An Improved YOLOv8-Based Framework for Tunnel Leakage Detection and Structural Inspection
Year: 2026
Type: article
Abstract: Image-based detection of tunnel leakage can support structural inspection and maintenance management in underground construction. YOLO-based object detection methods have shown promising potential in tunnel lining defect inspection and other engineering applications, providing a foundation for their further application to tunnel leakage detection. However, dim lighting, partial occlusion, and complex backgrounds remain major challenges in identifying visible leakage regions. To address these issues, this study proposes YOLOv8-AS, an improved YOLOv8-based model for detecting and localizing visible tunnel leakage. The model integrates a generalized feature pyramid network (GFPN) for cross-scale feature fusion, RepViT blocks for efficient feature extraction, content-aware reassembly of features (CARAFE) for upsampling, and SIoU loss for bounding-box regression. The model was evaluated on a tunnel leakage image dataset through ablation experiments and comparisons with YOLOv8, YOLOv5, SSD, and Faster R-CNN. YOLOv8-AS achieved 90.0% precision, 82.6% recall, an F1 score of 86.1%, and a mean average precision (mAP) of 86.9%, compared with 76.0% for the baseline YOLOv8. The model file size was 5.4 MB. Visual results demonstrate its detection performance under low illumination, partial occlusion, and complex backgrounds. The ablation results reveal the contributions of different feature extraction, multiscale fusion, and localization strategies to tunnel leakage detection, providing methodological insights into the visual identification of irregular and partially occluded defects in complex tunnel environments. The proposed framework provides a basis for automated leakage localization and defect assessment, supporting structural inspection and maintenance management and contributing to the development of intelligent tunnel operation and maintenance technologies.
Source: Buildings
Authors: Qianyi Zhao, Hao Zhang, Haowen Jia
Institutions: Southeast University, Shijiazhuang Tiedao University
Language: English
Subfield: Civil and Structural Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7226033414
