To address the detection challenges of progressive corrosion in buried oil pipelines under the geological conditions of high porosity, strong water sensitivity, and high soluble salt content in loess areas, and to reveal the electromagnetic response mechanism of pipeline corrosion and establish accurate non-destructive testing methods, this study combines ground-penetrating radar (GPR) measurements with numerical simulation techniques. Using the finite-difference time-domain (FDTD) method, a two-dimensional numerical model of progressive pipeline corrosion is constructed with the gprMax2D software to analyze the dynamic variations of permittivity and conductivity in the corrosion halo (soil-iron-water-air mixture) during the corrosion process. The three instantaneous attributes of radar signals are employed to reveal the physical evolution of corrosion signals. Based on this theoretical foundation, an end-to-end inversion model based on U-Net is built to establish a nonlinear mapping between B-scan images and permittivity distributions, and a hybrid loss function combining L1 and multi-scale structural similarity (MS-SSIM) is introduced to achieve accurate inversion of corrosion depth. Based on measured parameters in loess areas, 1,000 pairs of B-scan images and ground-truth permittivity maps are constructed, incorporating randomized designs of multiple parameters such as burial depth, horizontal position, and corrosion degree. Data augmentation is performed via flipping and Gaussian noise injection. The results show that pipeline corrosion significantly alters the electromagnetic parameters of the surrounding soil, leading to regular changes in the time delay, intensity, and waveform characteristics of radar reflection signals; the greater the corrosion depth, the more severe the signal disturbance and the more frequent the phase jumps. The consistency between forward-simulated images and measured radar images verifies the accuracy of the numerical model. The inversion model achieves a mean absolute error (MAE, normalized) as low as 0.01798 and a structural similarity index (SSIM) of 0.873 for corrosion depth prediction. This study validates the feasibility of the end-to-end inversion method in simulation scenarios while systematically revealing the inherent limitations of purely forward-data-driven models in practical engineering applications, providing a clear problem orientation and data foundation for subsequent research on domain-adaptive inversion methods.