Regularized Tensor Representative Coefficient Model for Hyperspectral Target Detection

Wenting Shang, Mohamad Jouni, Zebin Wu, Yang Xu, Mauro Dalla Mura, Zhihui WeiIEEE GRSL 2023 · OpenReview

TL;DR

This paper introduces a novel tensor-based approach, TVLrS-TRC, for hyperspectral target detection. By modeling hyperspectral data as a tensor and exploiting its inherent low-rank and sparse structures, the method achieves superior detection accuracy and background suppression compared to existing matrix-based and tensor-based techniques.

What Problem Does This Paper Address?

Hyperspectral imaging (HSI) offers rich spectral information, making it invaluable for applications like remote sensing, security, and environmental monitoring. A key challenge in HSI is target detection (TD): identifying specific materials within a scene based on their unique spectral signatures. Traditional TD methods often rely on matrix representations of HSI data, which can lead to the loss of crucial spatial information and fail to capture the intrinsic multimodal structure of HSI cubes. While some tensor-based methods have emerged, they may not fully exploit the physical characteristics of the representation coefficients or effectively model both background and target properties. This paper aims to bridge this gap by proposing a tensor-based model that preserves spatial information and leverages the underlying structures of both background and target components for more robust and accurate detection.

Methods

The core of this paper is the Regularized Tensor Representative Coefficient Model (TVLrS-TRC). The authors propose to decompose the observed HSI tensor ($Y$) into a background tensor ($B$) and a target tensor ($E$), where $Y = B + E$. Crucially, both $B$ and $E$ are represented as the mode-3 product of a Tensor Representation Coefficient (TRC) tensor and a spectral dictionary matrix.

  • Background Modeling: The background TRC ($X$) is assumed to have a low-rank property and exhibit spatial smoothness. To enforce this, the tensor nuclear norm (TNN) is used to characterize low-rankness, and Total Variation (TV) regularization is applied to promote spatial piece-wise smoothness. The spectral dictionary for the background ($U$) is constructed using the Kernel Spectral Angle Mapper (KSAM).
  • Target Modeling: The target TRC ($Z$) is assumed to be sparse, reflecting the fact that targets typically occupy a small portion of the scene. This sparsity is enforced using the $l_1$-norm. A prior target dictionary ($W$) is incorporated, as TD is a supervised problem.
  • Optimization: The overall optimization problem is formulated as a minimization of TNN, TV norm, and $l_1$-norm, subject to the tensor decomposition constraint. This is solved using the Alternating Direction Method of Multipliers (ADMM), which breaks down the complex problem into a series of smaller, manageable sub-problems.

Datasets

The proposed method was evaluated on two real-world hyperspectral datasets:

  • HYDICE: A dataset collected over an urban area. The experiments used a cropped region of interest (80x100 pixels) with 162 spectral bands after preprocessing.
  • AVIRIS: A dataset collected over San Diego, California. A 150x150 pixel region was selected with 191 spectral bands after preprocessing.

The paper does not introduce new benchmarks but compares its performance against existing state-of-the-art detectors on these established datasets.

Key Insights and Ideas

  • Tensor Representation is Key: Moving beyond matrix representations is crucial for HSI analysis because it allows for the simultaneous exploitation of spectral and spatial information without data distortion.
  • Leveraging Physical Characteristics: The novelty lies in explicitly modeling the physical characteristics of the representation coefficients (TRCs) for both background and target. The low-rank nature of background TRCs and the sparsity of target TRCs are fundamental assumptions that lead to improved performance.
  • TV Regularization for Efficiency: Applying TV regularization to the background TRC, rather than the entire background tensor, is a computationally efficient way to enforce spatial smoothness.
  • KSAM for Background Dictionary: Using KSAM for background dictionary selection is a principled approach that leverages spectral similarity to build a robust background model.

Key Results

The TVLrS-TRC method demonstrated superior performance on both HYDICE and AVIRIS datasets compared to several established detectors, including ACE, CEM, CSCR, and DM-BDL.

  • Quantitative Performance: The paper reports the highest Area Under the Curve (AUC) values for the proposed method across various ROC curve metrics (e.g., (PD, PF), (PD, T), (PF, T)). For instance, on the HYDICE dataset, TVLrS-TRC achieved the highest (PD, PF) value. On the AVIRIS dataset, it also showed the highest (PD, PF) and competitive (PD, T) values.
  • Qualitative Performance: Visualizations (2D plots) showed that the proposed method effectively detects targets while suppressing background noise, with target regions clearly preserved. The ROC curves for TVLrS-TRC consistently approached the ideal upper-left corner, indicating excellent detection accuracy and background suppression.

Limitations and Future Work

  • Parameter Tuning: While the paper provides optimal parameter settings for the tested datasets, the sensitivity of the method to parameter choices ($\lambda$ and $\beta$) might be a limitation. Further research could explore adaptive parameter selection strategies.
  • Dictionary Construction: The performance is dependent on the quality of the background and target dictionaries. While KSAM is used for background, and prior information for the target, exploring more advanced dictionary learning techniques could further enhance robustness.
  • Computational Complexity: Although ADMM is used for optimization, the computational cost of tensor operations can still be significant for very large HSI cubes. Future work could focus on further optimizing the algorithm for speed and scalability.
  • Extension to Other Applications: The authors suggest that the proposed tensor-based framework could be extended to other HSI analysis tasks beyond target detection, such as target recognition and unmixing.

Summary updated

Abstract

Target detection based on hyperspectral image (HSI) representations has drawn wide attention given its wide variety of features. The matrix-based approach inevitably loses spatial information and fails to explore the intrinsic multimodal structure of an HSI cube. In this letter, we propose a regularized tensor-based model without altering the data structure. We assume that an observed third-order HSI tensor is decomposed into the sum of a total variation (TV)-regularized low-rank background tensor and a sparse (TVLrS) target tensor. The two tensors are represented as the mode-3 product of a third-order tensor, called the tensor representation coefficient (TRC), and a spectra dictionary matrix. Then, the model is coined as TVLrS-TRC. The background TRC has a low-rank property, contributing to the low-rankness characterization in our model. Moreover, as the size of the background TRC term is smaller than the background tensor, characterizing its local smoothness via TV regularization reduces the computational cost compared with that of the background tensor. Extensive experiments on two real hyperspectral datasets demonstrate the advantage of the proposed method compared with the state of the art.
BibTeX
@article{shang2023regularized,
  author = {Wenting Shang and Mohamad Jouni and Zebin Wu and Yang Xu and Mauro Dalla Mura and Zhihui Wei},
  title = {{Regularized Tensor Representative Coefficient Model for Hyperspectral Target Detection}},
  journal = {IEEE Geosci. Remote. Sens. Lett.},
  year = {2023},
  url = {https://openreview.net/forum?id=7MYfDkLGuKN},
}
Figure