Abstract
Hyperspectral images often feature complex land-cover distributions, making effective background representation a significant challenge for anomaly detection. This paper proposes a novel framework: Sparsity Score Estimation for Hyperspectral Anomaly Detection (SSEAD). By leveraging sparse learning, the framework constructs an overcomplete dictionary and estimates the frequency of each dictionary atom used in reconstruction—termed the Atom Usage Probability (AUP). These frequencies are then transformed into a sparsity score for each pixel. To enhance performance, the framework incorporates dictionary-based background feature transformation and iterative reweighting. Experimental results on two real-world hyperspectral datasets demonstrate that SSEAD achieves superior performance compared to state-of-the-art methods.
Introduction
Hyperspectral sensors capture hundreds of narrow spectral bands, providing a unique spectral signature for every pixel. Anomaly Detection (AD) in these images aims to identify pixels that deviate significantly from their surroundings, such as man-made objects in natural environments, without prior knowledge of the target's signature.
Traditional methods, like the Reed-Xiaoli (RX) detector, often struggle because they rely on global statistics that can be easily contaminated by the anomalies themselves. Furthermore, the high dimensionality and redundancy of hyperspectral data make it difficult to distinguish subtle anomalies from complex, heterogeneous backgrounds.
Methodology: The SSEAD Framework
The SSEAD framework addresses these challenges by shifting the focus from direct pixel modeling to sparse representation.
Dictionary Learning and Atom Usage
The process begins by learning an overcomplete dictionary—a set of representative spectral "atoms" that can reconstruct any pixel in the scene. The core innovation of SSEAD is the calculation of the Atom Usage Probability (AUP), which measures how frequently each atom is utilized across the entire image. Atoms that are used frequently are likely representative of the background (e.g., vegetation or soil), while those used rarely are indicative of anomalies.
From Rarity to Sparsity Score
Borrowing from information theory, the framework calculates the Negative Log AUP (NLAUP) for each atom. As shown in Figure 1, an atom with a low frequency of use (low AUP) results in a high NLAUP value, effectively quantifying the "information surprise" of that atom's presence.
Each pixel is then assigned a Sparsity Score based on the NLAUP values of the atoms required for its reconstruction. A pixel that relies on "rare" atoms receives a high anomaly score. To further refine the results, the authors employ background feature transformation to suppress common signatures and iterative reweighting to better separate background atoms from anomaly-related ones.
Experimental Evaluation
The SSEAD framework was evaluated using two benchmark datasets: the HYDICE (Urban) scene, containing 162 spectral bands with 10 man-made vehicles as targets, and the AVIRIS (Airport) scene, featuring 189 spectral bands with 3 aircraft as targets.
Performance and Separability
The framework was compared against seven state-of-the-art detectors, including RX, Kernel-RX (KRX), and Collaborative Representation-based Detector (CRD).
As illustrated in the ROC curves (Figure 2), SSEAD consistently achieved higher detection rates at lower false-alarm rates than its competitors. Furthermore, separability analysis indicated that SSEAD provides a clearer distinction between background and anomaly scores, reducing ambiguity for end-users.
Conclusion
SSEAD offers a robust, mathematically elegant approach to hyperspectral anomaly detection by quantifying the "rarity" of spectral components. While the method demonstrates superior accuracy, future work is needed to address the computational intensity of sparse coding for real-time applications and to develop automated parameter tuning for diverse environments.
Summary updated