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

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.