Structured Background Modeling for Hyperspectral Anomaly Detection

Fei Li, Lei Zhang, Xiuwei Zhang, Yanjia Chen, Dongmei Jiang, Genping Zhao, Yanning Zhang2018 · DOI

TL;DR

This paper introduces a novel hyperspectral anomaly detection method that leverages the block-diagonal structure of background data. By modeling multi-mode background characteristics through clustering and spatial-spectral feature learning, the proposed method significantly improves detection accuracy compared to existing state-of-the-art techniques.

What Problem Does This Paper Address?

Hyperspectral images (HSIs) offer rich spectral information, making them valuable for applications like target detection. However, anomaly detection in HSIs is challenging due to the cluttered nature of imaging scenes, which leads to complex, multi-mode background distributions. Traditional methods often struggle to accurately model these backgrounds, leading to suboptimal detection performance. Existing representation-based methods, while promising, often overlook the intricate structure of the background, treating it as a monolithic entity. This paper aims to address this by developing a method that explicitly models the multi-mode structure of the background, leading to more robust and accurate anomaly detection.

Methods

The core of the proposed method lies in its novel approach to background modeling. It begins by representing each pixel using a local spatial-spectral patch, capturing both spatial context and spectral information. These patches are then clustered to group similar background pixels, effectively addressing the multi-mode nature of the background. For each cluster, a spatial-spectral dictionary is learned using Principal Component Analysis (PCA). This process is designed to capture the dominant features of each background cluster while suppressing anomalous information.

The key innovation is the exploitation of a block-diagonal structure that emerges when background pixels are represented by these learned dictionaries. The authors theorize and demonstrate that background pixels can be well-represented by their corresponding cluster dictionaries, leading to a block-diagonal structure in the representation matrix. Anomalies, on the other hand, are expected to be poorly represented by these dictionaries, exhibiting a sparse structure.

This observation forms the basis of a Block-Diagonal Structure Based Low-Rank Representation (BDSLRR) model. The model aims to decompose the HSI into a low-rank background component and a sparse anomaly component. The optimization of this model is achieved using the Alternating Direction Method of Multipliers (ADMM) algorithm. The final anomaly detection is performed by analyzing the sparsity of the resulting anomaly matrix.

Datasets

The proposed method was evaluated on five datasets:

  • Synthetic Hyperspectral Dataset: Created by implanting synthetic sub-pixel anomalies into a real HSI scene from the San Diego airport area.
  • Real San Diego Airport Dataset: A subset of the San Diego airport image, with small aircraft used as anomalies.
  • HYDICE Urban Dataset: A publicly available dataset with a spectral resolution of 10 nm and spatial resolution of 2x2m.
  • Pavia Center (PaviaC) Dataset: Acquired by the ROSIS sensor, this dataset features vehicles and bare soil as anomalies against a background of bridge, water, and shadows.
  • Target Detection Blind Test Project Dataset: Collected over Cook City, Montana, this dataset includes various fabric panel and vehicle targets.

Key Insights and Ideas

  • Block-Diagonal Structure as a Background Signature: The paper's central idea is that by modeling background clusters with learned dictionaries, a block-diagonal structure emerges in the representation matrix. This structure is more robust than a simple low-rank assumption, especially in complex HSI scenes.
  • Spatial-Spectral Features for Robust Dictionary Learning: Incorporating spatial information through local patches before dictionary learning significantly enhances the dictionary's ability to capture relevant background features and suppress anomalies.
  • PCA for Clean Dictionary Extraction: Using PCA for dictionary learning, followed by discarding less significant components, effectively removes anomalous information that might have been initially included in a cluster.
  • ADMM for Efficient Optimization: The ADMM algorithm provides an effective and efficient way to solve the complex optimization problem posed by the BDSLRR model.

Key Results

The proposed BDSLRR method demonstrated superior performance across all tested datasets.

  • Synthetic Dataset: BDSLRR achieved an AUC of 0.99711, outperforming all other compared methods.
  • San Diego Dataset: BDSLRR achieved the highest AUC, confirming its ability to outperform traditional and state-of-the-art detectors.
  • Urban Dataset: BDSLRR achieved the highest AUC score, indicating superior detection probability.
  • Pavia Dataset: Despite other methods achieving AUC scores above 0.95, BDSLRR achieved the highest values, demonstrating its robustness and stability.
  • Blind Test Dataset: BDSLRR achieved an AUC of 0.84011, outperforming other methods, especially on more complicated and larger-sized datasets.

The paper also highlights that the proposed method achieves better measurement accuracy (lowest SER and highest AER) compared to other algorithms across all datasets.

Limitations and Future Work

  • Parameter Tuning: While the proposed method is generally robust, the number of clusters requires manual tuning for each dataset. However, the authors note that the optimal value often falls within a reasonable range (6-20), making it applicable to real-world scenarios.
  • Computational Complexity: Although the ADMM algorithm is efficient, the overall computational complexity is moderate. Future work could focus on further optimizing the computational efficiency.
  • Exploiting Other Structures: The paper focuses on the block-diagonal structure. Future research could explore other potential structures within HSI data for anomaly detection.

Summary updated

BibTeX
@misc{li2018structured,
  author = {Fei Li and Lei Zhang and Xiuwei Zhang and Yanjia Chen and Dongmei Jiang and Genping Zhao and Yanning Zhang},
  title = {{Structured Background Modeling for Hyperspectral Anomaly Detection}},
  year = {2018},
  doi = {10.3390/s18093137},
}
Figure