Deep Learning and Geometric Modeling for 3D Reconstruction of Subsurface Utilities from GPR Data.

Area of Science:

• Geophysics
• Computer Vision
• Geospatial Analysis

Background:

• Accurate underground utility mapping is crucial for safe excavation.
• Ground Penetrating Radar (GPR) interpretation for subsurface utilities is complex.
• Existing methods often struggle with real-world data complexity and scalability.

Purpose of the Study:

• To develop a novel deep learning pipeline for 3D reconstruction of buried linear utilities from GPR B-scan data.
• To compare the performance of YOLOv8, YOLOv11, and Mask R-CNN for utility reflection detection.
• To create a practical and scalable solution for subsurface utility mapping.

Main Methods:

• Utilized high-resolution GPR B-scan data from a real-world dataset.
• Employed YOLOv8, YOLOv11, and Mask R-CNN for bounding box and keypoint detection of hyperbolic reflections.
• Applied 3D DBSCAN clustering and RANSAC-based line fitting for spatial trajectory approximation.

Main Results:

• Mask R-CNN achieved the highest keypoint F1-score (0.822) and bounding box F1-score (0.867).
• The 3D reconstruction pipeline yielded an average RMSE of 0.06 for fitted 3D utility paths.
• The hybrid model demonstrated superior performance on real-world data compared to YOLO models.

Conclusions:

• The proposed deep learning pipeline offers a practical and scalable solution for subsurface utility mapping.
• Direct summit point detection and a geometric 3D reconstruction pipeline overcome limitations of prior studies.
• This method enhances generalizability to complex real-world GPR data scenarios.