Presentation Information
[B-1A-18]ML Prediction of Deep Soil Permittivity from AMeDAS Data for Correcting GPR Depth Estimation Errors
〇Jun Sonoda1 (1. National Instituite of Technology, Sendai)
Keywords:
Ground-penetrating radar,Deep soil permittivity,Machine learning,AMeDAS,Depth-error correction
Ground-penetrating radar (GPR) requires the relative permittivity of soil for estimating the depths of buried pipes and road cavities. However, because direct in situ measurement is often difficult, empirical values are commonly used, resulting in depth estimation errors. In this study, we investigated machine-learning-based prediction of deep soil permittivity and correction of GPR depth errors using soil permittivity measured from 2019 to 2025 by TDR sensors installed at depths of 1 m and 2 m beneath asphalt pavement at the Hirose Campus of National Institute of Technology, Sendai College, together with AMeDAS precipitation, air temperature, and sunshine duration data. Ridge regression, random forest, and gradient boosting models were compared, and their generalization performance was evaluated using a year-based walk-forward validation. The results showed that ridge regression effectively reproduced seasonal variations and rainfall responses observed in the measured permittivity. At a depth of 2 m, the mean depth error obtained using an empirical permittivity value was reduced from 17.1 cm to 1.1 cm. These results demonstrate the potential of reducing GPR depth errors by up to 94% using only publicly available meteorological data.