DOWSER: A DINO-based One-shot Water Seepage Recognition network for TBM tunnel inspection

Zehao Ye, Neng Wang, Huamei Zhu, Qian Zhang, Zili Li, Jelena Ninić

The industrial deployment of supervised models for tunnel damage detection is limited by the high cost of data collection and pixel-level annotation. Although deep learning has advanced defect detection, its heavy data requirements remain impractical for dynamic construction environments. To address this issue, we propose DOWSER ( D INO-based O ne-shot W ater S eepage R ecognition), a framework designed to achieve high-precision seepage segmentation with only a single annotated sample. By exploiting the highly structured and geometrically coherent environment of TBM tunnels, DOWSER utilizes a frozen DINOv3 backbone and a Dual-Stream architecture to unifies a non-parametric Prototypical Stream that generates smooth probabilistic gradients for robust generalization, combined with a parametric Multilayer Perceptron (MLP) Stream for effective background suppression. Experimental results show that, using only one labelled image and 2 min for automated model construction, DOWSER achieves an mIoU of 86.42% on the validation set, reaching 95% of the performance of fully data-trained supervised models. This achievement establishes an automated, “cold-start” workflow compatible with digital twins, bridging the gap between algorithmic innovation and practical, efficient on-site decision-making.