Abstract: To enhance the efficiency and accuracy of trunk radar wave interpretation, this study proposes the application of an improved U-Net deep learning algorithm for radar wave inversion of internal trunk structures, achieving the inversion estimation of both internal structural features and dielectric constant distributions. Numerical models of trunk stratification and hollow decay were constructed, and theoretical datasets were generated through forward modeling. The U-Net inversion algorithm was improved by integrating an Atrous Spatial Pyramid Pooling （ASPP） module to expand the receptive field, establishing an inversion mapping from radar wave scan images to dielectric constant distribution images. Simulation results demonstrated that the proposed method enables rapid inversion of internal dielectric constants in tree trunks with high precision, achieving high accuracy with a mean squared error （MSE） of 1.109 in numerical reconstruction and multi-scale structural similarity （MS-SSIM） of 0.979 in cross-scale fidelity. In practical detection of ancient Platycladus orientalis trunks, the inverted geometric contours aligned with the internal defect ranges analyzed by existing commercial tree radar systems. Furthermore, the method provided precise and detailed dielectric constant distributions within trunks, offering an effective detection tool for subsequent research on the spatial distribution of trunk internal physical properties.