Multispectral Anomaly Detection: Comparison of sensor bands in conventional and machine learning approaches

Abstract. Detecting camouflaged objects in UAV-based reconnaissance imagery is challenging, particularly beyond the visible spectrum where labeled data, required for training advanced methods, is limited. Multi-spectral sensing and synthetic thermal imagery promise improved anomaly-based camouflage detection. This work systematically compares sensor-band configurations and anomaly detection paradigms, including simulated long-wave infrared (LWIR) references, for multi-spectral camouflage detection. We construct 15 datasets of multi-spectral image stacks by not only combining real imagery, but also real and synthetic LWIR, generated via 3D scene reconstruction and heat-transfer simulation. Four detectors, ranging from statistical-based, conventional machine-learning, and deep learning, are evaluated against the image band combinations, focusing on methods suitable for near real-time operation. Results show that complementary RGB, near-infrared, and LWIR information, exploited by lightweight deep learning detectors, enables accurate multi-spectral camouflage detection, and the applied thermal simulations for synthesizing LWIR are not yet accurate or well-integrated enough to serve as beneficial reference channels, suggesting that the current network and training setup cannot yet exploit the simulated channel. This motivates for future improvement of thermal models, reduction of reality-to–simulation gaps, and tailoring deep learning architectures or training strategies to harness synthetic data.