New hyperspectral anomaly detection method robust to mixed noise types
Sato & Ono's convex optimization separates background, anomaly, and noise in HS images
A new paper from Koyo Sato and Shunsuke Ono tackles a persistent weakness in hyperspectral anomaly detection: noise robustness. While most existing methods assume clean Gaussian noise or ignore it entirely, real-world HS images suffer from sparse noise, stripe noise, and calibration errors that severely degrade detection. The authors formulate a constrained convex optimization problem that jointly decomposes an image into three components—background, anomaly, and mixed noise—and then recover the anomaly map directly. They solve it with an efficient preconditioned primal-dual splitting algorithm, allowing the model to scale to large hyperspectral cubes without sacrificing accuracy.
Experiments across seven real hyperspectral datasets (including urban, agricultural, and industrial scenes) show the method achieves detection accuracy comparable to top approaches like LRX and CRD on clean images. But when artificial mixed noise (Gaussian + sparse + stripe) is added, the new method maintains >90% detection rates while others drop below 60%. The work is especially relevant for satellite and drone-based sensing where sensor imperfections are common. The full paper is available on arXiv (2401.14814) and submitted to IEEE JSTARS.
- Method decomposes HS images into background, anomaly, and three noise types (Gaussian, sparse, stripe) via constrained convex optimization
- Uses preconditioned primal-dual splitting algorithm for efficient large-scale optimization
- Tested on 7 real datasets; maintains >90% detection under mixed noise vs. <60% for state-of-the-art methods
Why It Matters
Enables reliable anomaly detection in real-world hyperspectral imaging where sensor noise is inevitable, improving satellite and drone surveillance.