Deep learning-based inversion for soil moisture estimation

Traditional approaches for soil moisture determination, the gravimetric ovendrying method, are destructive and limited to small-scale applications. Although remote sensing offers large-scale coverage, its accuracy is often affected by environmental factors. Ground-penetrating radar (GPR) is widely employed in near-surface soil characterization; however, existing approaches often suffer from limited inversion accuracy. To address these limitations, this study proposes a nonlinear inversion method based on the U-Net deep learning architecture. Synthetic GPR datasets were produced via finite-difference time-domain (FDTD) electromagnetic modeling, and the Topp relation was applied to transform dielectric permittivity into volumetric water content. The experimental outcomes indicate that the proposed approach attains high inversion accuracy and demonstrates robust generalization performance.