ActiveGPR: Multi-Scale Geometric Reasoning for Adaptive Subsurface Sensing and Imaging

Abstract

Subsurface pipe mapping with ground penetrating radar (GPR) remains challenging because overlapping pipe structures interfere with reflected signals and degrade reconstruction quality, while blind or grid-based scanning wastes measurements in sparse environments. We propose ActiveGPR, a unified framework for adaptive subsurface imaging and scan planning. TEUNet operates on local A-scan arrays to recover patch-level 3D pipe geometry, while a masked autoencoder provides global map priors and a topology network predicts local continuation directions from scanned patches. These cues are fused into a belief state and used by a reinforcement learning policy to select scan locations under a fixed budget. In controlled gprMax simulations on 40 test maps, the feature-fusion policy reaches 0.902 coverage by step 60 and 0.987 final coverage, compared with 0.569 and 0.796 for Grid Search.