Deep Calibration and Operator Learning for Ground Penetrating Radar Imaging

The accurate imaging of underground scenes using wave-based sensor technologies, such as ground penetrating radar, presents challenges due to ill-posedness, formulation complexities, and computational demands. In this paper, we propose a machine learning-based approach that leverages a learned forward model to simulate wave-object interactions inspired by physics principles as well as the calibration to realistic antenna configurations. Our approach combines a learned wave propagation model, referred to as Born FNO, with a deep calibration network that maps a point-receiver scattered wavefields to the response of a desired receiving antenna architecture. We evaluate our method on a simulated dataset that includes multiple ground layers and complex target structures. We demonstrate that our proposed calibration network enables the reconstruction of permittivity distributions and outperforms a linear calibration operator trained on the same dataset by over 4.5 dB in peak signal-to-noise ratio.