Beyond RGB: out-of-distribution detection for hyperspectral material recognition

Abstract

Hyperspectral imaging (HSI) extends conventional RGB imaging by measuring light in tens to hundreds of narrow spectral bands, often extending into the short-wave infrared. The resulting wavelength-dependent responses provide information about material properties and chemical composition, enabling material recognition beyond RGB appearance alone. Machine-learning-based HSI systems are typically trained under a closed-set assumption and may confidently assign previously unseen materials to known classes. Similar to RGB imaging, reliable out-of-distribution (OOD) detection is therefore crucial for trustworthy hyperspectral recognition. In this work, we investigate whether post-hoc OOD detectors developed primarily for RGB image classification transfer to hyperspectral material recognition. Using a standardized hyperspectral line-scan setup, we acquire object-level data on polyethylene (PE), polypropylene (PP), and polyethylene terephthalate (PET), as well as previously unseen materials. A fixed spectral-spatial R(2+1)D network is trained on reduced hyperspectral data obtained from three expert-selected spectral regions, as typically used for such a classification task. We evaluate four categories of post-hoc OOD detection methods: output-based, feature-based, model-enhancement, and intermediate-feature methods on sensor-matched near-OOD objects and complementary cross-domain and synthetic shifts. The investigated OOD detection methods exhibit markedly different transfer behavior. Feature-based and intermediate-feature methods generally outperform output-based and model-enhancement approaches. Moreover, method rankings differ across AUROC, FPR95, and FPR99, showing that global ranking metrics alone are insufficient for deployment-oriented method selection. The results presented highlight the importance of representation-aware OOD detection and evaluation at high ID-retention operating points for reliable hyperspectral material recognition.