This paper focuses on the specific case of anomaly detection from hyperspectral ground-based images in the context of hidden targets in the forest. Compared to nadir viewing, several limitations are identified making it difficult to use conventional anomaly detection methods. Since ground-based image simultaneously contains sky and background, very high contrasts are observed on the transition pixels between the sky and the treetops. Mixed pixels in transition have a very different spectrum from the sky spectrum and the vegetation spectrum. Consequently, detection methods based on distance from a model or local pixel density will interpret mixed pixels as anomalies. To overcome these limitations, this paper proposes a simple processing, to complement standard methods such as distance-to-model methods, for identifying mixed pixels and assigning them low detection score. The approach consists in isolating potentially mixed pixels and calculating a distance between these pixels and the closest point on the axis formed by neighboring classes. The benefit of the proposed method (named ADMC) is illustrated using hyperspectral data acquired during the DEBELA European project field campaign, where the anomalies are military targets concealed into the vegetation. In the illustrative image, the standard detection score for treetop pixels is higher than that of the targets. The proposed method drastically reduces the score of these mixed pixels, dividing the rate of false positives before detection by ten.

© (2024) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.

Stephanie Doz, Philippe Déliot, and Pierre-Yves Foucher, "Specific processing of mixed pixels for anomaly detection, case study on DEBELA campaign," Proc. SPIE 13200, Electro-Optical and Infrared Systems: Technology and Applications XXI, 132001D (Presented at Security + Defence: September 19, 2024; Published: 1 November 2024); https://doi.org/10.1117/12.3033669.