Hyperspectral Anomaly Detection via Sparse Dictionary Learning Method of Capped Norm

Yuan Yuan, Dandan Ma, Qi WangIEEE Access 2019 · DOI

Introduction

Hyperspectral Anomaly Detection (HAD) is a critical task in remote sensing, aimed at identifying pixels that deviate significantly from their spectral surroundings. These anomalies often represent small, man-made objects like vehicles or structures hidden within natural backgrounds. For decades, the Reed-Xiaoli (RX) algorithm has served as the benchmark. However, RX relies on a Gaussian distribution assumption that rarely holds in complex real-world scenes and is highly susceptible to "pollution"—where the anomalies themselves distort the background statistics, making them harder to detect.

The SDLCN Approach

To overcome these limitations, this research introduces SDLCN (Sparse Dictionary Learning of Capped Norm). Instead of relying on rigid statistical models, SDLCN learns a representative dictionary of the background. The core principle is that background pixels can be reconstructed as a linear combination of a few "basis atoms" from this dictionary, while anomalies will exhibit high reconstruction errors.

The Capped Norm Innovation

Traditional learning methods use quadratic loss functions that attempt to minimize all errors, inadvertently forcing the model to learn anomalies if they are present in the training set. SDLCN employs a capped norm loss function with a threshold ($\epsilon$). When a reconstruction error exceeds this threshold, the loss is capped, effectively instructing the model to ignore the pixel as an outlier. This ensures the dictionary remains "purely" background-focused.

Clustering-based Selection

To ensure the dictionary captures the full diversity of the background, the method utilizes K-means++ clustering. By selecting training samples from every cluster, the model accounts for rare background materials, reducing false positives.

Datasets and Evaluation

The performance of SDLCN was validated using several high-dimensional HSI benchmarks, including the HYDICE Urban scene and AVIRIS sub-images from San Diego.

A sample hyperspectral scene used for anomaly detection evaluation.

These datasets present various challenges, from urban clutter to complex coastal environments, featuring targets like airplanes and man-made panels.

Key Results and Discussion

The SDLCN method demonstrated superior performance against eight state-of-the-art competitors. In the AVIRIS1 dataset, SDLCN achieved an Area Under the Curve (AUC) of 0.9659, significantly outperforming traditional dictionary learning (0.8025).

A ground truth anomaly map showing the target pixels to be detected.

Beyond accuracy, the method proved computationally efficient. For the Urban dataset, SDLCN completed its task in approximately 78 seconds, whereas non-parametric methods like SVDD required over 3,700 seconds. The resulting anomaly maps are notably cleaner, with distinct hits on targets and minimal background noise.

Conclusion

By integrating a capped norm into sparse dictionary learning, SDLCN provides a robust, adaptive solution for hyperspectral anomaly detection. While the method currently requires some hyperparameter tuning, its ability to ignore outliers during training marks a significant step forward in creating "pollution-proof" detection models. Future research will focus on automating parameter selection to create a fully autonomous detection framework.

Summary updated

BibTeX
@article{yuan2019hyperspectral,
  author = {Yuan Yuan and Dandan Ma and Qi Wang},
  title = {{Hyperspectral Anomaly Detection via Sparse Dictionary Learning Method of Capped Norm}},
  journal = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2894590},
}
Figure