Hybrid data generation and deep learning for GPR-based reconstruction of robotic-built underground structures

Abstract

There is substantial potential for future underground construction operations to be performed by autonomous robots. This paper proposes a 360-degree digital reconstruction framework for robotic-built underground structures using in-pipe rotating ground penetrating radar (GPR). Unlike conventional ground-level applications, placing GPR inside pipes significantly reduces signal attenuation by shortening the distance to the target, enhancing imaging accuracy. To overcome limited data, this paper proposes a high-fidelity in-pipe GPR generator that combines calibrated synthetic data with real-world pipe reflections, clutter, and random noises. Besides, a ‘stochastic-ellipse-union’ method models robot-constructed structures mathematically, ensuring dataset diversity. Moreover, a optimized 2D digital antenna model, calibrated to 97 % accuracy using a genetic algorithm, reduces radargram generation time by 99.2 % compared to traditional 3D methods. Benchmark tests among seven DL models identified ResNet101-enhanced U-Net as optimal, achieving an intersection-over-union score of 0.937, proving the effectiveness of the framework in reconstructing robotic-built underground structures. • A framework for 360-degree reconstruction of underground structure maps. • A hybrid data generator to address the scarcity of in-pipe rotating GPR datasets. • Modelling irregular robotic-injected structures with enhanced diversity and realism. • A calibrated 2D digital antenna model of high realism and computational efficiency. • Seven networks benchmarked to identify the optimal architecture for reconstruction.