Hyperspectral Anomaly Detection Fused Unified Nonconvex Tensor Ring Factors Regularization

Wenjin Qin, Hailin Wang, Hao Shu, Feng Zhang, Jianjun Wang, Xiangyong Cao, Xi-Le Zhao, Gemine Vivone

TL;DR

This paper introduces HAD-EUNTRFR, a novel hyperspectral anomaly detection method that leverages Tensor Ring (TR) decomposition and a unified nonconvex regularization strategy. By effectively capturing both global and local spatial-spectral correlations in the background, HAD-EUNTRFR significantly improves anomaly detection accuracy compared to existing state-of-the-art methods.

What Problem Does This Paper Address?

Hyperspectral images (HSIs) contain rich spectral and spatial information, making them valuable for various applications. Hyperspectral Anomaly Detection (HAD) aims to identify rare pixels (anomalies) that deviate from the surrounding background. A key challenge in HAD is effectively modeling the complex background, which exhibits both global correlations and local smoothness across spectral and spatial dimensions. Existing methods often struggle to capture these characteristics simultaneously, leading to suboptimal detection performance. Specifically, many methods either fail to fully exploit the structural information within the HSI data (e.g., by converting 3D HSI cubes to 2D matrices) or use convex approximations that limit their ability to capture intricate data structures.

Methods

The proposed HAD-EUNTRFR method tackles this problem through a multi-stage approach:

  1. Background-Anomaly Separation: The initial HSI data is decomposed into background and anomaly components using a tensor robust principal component analysis (TRPCA) framework.
  2. Tensor Ring (TR) Decomposition: The background component is then processed using TR decomposition. This technique is chosen for its ability to efficiently capture high-dimensional spatial-spectral correlations in a flexible and interpretable manner.
  3. Unified Nonconvex Regularization (UNTRFR/EUNTRFR): A novel unified nonconvex regularizer is introduced. This regularizer, inspired by tensor singular value decomposition (T-SVD) and generalized nonconvex tensor correlated total variation (GNTCTV), simultaneously encodes the low-rankness and sparsity of the 3D gradient TR factors of the background. This approach aims to preserve the intrinsic low-rank and smooth properties of the background. An enhanced version (EUNTRFR) further improves robustness by decomposing gradient TR factors into low-rank and sparse components.
  4. Anomaly Component Regularization: A generalized nonconvex regularization term is designed to exploit the group sparsity of the anomaly component, further enhancing its detection.
  5. Optimization: The resulting doubly nonconvex model is solved using an efficient optimization algorithm based on the Alternating Direction Method of Multipliers (ADMM) framework.

Datasets

The authors evaluated their method on a variety of HSI datasets, including:

  • Synthetic: Salinas
  • Real-world: Pavia, Hyperion, HYDICE-Urban, San Diego, Airport-4, Urban-3, Urban-4, Urban-5, Beach-3, Beach-4, Qingpu-I, Qingpu-II, Avon, Cri, Sequoia National Park.

Key Insights and Ideas

  • TR Decomposition for Richer Representation: The paper highlights the power of Tensor Ring decomposition in capturing complex spatial-spectral correlations in HSIs, offering a more flexible and interpretable representation than traditional matrix-based methods.
  • Unified Nonconvex Regularization for L+S Priors: The core innovation lies in the unified nonconvex regularizer that simultaneously enforces low-rankness (L) and sparsity/smoothness (S) on the gradient TR factors. This approach is more effective than separate convex regularizers or applying these priors to the original domain.
  • Gradient Domain Modeling: By operating on the gradient TR factors, the method effectively captures local structural information, which is crucial for distinguishing anomalies from background.
  • Robustness through Nonconvexity: The use of nonconvex penalty functions is shown to be more effective than convex approximations, allowing for better modeling of complex data distributions and improved detection performance.
  • ADMM for Efficient Optimization: The development of an ADMM-based algorithm ensures efficient and scalable solutions for the complex, nonconvex optimization problem.

Key Results

  • Superior Detection Accuracy: Experimental results on numerous benchmark datasets demonstrate that HAD-EUNTRFR consistently outperforms existing state-of-the-art (SOTA) methods in terms of detection accuracy, as measured by various AUC metrics.
  • Effective Background Suppression: Qualitative results (anomaly detection maps) show that the proposed method effectively suppresses background noise while clearly highlighting anomalous targets.
  • Robustness to Parameters: The method exhibits robustness to the choice of trade-off parameters (α and β) and TR rank, simplifying its practical application.
  • Ablation Studies: The paper includes extensive ablation studies that validate the effectiveness of the proposed regularization strategies (GNTCTV, EGNTCTV, UNTRFR, EUNTRFR) and the TR decomposition.

Limitations and Future Work

  • Parameter Tuning: While the method shows robustness, tuning parameters like TR rank and regularization coefficients can still be challenging. The authors suggest choosing TR rank parameters to keep the two sides smaller and the middle larger, and simplifying by setting r1=r3.
  • Computational Complexity: Although an efficient ADMM algorithm is developed, the computational complexity can still be significant for very large datasets.
  • Future Work: The authors plan to explore novel data-driven regularization strategies and develop a sketching framework for dimensionality reduction. They also intend to integrate these techniques with tensor-format deep neural networks for more advanced HAD solutions.

Summary updated

Abstract

In recent years, tensor decomposition-based approaches for hyperspectral anomaly detection (HAD) have gained significant attention in the field of remote sensing. However, existing methods often fail to fully leverage both the global correlations and local smoothness of the background components in hyperspectral images (HSIs), which exist in both the spectral and spatial domains. This limitation results in suboptimal detection performance. To mitigate this critical issue, we put forward a novel HAD method named HAD-EUNTRFR, which incorporates an enhanced unified nonconvex tensor ring (TR) factors regularization. In the HAD-EUNTRFR framework, the raw HSIs are first decomposed into background and anomaly components. The TR decomposition is then employed to capture the spatial-spectral correlations within the background component. Additionally, we introduce a unified and efficient nonconvex regularizer, induced by tensor singular value decomposition (TSVD), to simultaneously encode the low-rankness and sparsity of the 3-D gradient TR factors into a unique concise form. The above characterization scheme enables the interpretable gradient TR factors to inherit the low-rankness and smoothness of the original background. To further enhance anomaly detection, we design a generalized nonconvex regularization term to exploit the group sparsity of the anomaly component. To solve the resulting doubly nonconvex model, we develop a highly efficient optimization algorithm based on the alternating direction method of multipliers (ADMM) framework. Experimental results on several benchmark datasets demonstrate that our proposed method outperforms existing state-of-the-art (SOTA) approaches in terms of detection accuracy.
BibTeX
@article{qin2025hyperspectral,
  author = {Wenjin Qin and Hailin Wang and Hao Shu and Feng Zhang and Jianjun Wang and Xiangyong Cao and Xi-Le Zhao and Gemine Vivone},
  title = {{Hyperspectral Anomaly Detection Fused Unified Nonconvex Tensor Ring Factors Regularization}},
  journal = {IEEE Transactions on Geoscience and Remote Sensing},
  year = {2025},
  doi = {10.1109/tgrs.2025.3633714},
}
Figure