Domain-adaptive self-supervised learning for corrosion detection and 3D building information model mapping in steel tunnels

Abstract

Accurate detection and localization of steel corrosion in tunnel infrastructure remains a major challenge, particularly under conditions of variable lighting, limited accessibility, and visual domain shifts common in real-world inspection scenarios. This study presents a novel integrated framework that automates tunnel inspection by combining self-supervised deep learning, image-based three-dimensional reconstruction, and building information modeling (BIM)-based spatial damage localization. At the core of our approach is a Segformer-based, two-stage domain adaptation model, which leverages pseudo-labeling and confidence masking to improve generalization across visually diverse environments without requiring extensive labeled data. Unlike traditional supervised methods, our model achieves a mean intersection over union (mIoU) of 0.81 and an F1 score of 0.77, demonstrating superior robustness and generalization. Images captured via unmanned aerial vehicles and iPhones were processed to generate a dense point cloud, which was used to construct a three-dimensional (3D) BIM model of the tunnel structure. Corrosion regions were detected and precisely localized within the BIM coordinate system using a custom coordinate estimation method. The final outputs were compiled into a structured database for seamless digital asset management. Overall, the proposed framework offers a scalable, cost-effective, and highly adaptable solution that significantly reduces manual labor and inspection time, with strong potential for broader deployment in infrastructure condition monitoring and digital asset management.