Abstract: Root systems play a crucial role in plant individuals, and parameters including root size, distribution range and three-dimensional structure are of great significance for the research on root functions and belowground ecology. However, root systems are distributed in underground soil, with complex morphology and significant differences in size. As a non-destructive, high-efficiency geophysical exploration method with high positioning accuracy, ground penetrating radar （GPR） has been widely used in the non-destructive detection of root systems. Nevertheless, affected by factors such as the complex structure of tree root systems, the heterogeneity of underground media, and direct waves from ground reflection, the interpretation of GPR data has become extremely difficult. To solve the problems of GPR image interpretation and target detection of tree root systems, this study proposes a root system inversion algorithm based on a multi-scale convolutional generative adversarial network （MCGAN）. The algorithm is applied to invert and interpret GPR images of root systems, so as to reconstruct the spatial position distribution of root systems in real forest environments. First, a dynamic-weighted multi-scale convolutional layer （DWMC） is introduced into the generator of the MCGAN. By dynamically assigning the weights of each convolutional layer, it can efficiently extract complex features of different spatial scales from GPR B-scan images. Meanwhile, the adversarial training mechanism between the generator and the discriminator is adopted to generate high-precision structural maps of multi-target root systems. Second, forward numerical simulation of the radar detection process for tree root systems was carried out using the gprMax3.0 simulation software, and 4 000 sets of root system models and their corresponding GPR data were established based on the real structure of tree root systems. Subsequently, a field forest experiment was conducted in Qingping Forest Farm, Gaotang County, Shandong Province. Populus tomentosa, a typical plant in the forest area, was selected as the research object. GPR scanning was performed on the root systems using the grid method, and profile excavation was carried out along the detection lines, with 132 sets of real field data obtained. Then, preprocessing and data augmentation were performed on the field data, including direct wave removal, depth gain adjustment and median filtering, to eliminate noise and enhance the reflection intensity of root systems. Finally, indicators including the structural similarity index measure （SSIM）, mean square error （MSE）, and mean absolute error （MAE） were adopted to evaluate the feasibility and accuracy of the proposed root inversion network. The research results show that: 1） The MCGAN algorithm can achieve accurate inversion of multi-target root systems, with the SSIM of root inversion reaching 0.874 6, and the MSE and MAE of 137.442 1 and 1.217 9, respectively. 2） Compared with traditional inversion algorithms, the MCGAN network can more accurately capture the complex features of targets at different scales through the multi-scale feature extraction and dynamic weight allocation mechanism, and is more adaptable to the intricate distribution characteristics of tree root systems. 3） The DWMC module can adaptively assign the feature weights of each convolutional branch according to the actual feature distribution of the input image, which solves the problems of inaccurate identification of small-target root systems and high susceptibility to noise caused by fixed weights. This study provides a feasible and effective method for the inversion of underground multi-target root systems, contributes to the non-destructive detection of tree root systems, and provides certain support for the research on the three-dimensional architecture of tree root systems in forest areas.