advancements include:

• Modeling Nonlinear Mixing: Extension beyond linear models to accommodate intimate mixtures or nonlinear interactions in complex scenes.
• Spectral Variability Incorporation: Integration of explicit spectral variability models, possibly through bundle approaches or generative models.
• Automated Endmember Number Selection: Development of unsupervised schemes for endmember count estimation, leveraging information criteria or clustering-based heuristics.
• Adaptation to Multimodal Data: Fusion of hyperspectral information with other modalities (LiDAR, SAR) in AA frameworks for richer scene understanding.
• Hybrid Optimization Strategies: Coupling entropic descent with stochastic or block-coordinate mechanisms for even greater scalability and robustness.
• Integration with Online or Streaming Data: Adaptation to incremental or real-time HU as sensor deployments expand.

Conclusion

Entropic Descent Archetypal Analysis (EDAA) introduces a principled, scalable, and interpretable solution for blind hyperspectral unmixing under linear mixing constraints, circumventing pure pixel limitations and sidestepping the pitfalls of prior AA or NMF optimization. The combined computational and accuracy improvements, along with hyperparameter robustness and reproducibility, mark EDAA as a practical tool for remote sensing research and operational deployment (2209.11002).