Self-Supervised and Unsupervised Multispectral Anomaly Detection for Unknown Substance and Surface Defect Identification

Cansu Beyaz, Mohamed Farag, Peer Schütt, Tobias Hecking, Jonas Grzesiak, Christoph Geiß, Ribana Roscher
Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), PMLR 307:27-38, 2026.

Abstract

Autonomous systems and environmental monitoring require reliable detection of unknown hazardous materials to prevent traffic accidents and ecological damage resulting from chemical spills, fuel leaks, and agricultural runoff. Traditional detection methods, such as gas chromatography, pose contamination risks and cause delays, while laser-based techniques rely on prior localization of potential hotspots. This paper addresses the automatic detection of unknown materials (e.g., fertilizer, sand, soil) and surface anomalies (e.g., cracks, holes) without requiring labeled anomaly examples during training. We employ unsupervised and self-supervised deep learning methods to learn normal patterns and identify deviations. Our approach evaluates four models: convolutional and vision transformer-based autoencoders, and two self-supervised methods, SimCLR and Barlow Twins. Experiments conducted on multispectral road images from the German Aerospace Center and the MVTec hazelnut dataset demonstrate that the ViT-based autoencoder outperforms its convolutional counterpart, while Barlow Twins achieves superior anomaly localization compared to SimCLR. These results highlight the potential of efficient deep learning models for enhancing road safety and environmental protection through early detection of potentially hazardous substances before they cause harm.

Cite this Paper

BibTeX
@InProceedings{pmlr-v307-beyaz26a, title = {Self-Supervised and Unsupervised Multispectral Anomaly Detection for Unknown Substance and Surface Defect Identification}, author = {Beyaz, Cansu and Farag, Mohamed and Sch{\"u}tt, Peer and Hecking, Tobias and Grzesiak, Jonas and Gei{\ss}, Christoph and Roscher, Ribana}, booktitle = {Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL)}, pages = {27--38}, year = {2026}, editor = {Kim, Hyeongji and Ramírez Rivera, Adín and Ricaud, Benjamin}, volume = {307}, series = {Proceedings of Machine Learning Research}, month = {06--08 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v307/main/assets/beyaz26a/beyaz26a.pdf}, url = {https://proceedings.mlr.press/v307/beyaz26a.html}, abstract = {Autonomous systems and environmental monitoring require reliable detection of unknown hazardous materials to prevent traffic accidents and ecological damage resulting from chemical spills, fuel leaks, and agricultural runoff. Traditional detection methods, such as gas chromatography, pose contamination risks and cause delays, while laser-based techniques rely on prior localization of potential hotspots. This paper addresses the automatic detection of unknown materials (e.g., fertilizer, sand, soil) and surface anomalies (e.g., cracks, holes) without requiring labeled anomaly examples during training. We employ unsupervised and self-supervised deep learning methods to learn normal patterns and identify deviations. Our approach evaluates four models: convolutional and vision transformer-based autoencoders, and two self-supervised methods, SimCLR and Barlow Twins. Experiments conducted on multispectral road images from the German Aerospace Center and the MVTec hazelnut dataset demonstrate that the ViT-based autoencoder outperforms its convolutional counterpart, while Barlow Twins achieves superior anomaly localization compared to SimCLR. These results highlight the potential of efficient deep learning models for enhancing road safety and environmental protection through early detection of potentially hazardous substances before they cause harm.} }
Endnote
%0 Conference Paper %T Self-Supervised and Unsupervised Multispectral Anomaly Detection for Unknown Substance and Surface Defect Identification %A Cansu Beyaz %A Mohamed Farag %A Peer Schütt %A Tobias Hecking %A Jonas Grzesiak %A Christoph Geiß %A Ribana Roscher %B Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL) %C Proceedings of Machine Learning Research %D 2026 %E Hyeongji Kim %E Adín Ramírez Rivera %E Benjamin Ricaud %F pmlr-v307-beyaz26a %I PMLR %P 27--38 %U https://proceedings.mlr.press/v307/beyaz26a.html %V 307 %X Autonomous systems and environmental monitoring require reliable detection of unknown hazardous materials to prevent traffic accidents and ecological damage resulting from chemical spills, fuel leaks, and agricultural runoff. Traditional detection methods, such as gas chromatography, pose contamination risks and cause delays, while laser-based techniques rely on prior localization of potential hotspots. This paper addresses the automatic detection of unknown materials (e.g., fertilizer, sand, soil) and surface anomalies (e.g., cracks, holes) without requiring labeled anomaly examples during training. We employ unsupervised and self-supervised deep learning methods to learn normal patterns and identify deviations. Our approach evaluates four models: convolutional and vision transformer-based autoencoders, and two self-supervised methods, SimCLR and Barlow Twins. Experiments conducted on multispectral road images from the German Aerospace Center and the MVTec hazelnut dataset demonstrate that the ViT-based autoencoder outperforms its convolutional counterpart, while Barlow Twins achieves superior anomaly localization compared to SimCLR. These results highlight the potential of efficient deep learning models for enhancing road safety and environmental protection through early detection of potentially hazardous substances before they cause harm.
APA
Beyaz, C., Farag, M., Schütt, P., Hecking, T., Grzesiak, J., Geiß, C. & Roscher, R.. (2026). Self-Supervised and Unsupervised Multispectral Anomaly Detection for Unknown Substance and Surface Defect Identification. Proceedings of the 7th Northern Lights Deep Learning Conference (NLDL), in Proceedings of Machine Learning Research 307:27-38 Available from https://proceedings.mlr.press/v307/beyaz26a.html.

Related Material