Leakage Detection in Subway Tunnels Using 3D Point Cloud Data: Integrating Intensity and Geometric Features with...

Area of Science:

• Geosciences
• Civil Engineering
• Computer Science

Background:

• Tunnel leakage detection from 3D point clouds is challenging due to noise and complex morphology.
• Existing methods struggle with accurate feature extraction and segmentation of leakage patterns.

Purpose of the Study:

• To develop an accurate and efficient method for detecting tunnel leakage using mobile laser scanning (MLS) point clouds.
• To integrate intensity and geometric features for improved classification performance.
• To compare the proposed XGBoost method against other machine learning and deep learning models.

Main Methods:

• Noise filtering using the RANSAC algorithm to remove non-leakage tunnel objects.
• Feature extraction combining intensity and k-neighborhood geometric features.
• Optimal neighborhood scale selection based on F1-score.
• Binary classification using the XGBoost classifier for leakage detection.

Main Results:

• The proposed method achieved high F1-scores (91.18% and 97.84%) on two datasets.
• Demonstrated robust generalization across four heterogeneous datasets.
• XGBoost outperformed Random Forest, AdaBoost, LightGBM, and CatBoost in accuracy and efficiency.
• Superior performance compared to PointNet, PointNet++, and DGCNN in both accuracy and computational efficiency.

Conclusions:

• Integrating geometric features significantly enhances leakage detection accuracy in MLS point clouds.
• The XGBoost-based approach offers a robust and efficient solution for tunnel leakage detection.
• The method shows strong potential for practical applications in tunnel inspection and maintenance.